Preface

In spring of 1996, Henry Gleitman taught his 100th introductory psy-chology lecture course. This happy event provided the opportunity forHenry and Lila Gleitman’s students and colleagues to reflect on the con-tributions the two of them have made over their distinguished careers.Such reflection led to a convocation in the spring in Philadelphia; theconvocation led to the writing of these essays in honor of Henry andLila Gleitman.

The essays contained in this volume are organized into three parts.Part I contains an essay by the editors, outlining the history of Henryand Lila’s careers, both singly and collaboratively, and the impact theyhave had on the fields of perception, language, and cognition. Readerswho have not had the pleasure of knowing Henry and Lila might wantto know a bit about them, and knowing a bit about them will, perhaps,help readers to appreciate the essays that follow. Part II contains essaysprincipally addressing Henry’s contributions as a teacher and scholar.These essays are only slightly modified versions of the addresses pre-sented by Henry’s colleagues during the celebration of his 100th psy-chology course, and with the exception of the last chapter by Lamm, areorganized chronologically by the dates during which the contributorswere chairs of psychology at Penn. They include an early history ofHenry’s teaching at Swarthmore, his influence on the development ofpsychology at Penn, and the trials and tribulations accompanying thelengthy gestation of his book, Psychology.

Part III principally contains essays from former students of Henryand Lila’s, describing their current research and its origins in theGleitman “seminar” (described in more detail in the introduction).Attesting to the continuing impact of the seminar, the last two essaysare written by current faculty members at Penn, who have benefitedfrom the seminar much as past graduate students have. The essays inPart III, like those in Part II, are organized chronologically, from the ear-liest students to the latest. This last part is the lengthiest, but still reflects

only some (by no means all) of the impact that Henry and Lila have hadon the field, through their students.

We hope that readers of this volume will take as much pleasure inreading these essays as we, their colleagues, have had in putting to-gether this celebration.

x Preface

Part I

Introduction

Within the fields of psychology, linguistics, and cognitive science, thenames of Henry and Lila Gleitman are well known. However, as theGleitmans have often said, one can truly understand a particular contri-bution only within its historical context. We therefore present a briefhistory of the Gleitmans, both singly and collaboratively, with the hopethat understanding this history will enhance the reader’s enjoyment ofthe remaining contributions.

The Early Years

Henry was born in Leipzig, Germany, on January 4, 1925. He and hisfamily came to the United States in August of 1939, among the very lastJews to leave Germany. Henry attended City College of New York from1942 to 1946. This was the City College of no tuition and very high stan-dards, the City College that produced more Nobel Laureates than anyother undergraduate institution in the world. It was the City Collegewhere everything was hotly debated, everything from world politics tothe adequacy of behaviorism. It was, in short, one of America’s fore-most homes for the development of intellectuals. Henry took a B.S. inpsychology there.

Henry’s introductory psychology instructor was Kenneth Clark. Butthe lecturers who Henry says influenced him most were GardnerMurphy—in personality and history and systems—and the GestaltistMartin Scherer—in physiological and experimental. Murphy, accordingto Henry, was an elegant man with a very broad view of psychology,and Scherer was a man with enormous charisma and energy as a lec-turer. Henry set himself the task of combining those qualities.

Henry’s career at CCNY was a bit checkered. It is true that as a juniorHenry won the medal for the best senior psychologist at CCNY, bestingJulie Hochberg, fellow one-time physics major, now psychology major.But it is also true that Henry ran into trouble with botany and militaryscience. Henry reports that as an act of defiance he once ate the apple he

was supposed to dissect, an act that, one imagines, contributed to hispoor grade for the course. Henry’s botany has improved considerablyover the years under the influence of his (and Lila’s) passion for gar-dening. And, Henry reports, his military science too has improved; apassion for chess has helped that along. In any event Henry did get aB.S. and he took it to Berkeley.

At Berkeley, Henry worked with Edward Chace Tolman. Tolmanwas, of course, the cognitive behaviorist willing to do battle with ClarkHull on his own turf, in the animal laboratory. There Henry, characteris-tically, ignored the fluff and rhetoric that often surrounded debatesabout behaviorism to produce research asking a serious question: Exactlywhat does an animal learn when some behavior is reinforced? (SeeRescorla’s chapter, in which he describes this experiment, and his ownrefinement of it, in detail.) In any event, Henry’s stay at Berkeley wasbrief, a mere three years to his Ph.D. During that three-year period, inthe summer of 1947, Henry returned to New York to visit his family, andthere he taught introductory psychology for the first time in the sum-mer sessions of CCNY and Brooklyn College.

Meanwhile, Lila Lichtenberg was born on the Upper West Side ofManhattan on December 10, 1929, less than two months after the mar-ket crashed. As a consequence of the crash, the family moved to moremodest surroundings in Brooklyn, where Lila attended PS 153. Lila’s fa-ther was a self-employed structural steel detailer, and as times becamebetter, the Lichtenbergs moved to the “Casa del Ritz,” which Lila proudlyincluded as the return address on all of her correspondence. As anotherconsequence of financial solidity in the household, Lila attended sum-mer camp in Vermont. Camp Kee-Wa-Kee was the site of Lila’s firsthonor: She became a member of the Blue Dragons in 1939, an honor thatstill holds pride of place on her vitae.

Lila did not plan to attend college, and stubbornly refused to apply toany. Despite her profound efforts in this direction, she was accepted atboth Brooklyn College (through no fault of her own, her scores on amandatory exam had been sent in by her high school) and AntiochCollege (which was the compromise as it was work-study). She enteredAntioch in 1947. Her mother was delighted because, as she told Lila,“People who go to college can talk to anyone about any subject.”(Events certainly proved her mother right.) The program at Antioch re-quired that students devote half time to outside (nonacademic) work.Although Lila did attend classes, the more memorable parts of her col-lege career were her work. She held jobs as an occupational therapist atthe Delaware State Hospital for the Insane (where she attempted toteach amnesic patients to perform plays with marionettes), as a reporter(doing press releases for the European Reconstruction Agency, under

2 Introduction

the Marshall Plan, in Washington, D.C.), and as the editor of Antioch’scollege magazine, The Idiom (where she penned some of the more left-wing editorials).

After graduating, Lila and another writer for The Idiom (nicknamed“Hyphen” for his compound name) went to New York City as literarylights. Hyphen got a job as an assistant editor at Doubleday. As Hy-phen’s female equivalent at Antioch, Lila got the female equivalent ofhis job: She became a dictaphone operator. Lila reports that she only no-ticed the discrepancy in job assignments some years later; and at thetime, they were both very happy with their “jobs in publishing.” Fromthere, Lila moved up to “Gal Friday” at the Journal of the American Water-works Association, where she ran the journal.

One weekend, Lila joined some friends at Dartmouth, where she metEugene Galanter, an assistant professor of psychology at the Universityof Pennsylvania. That weekend, they decided to get married, and did sothe following Saturday. They moved into an apartment at 36th andSpruce, and Lila—as a faculty wife—began to take courses in Greek aspart of a program in classics.

At the time, the great Indo-European scholar Henry Hoenigswaldtaught courses in both linguistics and classics. Lila took Greek fromhim, and spent hours translating text. Hoenigswald recognized thatLila loved the parsing of Greek sentences most of all, and encouragedher to work with Zellig Harris in the linguistics department. Followinghis advice, she became a graduate student in linguistics, working onHarris’s grant, “Tranformation and Discourse Analysis Project” (TDAP).The central problem here was to understand the relationship betweensentences in a discourse, and in particular, how any given item movesthrough the discourse, for example, changing from “new” informationto “given.” The fact that a single item could change its function oversentences led to the problem of how one could—in principle—relate thedifferent occurrences of the item to each other over sentences. If “Bill”occurs in one sentence as subject and then shows up as object in another,how could these links be described? Harris’s idea was that each of therelevant sentences could somehow be related to a central, or “kernel”sentence—with “transformations” relating the kernel to each of its real-izations. Thus the TDA Project sought to relate sentences to each otherusing a mechanism that would—in some form—come to play a key rolein future advances in linguistics and psycholinguistics.

As part of her graduate training, Harris advised Lila to learn how towork the UNIVAC2 computer, which had been donated to the univer-sity by Univac in recognition of the contributions of Eckert and Mauchlyto the development of a general purpose digital computer. The UNIVACoccupied the entire first floor of the David Rittenhouse Laboratory. It

Introduction 3

also had astonishing computing power for the time, though consider-ably less than our current handheld calculators. Working on Harris’sproject was a group of brilliant graduate students (Lila Gleitman, BruriaKauffman, Carol Chomsky, and Naomi Sagar) as well as an engineer en-listed by Harris—Assistant Professor Aravind Joshi. Joshi was fasci-nated with the problem of how to use a computer (designed, after all,for number crunching) to understand language. Harris enlisted Joshi todevelop a parser—an automaton that would be able, to a limited de-gree, to comprehend the running text of a natural language usingHarris’s analytic methods of distributional analysis. In doing so, Joshibecame the first computational linguist. The parser that emerged fromthis project, Joshi recently reminded us, still rivals (or outperforms) thebest current parser in the field.

But independently, Lila was beginning to wonder whether distribu-tional analysis could really properly capture the organization of lan-guage or language learning. Noam Chomsky, a recent graduate of Pennwho had also studied under Harris, suggested that the entire enterprisewas doomed to failure, and he provided Lila with a copy of his recentbook to read. She read Syntactic Structures secretly, and clearly recog-nized how Chomsky’s approach reformulated problems in the organi-zation of language and language learning.

The Middle Years

On taking his Ph.D., Henry moved back to the East Coast to take up aposition as an assistant professor at what was then the Mecca in exile ofGestalt psychology, Swarthmore College. There, Henry joined a facultywith, inter alia, Solomon Asch, Wolfgang Kohler, and Hans Wallach(perhaps the person Henry admires most as an experimentalist). Henrystayed at Swarthmore for fifteen years. While there, he undertook manyprojects. For one, he worked with Solomon Asch on his famous studiesof conformity and independence. For another, he began work with twograduate students (Ulrich Neisser and Jacob Nachmias) on what theysaw as a three-part series, a series that they envisioned as the definitivedeconstruction of Hullian learning theory (see Nachmias’s chapter, thisvolume, for more on this topic). And Henry began to grow famous forhis learning theory seminars—evening seminars, of course. (AlthoughHenry, like his mentor, believes in the possibility of serious cognitive ac-tivity in the minds of rats, he certainly does not believe in the possibilityof serious mental activity before noon.) These seminars began in theevening but they ended whenever the topic was exhausted, be thatthree, four, or five hours after they started. They certainly did not endwhen Henry was exhausted—Henry does not become exhausted while

4 Introduction

engaged by psychology, poker, or God knows, the theater (see Nach-mias and Rescorla on these Swarthmore learning seminars). Henrybegan something else while at Swarthmore—Psychology, the book.

At Swarthmore, Henry made contact with Don Lamm. Don wasHenry’s first editor. But before Henry could finish the book—someeighteen years later—Don became President of W. W. Norton. Still,Lamm remained an über editor for the book, always available for goodcounsel. (See the Lamm chapter for how this relationship was estab-lished and how the book came to be.)

Also at Swarthmore, Henry met Lila. (And, for those who have readrecent work by the Gleitmans, Lila met Henry. They met each other.)Through Gene Galanter (whom Lila had divorced after one year), Lilahad become acquainted with a number of faculty in psychology atPenn, and there was a close connection between the Penn and Swarth-more psychology faculties. Henry immediately fell in love with the ele-gant and brilliant Lila, and they were married.

In 1963, Henry and Lila left Swarthmore—not for Penn, but for Cornell.Although this might have been disruptive to Lila’s graduate training,the timing was not all bad, for Lila had begun to leave the Harris foldand had just written her intended thesis—a transformational analysis ofconjunction entitled “Coordinating Conjunctions in English” (pub-lished two years later, in 1965). Already a distinguished linguist but notyet a Ph.D., she accompanied Henry to upstate New York.

At Cornell, Henry continued his investigations of memory. As Henrytells it, he mainly investigated whether a person could remember whatthe sun looked like without seeing it for an entire academic year. Ithaca,as it turned out, was not Henry and Lila’s cup of tea: The final momentinvolved a horse and their bedroom window. Henry was willing, there-fore, to return to Philadelphia to become professor of psychology atPenn, as well as the chairman of the department of psychology. He wouldalso, of course, continue writing The Book. The Gleitmans returned toPenn in 1965.

These were exciting times, generally and specifically, at Penn psy-chology. In the late 1950s, the trustees and provost had decided to re-vamp the School of Arts and Sciences. In part because of the influence ofProvost David Goddard, the first department to be revamped was psy-chology. A physicist, Robert Bush, was brought in as chair to do the re-vamping. Under his leadership, a revolution, a bloody one from somereports, was wrought at Penn in the early 1960s. Bush hired many lumi-naries—Dottie Jameson and Leo Hurvich, Duncan Luce, Jack Nach-mias, Dick Solomon, and Philip Teitelbaum, among others. But as hecame to the end of his term as chair there was a need for an appointmentthat would combine research excellence, administrative acumen, and

Introduction 5

brilliant teaching of undergraduates. Thus was Henry called fromIthaca to become the chairman whose task it was to solidify the revolu-tion. It was Henry, then, who stabilized the department, who gave it itstraditions, and who made it into the place that Henry and Lila’s gradu-ate students (who have written most of the essays in this volume)would later find.

During this time, Lila and Henry had two daughters, Ellen and Claire.A scholar of learning, Henry was quite sure that babies had no capaci-ties other than eating, sleeping, and crying. But Lila, a scholar of lan-guage, recognized in her young children something quite remarkable.They learned language rapidly, with no obvious effort, and in the frankabsence of any explicit tutoring. Her close friend, Elizabeth Shipley, wasalso a young mother, as well as a psychologist trained under DuncanLuce. And she agreed with Lila: There was something quite remarkableabout language learning in infants.

Lila and Liz began to systematically study their children’s languagelearning (see chapter by Shipley, this volume), and with Carlotta Smith,won their first grant to support the work. Their first publication on thetopic was “A study in the acquisition of language: Free responses tocommands” (Shipley, Smith, and Gleitman 1969). The paper, seminal inits theoretical and experimental sophistication, laid out many of the keyissues that frame research in language learning today. The experimentitself was among the first to set up conditions for tapping into children’sknowledge of language without relying on spontaneous speech to doso. The method and findings—which revealed greater competence thanwas evident from children’s spontaneous production—were argued toprovide a basis for understanding the organization of language inyoung children. The theoretical context contrasted strong nativist posi-tions of that time such as Chomsky’s with strong empiricist positionsthen held by Bloomfieldian linguists and psychologists (see Newport’schapter, this volume, for more on this contrast). Although the authorsleaned toward a nativist stance, they firmly argued that considerablymore empirical evidence was needed before understanding the preciseinteractions between the child’s innate endowment—whether specificknowledge or data-manipulating tendencies or both—and the learningenvironment that was available to the child.

At around the same time, Lila decided to finish her Ph.D. Her friendand mentor, Henry Hoenigswald, convinced her to do her dissertationunder Henry Hiz in the linguistics department. She had already pub-lished the conjunction paper (her intended thesis), but was summarilytold that this would not “count” for a thesis because it had already beenpublished. She turned to study the structure of compound nouns,which she investigated from a theoretical perspective (doing formal lin-

6 Introduction

guistic analyses to characterize the nature of these compounds) andfrom a psychological perspective (eliciting people’s intuitions about thedifferences in meaning between pairs such as “black bird-house” and“black-bird house” in order to discover the linguistic structure). (As Lilahas taught her students, these two approaches are two sides of the samecoin.) When she presented these ideas to her committee in linguistics,she was told that the formal linguistic analyses would make a fine dis-sertation; but that she would not be able to present the results of her ex-periments on intuitions, because that was psychology. So Lila presentedthe intuitions in a format more congenial to the linguist’s ear and eye:sprinkled throughout the text and labeled sequentially with numbers.Thus she received her Ph.D. in linguistics from Penn in 1967.

Lila took a position as an associate professor of linguistics at Swarth-more College in 1968, and stayed there for four years, serving asSwarthmore’s linguistics department and teaching an entire generationof budding linguists and psycholinguists, including Lance Rips, GaryDell, John Goldsmith, Muffy Siegal, Emily Bushnell, Robert May, andElissa Newport (who, though a graduate student at Penn, drove toSwarthmore three times a week to learn linguistics by taking all of Lila’scourses). This second tour of duty at Swarthmore was more rewardingthan the first (as faculty wife), especially since this time she did not haveto wear white gloves to Sunday tea.

The Modern Era

The beginnings of The Modern Era are marked by two events: the pub-lication of Lila and Henry’s first collaborative effort, and the initiationand subsequent flourishing of “The Seminar.”

The first collaborative effort was in part a result of the linguistics de-partment’s dictum to Lila: No experiments in the thesis. But Lila had, infact, conducted experiments, and the data appeared to hold rich infor-mation about the organization of compounds. As Henry and Lila dis-cussed and debated some of the results, new questions arose, togetherwith elegant experimental designs and data analyses that were (andstill are) Henry’s hallmark. The result was their first joint publication,Phrase and Paraphrase (1970).

The developing seminar was a natural outcome of such collaboration.When Henry came to Penn in 1964 there was an ongoing seminar onmemory. But by 1970 that seminar had become Henry’s research semi-nar, and by 1972 Lila had joined it and it became Henry and Lila’s research seminar dedicated to educating graduate students (and first at-tended by Heidi Feldman, Sandra Geer, Susan Goldin-Meadow, JohnJonides, Peter Jusczyk, Deborah MacMillan, Elissa Newport, Marilyn

Introduction 7

Shatz, and Liz Shipley). Throughout the first decade of this seminar,Henry was, of course, still writing The Book—a process that set thestandard for all students who would later be engaged in scholarly writ-ing. Some time during the late 1980s, the seminar became the “CheeseSeminar” (since all along, various gourmet cheeses accompanied thediscussions of research), and by the 1990s, simply “Cheese.” It was atthese evening seminars that many of the contributors to this volumelearned how to do psychological research, learned to love psychology,and learned to love triple cremes. (They did not need to learn to loveHenry and Lila, since that is innate.) It was during the seminar that stu-dents presented budding ideas (always made into a real idea by Henry),possible experimental designs, theoretical puzzles, and job talks.Critically, it was also during these seminars that students learned to aska good question, and to know what a good answer might be—even ifthey did not have that answer. Students of Henry and Lila consider theseminar to represent the core of their training, as many of the chaptersattest.

The learning at these evening seminars has not just been for psychol-ogists, since students in the School of Education, the linguistics depart-ment, and, lately, the computer science department have also beenwelcome. During most years, the research seminar has been a joint pro-ject of Henry and Lila, but Henry has not always participated. Illnessand the demands of directing a play have sometimes kept Henry away.In those years the seminar has often been a Liz Shipley and Lila Gleit-man course.

Collaborative effort in research has always been completely naturalto Henry and Lila, and it continued during the 1970s, largely as a conse-quence of the seminar. During the 1970s, both Henry and Lila collabo-rated with Paul Rozin, a young psychobiologist in the Penn psychologydepartment (see the chapter by Rozin, this volume). For example, in1971, Henry published a study with Paul Rozin on goldfish memory asa function of temperature during the retention interval. Yes, goldfishmemory as a function of temperature during the retention interval.

A perpetually pressing issue has been whether forgetting is a matterof the mere passage of time, or is instead the result of interference fromthe intervening events that time allows. The trouble is, of course, thattime and events are closely correlated. Seemingly, what is needed to an-swer this question is a time machine—a device that can make time gofaster or slower while keeping the events that occur constant. Time ma-chines are hard to find; but Henry and Paul realized that biochemicalprocesses are a function of time and temperature. So if two organismsare at different temperatures for the same interval of time, from a bio-chemical point of view, this is equivalent to time’s moving faster for the

8 Introduction

hotter organism, but experiencing the same events. This experimentcannot be done with humans; changing external temperature for warm-blooded animals does not change their metabolism. But cold-bloodedanimals, like goldfish, can readily be warmed or cooled by changingtheir water temperature. So Rozin and Gleitman, lacking a time ma-chine, heated and chilled their fish. Clever!

Henry recently summarized the results for us: “If you want to be agoldfish who remembers, spend the retention interval in a refrigerator.”

The results of this first study were, according to Henry, quite strong,and were published in a scholarly paper. Henry’s hand was evident inan elegant and notable control that evaluated the possibility that failureamong the heated fish was due to their brains’ boiling. In this control,fish were heated at 90 degrees for 60 days and were then given threedays to readjust. These fish did as well as those who who learned at acool 60 degrees. But an attempted replication failed: Some of the fishsuffered from the Ick, and were unsuccessfully treated for the diseaseby a research assistant. The dead fish showed no hint of rememberinganything.

Lila, meanwhile, had moved from Swarthmore to Penn, as WilliamCarter Professor in the Graduate School of Education. There, she collab-orated with Paul Rozin on studies of reading—specifically, developingthe idea that an orthography based on syllables might be more “accessi-ble,” hence easier to learn, than an orthography such as the alphabet,which requires a highly abstract mapping from sound to individual let-ter. Their first publication, in 1973, posed the problem of reading as aproblem of mapping (or unmapping) orthography to a psychologicallyappropriate level of phonological representation. This inflamed manyeducators, who were entrenched in existing methods of teaching read-ing “for meaning,” which Lila and Paul pointed out was the equivalentof claiming a method of teaching driving “for destination.” Gleitmanand Rozin (and Rozin and Gleitman) went on to publish two landmarktheoretical and experimental papers documenting the logic of their ap-proach, using the history of writing systems and the psycholinguisticsof sound processing as supports.

But the bulk of collaboration was done with graduate students. In theearly 1970s, Henry and John Jonides worked on mechanisms of itemrecognition. John was the “senior student” in the seminar, the only per-son besides Henry who had his own clearly designated chair at theevening meetings. Henry and John were intrigued by the followingissue: It had been repeatedly demonstrated (by Ulric Neisser amongothers) that the visual features distinguishing one item from anotherhave a profound influence on people’s speed and accuracy of recogni-tion. The paradigm that became popular to explore these issues was

Introduction 9

visual search: requiring subjects to search for some prespecified targetitem (say, the letter X) among an array of other items (say, other letters).Using this paradigm, even the most casual experimentalist could findthat when the target item was physically different from its backgrounddistractors (say, an X among Os and Cs), search was faster and more ac-curate than when they were similar (say, an X among Ys and Ks). Thisresult, together with much additional evidence, has been amassed toargue for a featural theory of visual recognition, a theory that is still theleading contender today.

Going beyond these results, Henry and John explored the possibilitythat it was not just physical featural differences between characters thatmight influence visual recognition. In addition, they hypothesized, cat-egorial membership might be a distinguishing characteristic. Buildingon a previous result, they confirmed that visual search for a digit em-bedded among letters was faster and more accurate than for a letteramong letters (of course, this experiment could not have passedHenry’s muster had it not been exhaustively counterbalanced within abreath of its life for which items were targets and distractors). They thenwent on to show that it was not a physical difference between membersof the category “letter” or “digit” that differentiated them: A compara-ble pattern of results obtained when the target, the item “0,” was de-scribed to subjects as the letter “O” or the digit “0.”

Beyond this, Henry and John went on to explore the intricacies of vi-sual search based on categorial difference in several papers that fol-lowed. All of this earned them a reputation for having cornered themarket on alphanumeric stimuli, and it earned John his first academicjob.

During these years, Lissa Newport also began to work with both Lilaand Henry. She met Lila, as mentioned above, when Lila taught linguis-tics at Swarthmore: At the time, Penn did not teach generative linguis-tics, so Lissa commuted to Swarthmore to acquire appropriate trainingin linguistics for beginning work with Paul Rozin and Oscar Marin onaphasia. Soon after beginning this arrangement, however, Lissa sepa-rated from her then-husband and became an early member of the Gleit-man Hotel for Wayward Academics. Though she had at the time nevermet Henry (who was on sabbatical during her first year in graduateschool) and knew Lila only from classes, she was warmly invited to stayat the Gleitmans’ while she searched for a new place to live. Like manywho followed her, she found refuge in the Gleitman kitchen and livingroom, and, through countless hours of warm conversation, was nur-tured back from thoughts of quitting school to debates on the structureof the mind and nativist approaches to learning.

10 Introduction

Lissa also read the latest Gleitman, Gleitman, and Shipley grant pro-posal and grew interested in their discussion of approaching the nature-nurture question in language acquisition by studying mothers’ speechto children. The grant proposal suggested that perhaps mothers shapedtheir speech to children in accord with children’s abilities to compre-hend that speech, a suggestion also raised in Shipley, Smith, and Gleit-man (1969). If true, they went on, speech to children might providemore well-structured input for language learning than usually be-lieved, and this in turn might change our views of the extent and char-acter of innate knowledge required for acquisition.

With this possibility in mind, Lissa, Lila, and Henry began in 1972 tocollaborate on a study of fifteen mothers interacting with their youngchildren. Henry, always good with nicknames, christened this interestMotherese, and this became the term used widely in the field for speechto children (until the earnest 1980s and 1990s ruined a good phrase byturning it into the more politically correct caregiver speech).

The work progressed slowly: During the first six months, the seminarmembers heard Lissa reporting, on a weekly basis, “Still transcribing.”But the real problems had to do with conceptualizing the problem ofhow maternal input could help children learn language—beyond theobvious fact that it provided the model of the child’s native language.During this period, in the early 1970s, a number of dissertations had ap-peared on mothers’ speech to children, and all of these had shown thatMotherese exhibited sentences that were short and overwhelminglygrammatical. On the basis of these facts, many people in the field hadconcluded that these characteristics meant that Motherese was “simpleinput” to the child, and by inference, that this simple input must helpsolve the learning problem. But Lissa’s early analyses kept looking dif-ferent. While Motherese sentences were indeed shorter, the grammarrequired to describe them was not particularly simple: Mothers used awide range of sentence types to their children, including yes-no andwh-questions as well as imperatives, which required a full range of rathercomplex syntactic transformations to generate. By contrast, mothers’speech to an adult (Lissa) consisted almost entirely of simple active af-firmative declarative sentences, which were the kernel (and relativelyless complex) sentences of a transformational grammar. From a gram-matical point of view, then, Motherese was not so simple and did notappear to present a new solution to the language acquisition problem.These findings, and a discussion of their significance for acquisition the-ory, later became Lissa’s dissertation.

But Lissa, Henry, and Lila continued on together to ask a more im-portant question: How could one go beyond describing Motherese and

Introduction 11

find evidence about whether, in fact, it produced any changes in acqui-sition? As usual, this drew on Henry’s remarkable skills in analysis anddesign, particularly needed in this case because one couldn’t easilybring this problem into the lab and conduct an ordinary psychologicalexperiment.

They took two approaches to the question. First, they conducted anexperiment asking whether mothers’ frequent tendency to repeat them-selves, producing strings of related sentences, might help children toanalyze the grammar. Lissa and Henry designed a repetition experi-ment with a clever analysis to distinguish the benefits of merely havingmore opportunities to respond from potential benefits to learning overthe repeated presentations. The results: no learning. It began to dawnon them that perhaps these negative results were not so negative afterall: Perhaps Motherese did not change the problem of acquisition insuch a clear way.

The second line of work asked whether the individual differencesthat occured among the fifteen mothers would correlate with differ-ences in their children’s acquisition success over subsequent months.Since the study was originally designed to ask a different question (howdid mothers speak to children of different ages and linguistic abilities?),the differences between the children in initial linguistic ability were re-moved by performing double partial correlations. As the millennium ishere, it may be hard for younger readers to appreciate what this meant.Double partial correlations were first performed by using punch cardson a mainframe computer; later, the situation improved substantiallyby using an extremely expensive hand-held calculator that could add,subtract, multiply, divide—and had one memory, which enabled it tocompute simple correlations, which could then in turn be combined toproduce double partial correlations.

To Henry, Lissa, and Lila’s surprise, the results did not show overallrelationships between aspects of maternal speech and their children’slearning: Those mothers who produced the most simplified Motheresedid not produce the most gifted learners. However, there were a num-ber of other significant correlations, which Lila and Lissa quickly real-ized fit into a very different conceptualization—one that also had beensuggested by Shipley, Smith, and Gleitman. First, the acquisition of nounsand verbs—the argument structure and open class words for them—did not show effects of variation in maternal input. These seemed toemerge on their own timetable, without strong relations to the details ofmaternal input. But second, the acquisition of closed class elements,such as verbal auxiliaries—those grammatical elements that vary mostacross languages—did correlate with aspects of maternal speech. Critic-ally, the aspects of maternal speech that most strongly correlated with

12 Introduction

the learning of verbal auxiliaries were the positions in which theseitems appeared in Motherese. In Shipley, Smith, and Gleitman, childrenattended most prominently to certain parts of input sentences, espe-cially sentence beginnings. In accord with this, the Newport, Gleitman,and Gleitman results showed that mothers who produced the most aux-iliaries at sentence beginnings—by using more yes-no questions andfewer imperatives—had children whose own sentences were decoratedby verbal auxiliaries the earliest. In short, Motherese simplification didnot appear to change the character of language learning. Instead, the re-sults turned the motivating force in acquisition back to the predisposi-tions of the child.

In many quarters, this was not a popular claim. As Lila has pointedout in recent years, American psychology does not find comfort in na-tivism. Apparently, as she has succinctly put it, “Empiricism is innate.”But the findings did help to begin a line of research, by the Gleitmansand their collaborators and students, that systematically investigatednatural variations in the variables contributing to acquisition. The out-comes of this research consistently revealed that the mind of the child—and only secondarily her input from the external world—formed themost substantial contributions to the acquisition process. In contrast tothe Newport, Gleitman, and Gleitman studies, these subsequent lines ofwork wisely examined more extreme variations in input or in internalvariations than had been measured in the double partial correlations ofthat first work.

Newport, Gleitman, and Gleitman’s (1977) work produced the firstarticle published in collaboration with students in the research seminar.Also ongoing during this period was work by the Gleitmans with HeidiFeldman and Susan Goldin-Meadow on the creation of language bycongenitally deaf children; and a paper by Feldman, Goldin-Meadow,and Gleitman appeared in 1978. The papers with Jonides, with New-port, and with Feldman and Goldin-Meadow mark the coming-of-ageof the first generation of research seminar participants. They also markLila’s first work on understanding language acquisition by looking atchildren for whom the usual inputs to language acquisition are ab-sent—a theme that threads the work of seminar participants over sev-eral generations (see the chapters in this volume by Newport; Goldin-Meadow; Landau).

By 1978, Henry was still, of course, writing The Book. And the semi-nar had grown to include about a dozen students from psychology andfrom education. Although Lila’s original appointment was in the Schoolof Education, in 1979 she emigrated to the departments of psychologyand linguistics in the School of Arts and Sciences, where she later be-came the Marcia and Steven Roth Professor. A new faculty member at

Introduction 13

Penn—Liz Spelke—also joined the seminar, with her students PhilKellman and Hillary Schmidt, both interested in infant perception. Theseminar now also included other students who did not work directlywith either Lila or Henry (such as Dan Riesberg and Phil Kellman; seetheir chapters) but who saw the seminar as a critical event in their grad-uate education. Also present were students working mostly with Henry(Judy Foard; Jerry Parrott; see Parrott’s chapter), and those workingmostly with Lila (including a number of the students from education,such as Julia Dutton, Barbara Freed, Pam Freyd, Kathy Hirsh-Pasek,and George Meck).

Parrott’s work with Henry concerned a topic dear to Henry’s heart:the nature of a special group of emotions that included humor, playful-ness, curiosity, and the appreciation of beauty. Henry was deeply inter-ested in these emotions as they were, after all, what he regularly calledup in his capacity as director when he—every several years—directed aplay at Penn or elsewhere. Henry and Jerry hoped that experiments onthe role of expectation and surprise in simple humor might help themunderstand something about the more complex types of expectationand surprise found in comedy and drama. This led them to study adults’surprise and humor responses to unexpected elements of animatedevents. Discovering that Penn undergraduates’ responses were rathercomplex, however, Jerry and Henry turned to infants—who they as-sumed would, at least, constitute a “simpler preparation.” This devel-opmental work fit well within the seminar group at that time, for manyof the students were doing research on language learning, especially asframed by the deprivation paradigm. This included the acquisition oflanguage by deaf children and second language learning by adults. Twoother students in psychology were also intrigued with the deprivationparadigm, and eventually came to study language learning in childrenwith Down’s Syndrome (Anne Fowler) and children who were bornblind (Barbara Landau).

Barbara Landau came to graduate school particularly intrigued withthe recently published work by Feldman, Goldin-Meadow, and Gleit-man on the creation of language by deaf isolates—a case that providedstunning evidence for the emergence of a structured linguistic systemwith no formal linguistic input (see Goldin-Meadow’s chapter, this vol-ume). Having relatively little background in linguistics at the time,Barbara naively believed that the creation of a language in the absenceof linguistic input was not surprising, and could perhaps be explainedby considering the rich perceptual and conceptual environment inwhich the deaf children developed. Specifically, observations of objectsand events might allow the children to construct a semantic system thatwould be the foundation for language. Barbara and Lila had lengthy dis-

14 Introduction

cussions on the issue of whether one could, in fact, construct a languagewithout so observing the world; and these discussions led to the ques-tion of how the blind child could learn a language, given that she wouldhave to construct the semantics of a language in the absence of rich per-ceptual information afforded by the visual system. Lila cautiouslyagreed that this would be an interesting topic to pursue, and the two setout to study language learning in the blind child.

Something curious happened, however. Barbara had recruited threeblind children each around eighteen months, but none of them was pro-ducing much language. However, what they were doing was equallyfascinating: exploring and recognizing people and objects around them,navigating through their environments without hesitation, and inter-acting with the world in a way that suggested that they were construct-ing a rich spatial world. How could this occur?

Henry was immediately taken with this question, as was Liz Spelke,and with Barbara, they began to study the emergence of spatial knowl-edge in the blind child. Under the close guidance of Henry and Liz,Barbara’s informal observations of the capacities of the blind child werequickly turned into elegant experiments demonstrating the capacity ofthe blind child to learn spatial routes and make spatial inferences totravel along new routes in novel environments. These studies culmi-nated in several publications documenting spatial knowledge in a veryyoung child who was blind from birth; these papers reflected the fur-ther influence of Randy Gallistel, who guided the authors to under-stand this knowledge as a geometric system that emerged early in lifewith or without visual experience.

The existence of such a spatial knowledge helped Lila and Barbarapartly explain how the blind children learned language (as they did, be-fore long): If the blind child possesses spatial representations, then thesecould provide the foundation for the development of semantic repre-sentations of objects, locations, motions, etc. But perhaps less easily ex-plained were the striking observations that they made regarding aspecial portion of the blind child’s vocabulary: Among the earliest andmost productive words in the blind child’s vocabulary were words suchas look, see, and color terms. Intensive experimental investigation showedthat the semantic structure of these terms as used by the blind child wasquite rich and in many ways quite similar to those of sighted children.For example, the blind child used the term see for her own activity of ex-ploring with the hands (though she used the term for others to refer tovisual exploration).

It was the theoretical consideration of these phenomena, however,that led Lila and Barbara deeply into questions about the nature of con-cepts and meanings, and the mechanisms by which they could be

Introduction 15

learned by children. On the issue of concepts, Lila and Henry were alsoactively collaborating at the time with a postdoctoral member of the re-search seminar, Sharon Armstrong. Their question concerned the repre-sentation of everyday lexical concepts, such as “dog,” “apple,” etc.During the late 1970s and early 1980s, the field had been heavily influ-enced by the work of Eleanor Rosch and colleagues, who had arguedthat everyday lexical concepts have “prototype” structure. Rosch’sfindings confirmed the observation that people typically can judgemembers of a category as to their representativeness in the category, ortheir “goodness.” For example, robins are rated as “better” exemplarsof birds than penguins, and these differences across category membersseem to have reflexes in processing time as well as explicit judgments.

For Armstrong and the Gleitmans, however, the evidence did not log-ically prove that prototype representations were any more psychologi-cally real than were representations including the necessary andsufficient conditions for membership. They set out to disprove the argu-ments on logical as well as empirical grounds (and provided as well adevastating critique of featural theories in general). In 1983, they pub-lished “What some concepts might not be,” whose burden was to showthat the sort of evidence that had been collected to show that the con-cept “vegetable,” say, has prototype structure (rather than having theclassical structure of necessary and sufficient conditions), could also becollected for the concept “even number” (a clear example of a conceptwith a classical definition). Subjects are indeed willing to judge brusselssprouts to be worse examples of “vegetable” than carrots, but they arealso willing to judge 24 to be a less good example of “even number”than 4. If such evidence, the paper argues, is enough to convince youthat the mental representation of vegetable is prototypical rather thanclassical, it should convince you of the same for “even number”! Thusmembers of the Gleitman seminar, with their leaders, wrung theirhands over the prospects of characterizing everyday lexical items interms of features—prototype or not.

But on another front, Landau and Gleitman were still considering thepuzzle of how the blind child came to learning the meanings of visualverbs such as “look” and “see.” Acknowledging that they would proba-bly never be able to fully characterize the “meanings” of these terms,Lila and Barbara turned to a somewhat different question: How couldaspects of the meanings of a word be learned if the putative conceptualfoundation for the meaning was absent in the experience of the child?Theories of semantic acquisition typically assume that semantics issomehow transparently accessible to young children, through percep-tual and motoric interactions with the environment. For the case of aword such as “look,” the obvious mechanism for learning would be to

16 Introduction

link the hearing of the word with the experience of looking. But, pre-suming that knowing the meaning of “look” involves some experienceof actually looking, how could the congenitally blind child ever learn itsmeaning?

The theoretical analysis of this problem took several steps. The firstwas to consider the obvious: The child could have learned the hapticmeaning of “see” or “look” because her mother used it only in thosecontexts in which seeing or looking was possible—that is, when thechild was near some object. However, analysis of the contexts in whichthe mother uttered these words to the child revealed that such simplecontextual factors did not distribute themselves in a way that wouldhave allowed the child to group together the visual verbs as distinctfrom other verbs in the corpus. The next step was to consider the lessobvious: that the syntactic contexts in which the verbs appeared couldhave provided the child with additional information about their mean-ing. This analysis suggested that the joint use of syntactic and contex-tual information could indeed result in separation of the visual verbsfrom all others (as well as group other semantically related verbs to-gether coherently).

To this point, the idea that learners could use syntax to discover se-mantics was anathema in the field, which typically regarded the majormechanism of learning to occur in the opposite direction (i.e., the learneruses semantic categories to project syntactic categories). However, theidea was not novel to Lila: Indeed, its foundation could be seen as aproduct of her training under Harris. The central idea there was thatdistributional analysis was a powerful tool for deriving regularities inlanguage. In the case of verbs, the meaning of an individual verb shouldbe derivable from its distribution across all of its syntactic frames—anidea that was evident in Lila’s “Great Verb Game.”

In this parlor game, people were given a sentence containing onenovel verb and were asked to guess the meaning of that verb. With onesentence, this was quite difficult; but as additional new sentences (withthe same nonsense verb) were provided, it became easier and easier,until there was just one answer. The reason, theoretically, is that eachverb participates in a unique set of syntactic contexts; hence analyzingthe distribution of a novel verb across such contexts should yield aunique answer.

The general idea that syntax could aid in discovering the meaning ofa verb made a great deal of sense to Landau and Gleitman, who usedthe idea of such distributional analysis to evaluate whether the sen-tences that the blind child actually heard could provide the basis for in-ferring the meanings of the verbs therein. This analysis resulted in aspecific theory of how the blind child could acquire verbs such as

Introduction 17

“see”—and, by extension, how any child could acquire verbs such as“think,” “know,” and “believe,” whose immediate perceptual interpre-tation was not obvious. Although the idea of distributional analyses re-mained central to their theory, an important insight gained by thisstudy was that groups of semantically related verbs shared sets of syn-tactic frames. This meant that a distributional analysis of frames couldyield a “ballpark” semantic interpretation—that the verb to be learnedwas a verb of perception, of cognition, or a motion verb, for instance.These ideas were published by Landau and Gleitman in Language andExperience.

The theoretical idea of using syntactic context as a mechanism of verblearning—later dubbed “syntactic bootstrapping”—gave rise to numer-ous empirical predictions, which have recently been investigated byLila, Henry, and their students (and are reviewed in a 1997 paper by Lilaand Henry in Lingua). For example, can young children actually usesyntactic context to interpret novel verbs, and can they use these con-texts to infer new meanings for old verbs? (Yes, they can, as shown invarious papers written by the Gleitmans in collaboration with LettyNaigles, Cynthia Fisher, Susan Rakowitz, and Geoff Hall.) Do adultsrepresent verb meanings in such a way that one can recover semanticstructure from the syntactic contexts in which they occur? (Yes, theycan, as shown by Fisher with the Gleitmans.) Is the (nonlinguistic) con-textual environment for verb learning really such an unreliable basefrom which to infer meanings? (Yes. It is virtually impossible to predictthe presence of any specific verb from the nonlinguistic context sur-rounding its use; however, such context provides excellent informationfor predicting the presence of a specific noun; as shown by Gillette withthe Gleitmans.)

Honors and Awards

Finally, there are honors and awards. Henry has won all of the teachingawards for which he is eligible—all those that have to do with teachingpsychology (he has not to our knowledge won any awards for theteaching of chemistry or microbiology). He has won the University ofPennsylvania’s award (the Lindback), the School of Arts and Sciencesaward (the Abrams), and the American Psychological Association’saward (The Foundation Award). Had he not won them, it would havecalled their legitimacy into question. He is a Fellow of the Society ofExperimental Psychologists. And he has been the president of the twodivisions of APA he would have wanted to have been the president of,namely, Division 1 (General Psychology), and Division 10 (Psychologyof the Arts). It is not coincidental that these two divisions represent

18 Introduction

Henry’s brilliance as a teacher of psychology on the one hand and as adirector of plays on the other.

Lila has served as president of the Linguistic Society of America, is afellow of the Society of Experimental Psychologists and the AmericanAssociation for the Advancement of Science, and was recently namedthe Fyssen Foundation Laureate (the equivalent of the Nobel laureatefor language and cognition). She is currently codirector (with AravindJoshi) of the Institute for Research in Cognitive Science at Penn, whoseorigins date to the late 1970s, when the Sloan Foundation decided tostimulate development of the emerging field of cognitive sciencethrough Penn’s interdisciplinary faculty. The institute is a Science andTechnology Center of the National Science Foundation; it is the onlyScience and Technology Center grantee of the NSF in cognitive science,and it continues to have computerized parsing of natural language asone focus. So Zellig Harris’s vision in the late 1950s continues to be in-fluential through Lila, among others. Lila is also the editor of AnInvitation to Cognitive Science (volume 1), the first attempt to pull to-gether the various strands of research that constitute cognitive science.

These awards and honors reflect, after all, what Henry and Lila arecommitted to, and what they have successfully fostered in an entiregeneration of students: excellence in teaching and excellence in re-search. In turn, these reflect the single deepest common thread woventhrough their careers: Teaching and research, mentoring and collaborat-ing have always been bound together as the foundation from which allelse follows. For those of us who have benefited from this foundation,Henry and Lila’s intellectual presence has been an inspiration, and theirpersonal presence has changed our lives.

Postscript

Henry did finish The Book, with its first publication in 1981. Psychologyappeared in its fifth edition in the fall of 1998.

Introduction 19

               

 

Chapter 1

Der Urgleit

Jacob Nachmias

We are gathered here to reflect on the contributions of Henry and LilaGleitman to education at Penn. You have already heard much—and willhear more—about these very considerable contributions from studentsand colleagues who have had the good fortune to work closely with theGleitmans in recent decades. I think that the best way for me to make anonredundant contribution to these proceedings is to capitalize on thefact that, with one probable exception, I have known Henry Gleitmanfar longer than has anyone else in this room. Thus my remarks are alongthe lines of a memoir of the early years of Henry Gleitman’s academiccareer, the Swarthmore period. And in the spirit of those early years, Iwill entitle my talk “Der Urgleit.”

In 1946, Henry went to Berkeley to start his graduate studies. Hecompleted them in 1949, and in the same year, joined the faculty ofSwarthmore College. Anyone who has spent even a day in Berkeley willappreciate the strength of character exhibited by our hero when heopted to leave that charmed city across the bay from San Francisco afteronly three years in order to take a teaching job in the Delaware Valley.So when I first met him a year later, in the fall of 1950, he could not havegiven more than a couple of the 100-odd psychology 1 courses he hastaught to date. I had gone to Swarthmore to study at the feet of the thendemigods of perception—Hans Wallach and Wolfgang Köhler. And Idid just that, but, as it turned out, I actually spent far more time andprobably learned vastly more from two other individuals I had neverheard of before: one was Dick Neisser, my fellow Master’s student, andthe other was Henry Gleitman. Actually, it was much easier to study atthe feet of Henry Gleitman than most people: You did not even have tosit on the floor to do that, for Henry in those days was fond of perchingon any horizontal surface, particularly a radiator cover.

Henry was a phenomenon at Swarthmore in those days. With veryfew exceptions the Swarthmore faculty were solid, sensible, and seri-ous—as befits the faculty of a college with strong Quaker traditions. SoHenry could best be described as a sort of blue jay among brown owls:

He was vastly more colorful and louder. He was full of life, vitality, andmany talents: He acted, he directed plays, he sang outrageous Germantranslations of American ballads like “Frankie and Johnnie,” he was agourmet cook, he was an excellent cartoonist. But above all he taughtand he taught brilliantly. I don’t believe I actually heard him lecture atthat time, but I did sit in on two of his honors learning seminars. Theywere without a doubt the most intensive, exhilarating, and exhaustingintellectual experiences of my life; nothing before at Cornell or since atCambridge, Harvard, or Penn came close to them. Each session of theseminars started right after dinner, and went on well into the night, last-ing a good four or five hours. In those seminars we studied the writingsof the great learning theorists of the era—Hull, Tolman, Guthrie, andtheir disciples. The word studied does not begin to capture the flavor ofwhat we actually did. We read and reread, we analyzed, we dissected,we uncovered contradictions unsuspected by the original authors—orprobably by anyone else in the entire galaxy. We designed crucial exper-iments, some gedanken, some involving complicated, balanced designs,requiring armies of rats, to be run on ingenious runways or alleys orSkinner boxes. This was serious business: We really wanted to get to thebottom of things. There were no shortcuts, no time limits, no hand wav-ing. But it was also a lot of fun, with lots of laughter, and puns, and ban-ter, and food, and drink, and above all, camaraderie.

One of the outgrowths of those famous Gleitman learning seminarswas a Psychology Review paper, “The S-R Reinforcement Theory ofExtinction” by Gleitman, Nachmias, and Neisser. It was to be the first ofa series of papers intended to take apart the entire edifice of Hullianlearning theory, postulate by postulate. While we were working on theextinction paper, word reached us that galley proofs of Hull’s latestbook—A Behavior System—were available at Yale. The senior author ofthe GNN1 paper, as we called it, dispatched the two junior authors tolook through the galleys to make sure that the latest version of Hulliantheory was still subject to the criticisms we were making. Neisser and Itraveled to New Haven by a mode of transportation alas no longeravailable to impoverished graduate students, namely, the thumb. Whenwe got there, we discovered to our relief that the new book did not re-quire us to change a line of our critique.

Five years after I left Swarthmore, I returned as an instructor, andHenry and I were now faculty colleagues. But he was still very muchmy teacher. When I organized my first learning seminar, the memory—as well as the extensive reading lists (updated)—of those legendaryseminars led by Henry were my constant guides. But the most impor-tant thing I learned from Henry in that period was how to lecture—askill that alas, I seem to have lost in recent years. I learned by coteaching

24 Jack Nachmias

psychology 1 with him. Before then, I had never given a single lecture;my only prior teaching experience had been facing bored MIT under-graduates as recitation section leader for Bill McGill’s introductory psy-chology course. And here I was teamed up with a man who already hada formidable reputation as a lecturer! Ours was not the usual arrange-ment, where the course is neatly subdivided between the coteachers.True, Henry had his lectures and I had mine, but because of his some-what unpredictable commitments in New York at the time—he was a“cold warrior” working for Radio Free Europe—I had to be preparedto take over his lectures at a moment’s notice. Fortunately, the coursewas tightly organized—we had prepared detailed outlines, which werestrictly followed. Timing was everything: Each lecture was meant to lastprecisely one hour, and the goal was to finish the summary statementjust as the bell rang. It was this level of organization that made it possi-ble for Henry, arriving late from New York, to walk into the lecture hall,sit on the sidelines for a couple of minutes to make sure he knew exactlywhat point I had reached, and then take over from me without missinga beat.

Henry was not only my teacher and colleague at Swarthmore, butalso my stage director. As a graduate student, I had bit parts in Gilbertand Sullivan operettas, and as an instructor I had a small talking part inMolière’s Imaginary Invalid—yes, the faculty put on plays in those daysat Swarthmore. Since Henry did not know how to do anything byhalves, participation in a Gleitman production was approximately astime consuming as taking an honors seminar or teaching a course.There were numerous and protracted and quite spirited rehearsals; infact, one rehearsal was so spirited that I managed to sprain my ankle.However, Henry did succeed in getting his odd assortment of actors toput on quite creditable and memorable productions.

There is much more that I could recount about those early years, but Ihope that what I have said already helps to round out the picture of oneof the two remarkable psychologists we are celebrating this weekend.

Der Urgleit 25

Chapter 2The Wordgleits

Paul Rozin

My first exposure to a Gleitman was less than auspicious. One HenryGleitman had been selected by the professors to be the chairman of thePenn psychology department, starting in the fall of 1964. I had just ar-rived in 1963, and was full of enthusiasm for the wonderful, stimulat-ing, rapidly ascending psychology department assembled by itschairman, Robert Bush. I was a Bushophile. Much to my disappoint-ment, one Henry Gleitman was going to replace my fearless leader. Ihad met Henry at an EPA party in the spring before the takeover, andwas doubtful.

I soon appreciated that Henry Gleitman was just the man for the job.Bush had built a fine collection of researchers, and it fell to Gleitman toshape them into fine teachers. Henry quickly elevated the teaching ofpsychology, particularly of introductory psychology, into a major goal.He did this largely by his own example as a superb teacher; respect forteaching rose in the department. I was converted.

I soon realized that I had also gained a wonderful colleague and men-tor. We even collaborated on a research project on the decay of memo-ries: Could we slow down forgetting in goldfish if we cooled themdown in the retention interval? Our results were mixed, but they led toour coauthorship of a review paper on learning and memory in fish.

By Henry’s inspiring example, I became a psychology 1 teacher, a vo-cation that I proudly practice to this day. Meanwhile, my own main lineof research, on learning predispositions in rats, waned, and my desire todo something that might translate more directly into improving humanwelfare waxed. Henry, Dave Williams, and I led an evening seminar atFrank Irwin’s house in which we all, roughly simultaneously, shed ourSkinner boxes for work of diverse sorts on humans. For me, this meantan exploration of why the seemingly easy task of learning to read washard, and the daunting task of learning to speak was rather easy. Thedifficulty of early reading acquisition took its toll heavily in the innercity. I took a biological approach: We evolved as ear-mouth speakers,and the visual input was a new invention in our species. This line of

interest took me to my first intellectual contacts with Lila and a collabo-ration that lasted over five years, and included the design and testing ofa new reading curriculum. Lila and I had a swell time learning fromeach other, quipping and counterquipping, and, unfortunately, editingout each other’s clever lines from our joint publications. I have neverhad a more stimulating collaboration or collaborator.

Meanwhile, back at the stage, Henry was having a major effect on therest of my life through my children. He directed a play at the local ele-mentary school, where his daughter, Claire, and my daughter, Lillian,were students. It was HMS Pinafore. Lillian, at age 11, played Josephine,and her younger brother Seth, age 9, was recruited to be a member ofthe crew. It was a great success and instilled a great love of theater inthese two youngsters. So much so that Henry continued as theatercoach in frequent meetings of children from the cast in later months andyears. My wife, Liz, provided the musical coaching and background.And so was born a love for theater in Lillian and Seth. That led to themany pleasures of lead performances by both in junior high school,high school, and at Penn. The mark of Henry remains: Lillian took amaster’s degree in drama at the Tisch School at NYU, and now doescabarets and has written a musical review. Seth is now the artistic direc-tor of Interact Theater, a professional group in Philadelphia. Henrymade it happen.

Henry and Lila continued to be among my best friends, sharing theups and downs of life, and engaging me in interminable argumentsabout almost any issue. I haven’t collaborated with either of them forover fifteen years. But that doesn’t matter. They are members of my bio-logical family, and members of my academic family, my department.They have shaped my life as a teacher and scholar, and given directionto two of my children. They blur the distinction between friend andfamily, and are the best of both.

The Gleitmans are more than very good at words (and sentences,too). They are wordgleits, marvelous creatures that utter wonderfulword sequences; original (for the most part, never uttered before),pithy, trenchant, and delivered with panache. They generate oral andwritten words, and are both great at both. How many other couples canmake that claim?

It would be the opposite of an exaggeration (we don’t have an ade-quate word for this in English) to say that Lila and Henry were “one ina million.” Let’s take just Henry Gleitman (As Henry would no doubtsay at this point, I don’t mean “just” in the sense of “minimally” butrather, “only”). If Henry Gleitman was one in a million, that wouldmean there would be about 5,000 Henry Gleitman quasi-clones in theworld. Surely, there is only one Henry Gleitman. Evidence?

28 Paul Rozin

Well, first of all, Henry knows more about the intellectual accom-plishments of the Western world than anyone I have ever met. If all thepeople in the Western world were obliterated save one, Henry would bethe best bet for the lone survivor, in terms of saving as much as possibleabout what has been done. In my world, there isn’t a close second. Bythe way, the same holds in spades if we imagine the more imaginable(and to some, more appealing) prospect of the destruction of all livingpsychologists but one.

I’ve heard of a new way to measure importance or distinctiveness. Itcame from the Kennedy administration, and it was: How many phonecalls one would have to make to bring down a Latin American dictator-ship. A somewhat parallel measure more appropriate to Henry is howmany nontrivial characteristics of a person need be listed to establishhim or her uniquely as a human being. This is particularly easy forHenry, making him the prime candidate for being 1 in 5,000,000,000.

Consider the following:He’s the only person in the world (I think) who:

has taught 100 introductory psychology courseshas written an introductory psychology text and can whistle awhole movement of a Brahms symphonypublished papers separately, but never together, with both LilaGleitman and Paul Rozinis a director and a student of Tolman;and the list goes on.

While doing all this, Henry had time to have five car accidents, twochildren, play golf, direct perhaps twenty-five shows, and spend sixhours a day on the phone, coauthor one book of research and one ac-claimed scholarly textbook and many fine articles. And he did most ofthese things while conversing about weighty intellectual issues.

Henry is unique among academics. Poor Lila is a “regular” outstand-ing academic. She is only one among a few great graduate studentsponsors, past president of her field’s major professional society, andone of the few great psycholinguists in the world. Henry has chosen a“nonstandard” path. Thinking of the aims of academe as the creationand transmission of knowledge, we all know that we get paid to do thelatter (whether or not we do it well or with dedication), and rewardedfor doing the former. Henry has set a standard that few if any can equalon the transmission of knowledge. First of all, he has the critical prereq-uisite more than anyone else—he has the knowledge! Second, he is ded-icated to its transmission. Third, he mobilizes an incredible amount ofthought and energy to accomplish the transmission. Fourth, he is great

The Wordgleits 29

at the process of transmission, whether one-on-one, one-on-300, or, forthe case of the book, Psychology, one-on-1,000,000 or so.

Although Henry doesn’t appear on the New York Stock Exchange(we could all enjoy picking the right nonsense syllable for his three letter symbol), he has been one of the best investments in American his-tory. His profit, or his students’ profit, or perhaps, prophet, can be cal-culated in terms of income versus expense. I conservatively calculate,from his 100 psychology 1 courses alone, using current dollars, that hehas taught some 25,000 students (100*250), which, at Ivy league rates($2,000/course) generates $50,000,000 in tuition income. The costs inHenry’s salary, whatever it is (was) precisely, are well below the morethan $1,000,000/year that would balance this income. A return on in-vestment of over 10:1, for sure, and that doesn’t even count the knowl-edge of psychology, or enthusiasm for it, that Henry has transmitted.

And those who know Henry know that teaching psychology 1 is onlyone part of a monumental teaching effort. How much effort, you, oryour friendly economist might ask. My estimate follows:

Gleitman lifetime teaching timePsychology 1: 100 . 13 weeks . 3 hours/week = 3900 hoursWednesday night seminars: 30 years . 30/year . 5 hours = 4500hoursLong colloquium questions: 900 colloq. . 3 min. = 2700/60= 45hoursDrama coaching: 25 shows . 200 hours/show = 5,000 hoursAdvising: indeterminate but substantialAttending and advising on job talks: 10 hours/year . 30 = 300hoursSpeaking on behalf of teaching at Penn faculty meetings:(2 hours/year . 30 years) = 60 hours

Surely, this total of 13,805 hours is an underestimate, but it gives an ideaof the magnitude of the contribution.

This polymath, polyglot, polished but not polish (but close!) personcan play almost any role, say, for example, Louis XVI. Henry’s truehome should be at the head of one of the great courts of early modernEurope (see figure 2.1). Unfortunately, born when he was, Henry mustbe content to be the king of introductory psychology. And king he is.His image appears not only in his psychology text, but in others (see fig-ure 2.2).

Not satisfied with his own and Lila’s eminence in psychology, Henryhas introduced into The Book other signs of his lineage: Some are wellknown, but the fifth edition promises more (see figures 2.3, 2.4).

30 Paul Rozin

The Wordgleits 31

Figure 2.1.Henry Gleitman as courtier.

32 Paul Rozin

Figure 2.2.Konrad Lorenz and his ducks. (This is actually a picture of Konrad Lorenz, a dead ringerfor Henry Gleitman.)

The Wordgleits 33

Figure 2.3.The Gleitman family.

But the audience for Henry, his colleagues in psychology and collegestudents, is and has been much more limited than it has to be. New ver-sions of Psychology, for children and other deviants, may yet be forth-coming.

Henry is not just a superb orator, a highly educated person, and thequintessential psychologist. He is also a master at experimental design.His work with Jonides on Os and zeros is one of many examples. Thiswork, and related work by others, indicates that an object can be se-lected from an array of categorically different objects in a rapid, parallelsearch. However, objects sharing a common category are typicallyscanned serially. This important idea can be used to determine Henry’sown true category. To illustrate, one can ask how long it takes to findHenry in arrays of different types of objects. We have done so (of course,running balanced trials, with Henry’s image located at randomly differ-ent positions in different arrays), with arrays of such varied things asgarden equipment, fire hydrants, and reptiles. Most telling and mostcritical are the results from the two arrays presented below. Henry is de-tected by parallel/rapid scan when embedded in an array of photos ofprofessional basketball players (figure 2.5), but he is hard to detect, andmerits a serial search, when embedded in an array of great psycholo-gists (figure 2.6). This and related comparisons lead us to the inevitableconclusion that Henry is a great psychologist.

34 Paul Rozin

Figure 2.4.In an act of modesty, the identity of the child, grandchild Phillip, is largely obscured, butthe Gleitman visage somehow comes through.

The Wordgleits 35

Figure 2.5.Can you pick out Henry Gleitman from the Philadelphia 76ers?

Figure 2.6.Can you pick out Henry Gleitman from the eminent psychologists?

Henry Gleitman belongs, as the great twentieth-century introductorypsychology text author, in a chimeric relation to his nineteenth-centurypredecessor, William James (figure 2.7).

Who am I, Paul Rozin, to say and know all this about Henry and LilaGleitman? My credentials are impressive:

1. I am the only person who has published separately with Henryand Lila, but never together.2. I was promoted to Associate Professor with tenure under thechairmanship of Henry Gleitman.3. Lila Gleitman was brought to the Penn psychology departmentas professor, from the Graduate School of Education at Penn,under my chairmanship.4. Henry Gleitman is the theater father of two of my children: hewas their first teacher and director of theater and instilled a lovefor theatre in them that became a main theme of their lives.5. Lila Gleitman studied language development in two of mychildren, and reported that young Seth, when asked: “Two andtwo is four: Is there another way to say that?” received the re-sponse: “One and three is four?”6. Lila and Henry, Claire and Ellen, have been quasi family mem-bers to me and my family for some thirty years.

So, with all this contact, and all this affection (and mutual roasting atvarious celebratory events), what can I say about two of the most in-tense, indefatigable, informed, and intelligent people in the world?

As John Sabini notes, Henry has a saying: “If it isn’t worth doing, itisn’t worth doing well.” (The reverse also holds.) This reflects the inten-sity of both Henry and Lila (no halfway commitments here). There’s aparallel to this that I’d like to put forward: “If it isn’t worth feelingstrongly, it isn’t worth feeling at all.” These are passionate people:When they watch TV sports, nurture orchids, eat at Sagami Japaneserestaurant, relate to friends or to family or to students, there is an inten-sity, an enthusiasm that is rarely matched. That’s why it’s great to betheir student, their friend, and yes, even their orchid, house, or televi-sion set.

There are more sides to the Gleitmans than one can convey, even in abook. Lila, alone, is comfortable as both figure and ground, and thrives,along with her field, on temporary states of ambiguity, linguistic or vi-sual (figure 2.8).

I stand in a mix of awe and affection as I contemplate them and theirswath of influence and interactions on this earth.

36 Paul Rozin

The Wordgleits 37

Figure 2.7.Henry Gleitman and William James.

38 Paul Rozin

Figure 2.8.Some of the many sides of Lila Gleitman.

Chapter 3

Multiple Mentorship: One Example of HenryGleitman’s Influence

Robert A. Rescorla

Henry Gleitman has inspired a large number of undergraduates, grad-uate students, and colleagues in psychology. But few have been as for-tunate as I in interacting with Henry in all of these relationships.Indeed, I suspect that, with one or two exceptions, I have had more dif-ferent academic relationships with Henry than has anyone else. Con-sequently, I am pleased to have the opportunity to reminisce aboutsome of the things I have learned from Henry in those relationshipsover the course of nearly forty years.

For me, as for thousands of others, the first interaction with Henrywas in introductory psychology. I entered that course in the spring se-mester of my freshman year, 1958—Swarthmore did not allow first-semester freshmen to take it—eager to develop my expertise in the likesof Freud and Jung. I was initially appalled at having to attend a largelecture course with almost one hundred other students. My othercourses that term ranged in size from four to twenty students. But, likemany others, I quickly became completely engaged in the course asHenry brought to it his famous enthusiasm, his ability to highlight theessence of a concept, and his way of making the concepts memorable bya turn of phrase.

I recently looked back over my notes from the course, a total of 71pages for the 38 lectures. It was a full service psych 1, covering topicsranging from the “nervous system” to “emerging patterns in our soci-ety.” Of course in those days we had to use someone else’s book;Hilgard was the text and the bias was decidedly on experimental psy-chology. We dispensed with the nervous system in two lectures. Wespent seven lectures on sensation and perception and eleven on learn-ing. The latter did not count the two additional lectures spent on forget-ting and the five devoted to the topic of motivation, which emphasizedacquired motives. Personality was dealt with in three lectures, as wassocial psychology. The topic of Freud was not officially a part of theclass at all; rather that topic was reserved for three extra evening lec-tures, which were something of a spectacle for the whole campus.

Three particular things caught my eye as I reviewed these notes: First,the topic of language was sandwiched into two lectures between sec-tions called “complex learning” and “theoretical issues in complexlearning”—hardly a forecast of Henry’s future emphasis. Second, mynotes were especially fuzzy about a series of experiments on avoidancelearning apparently done at Penn, which I labeled in the margin as“Sullivan’s dogs.” Third, my notes clearly indicate that even thenHenry’s grandmother was the font of all wisdom in psychology.

My only disappointment in the course was Henry’s evaluation of myterm paper. I wrote on Köhler’s Mentality of Apes, believing that I hadcaptured the essence of the book in a few short phrases. Henry’s evalu-ation was even briefer: “Too discursive.” I was so shocked that anyonecould think me wordy that I completely changed my writing style to thepoint were some would say that it is now telegraphic and cryptic. Ofcourse, if I had known Henry as well then as I do now, I might have re-sponded with a homily about pots and kettles. Thankfully for all of us,this paper has been lost in my archives.

This course diverted me from my intended path toward the Meth-odist ministry. It set me on the way toward becoming an experimentalpsychologist. One of my offspring has commented that rarely since thedays of the Roman coliseum have so many Christians been saved froma terrible fate.

Two years later I showed up as an advanced student in Henry’s hon-ors seminar on learning. This seminar was famous for its exciting, spir-ited discussions, its long meeting hours, and its attendance by otherfaculty such as Jack Nachmias. We frequently argued Spence andTolman, Hull and Mowrer from 7:00 in the evening until well past mid-night. We read such secondary sources as Hilgard’s Theories of Learningand relevant portions of Osgood’s classic Experimental Psychology. Butwe also read an incredible number of original papers in original jour-nals by Spence, Lawrence, Bitterman, Kreshevsky, Meehl, Sheffield,Crespi, Tolman, Miller, Mowrer, Solomon (by now I had learned to spellhis name), Harlow, Guthrie, Estes, Asch, Melton, Underwood, Postman,Rock, etc. It was a tremendously deep and wide-ranging seminar. It metfourteen times and I took almost four hundred pages of single-spaced,typed notes. Obviously in the two years since I had taken introductorypsychology, Henry had learned a great deal more worth taking noteson!

It was this course that introduced me to the excitement of careful ex-perimental design and the importance of close logical reasoning. It setme on my choice of specialty within psychology. I still remember writ-ing a paper on Spence’s explanation of transposition and being excitedby the appearance of Mowrer’s 1960 book. I also remember being capti-

40 Robert A. Rescorla

vated by the experiments that Dick Solomon had done. It was thatwhich led me to go to Penn to work with Dick. Probably more than anyother class, this seminar changed my life’s course.

That summer, Henry gave me my first taste of real research. I workedwith him, Bill Wilson, and another undergraduate (Maggie Herman) ondelayed response learning in infant monkeys. Every day I hoppedaboard my motor scooter, the standard-issue form of transportation fora Swarthmore undergraduate, and sped over to Bryn Mawr to run themonkeys. That summer, Henry taught me the importance of attendingto every detail in the design and execution of experiments. We spenthours arguing over how to run the experiments and build the equip-ment. He also taught me the importance of understanding the specieswith which you are working. I developed a deep respect (and fear) offifteen-pound infant rhesus monkeys who carry the deadly virus B. Istill remember the time when one got loose while I was transportinghim to the experimental room: he ran down the hallway to the right andI ran down the hallway to the left, in search of someone willing to catchhim. I also learned that rats are not just furry miniature humans. Myfirst encounter with rats came that summer when some were deliveredto Henry’s lab for studies he was beginning on forgetting. The colonyroom seemed very hot to me and when I looked at the rats I becametruly alarmed about their health—they each seemed to have large tu-mors projecting from under their tails. I ran to Henry’s office to sum-mon him to see that they got proper medical attention, only to be morethan a little embarrassed as he noted that the males of many specieswere so equipped, though he admitted to the special endowment of ratsin this regard.

That summer produced my first publication, with Henry as the firstauthor and me as the last: “Massing and within-delay position as fac-tors in delayed response performance.” So now I found myself in a newrelationship with Henry, as co-author, and I learned two more lessons:In a serial position of four authors, if you are not first it pays to be last ifyou want to be noticed; and last authors tend to have little clout in deci-sions about writing.

It was several years before I had my next relationship with Henry, asmy first experience being a teaching assistant. In the meantime, I hadbecome a graduate student at Penn and he was recovering from a shortsojourn at Cornell, doing penance by accepting the chairmanship atPenn. He had been brought to Penn to renovate the undergraduate pro-gram in psychology and to re-instill enthusiasm for teaching in the de-partment, both of which he did, with effects that last to this day. But as agraduate student I had little interest in such matters. I was supported onan NSF graduate fellowship (these were the post-Sputnik times when

Multiple Mentorship 41

we all lived well) and saw my task as becoming the best researcher Icould. I deeply resented having to waste my time learning to teach. Ieven considered making a formal challenge to the departmental re-quirement that we all serve as teaching assistants, no matter what oursource of support. But in the end Henry prevailed. Just as Henry got mestarted down the road as a researcher, he got me started down the roadas a teacher. It was as a TA in his class that I began to experience thepleasure of teaching others. I also had, of course, the opportunity to ob-serve from a new perspective the master teacher. I like to think that Ipicked up a few tricks from him—although I never did master the art ofholding a class’s flagging attention for fifteen minutes by waving a cig-arette in my hand, the filter end pointed away from my mouth, acting asthough at any minute I might light that end by mistake. Henry knewhow to hook us on teaching. After the class was over, he confided toMichael Lessac and me that the undergraduates had praised us highly,something I suspect he still says to all his TAs.

My next relationship with Henry was having him as a member of mythesis committee. Dick Solomon was the main advisor and the newlyappointed assistant professor Paul Rozin was the other member. I havealways been grateful for a committee that basically left me alone. ButHenry taught me two important points in that context: (a) make yourdissertation a coherent story that focuses on one primary point andmakes it clearly; and (b) do not include extraneous material. The com-mittee in fact insisted that I drop four of the six experiments I hadwanted to include in my thesis and instead write up just the two centralones. Since then I have come to realize that in many ways the presenta-tion of one’s work is as important as the work itself, if you want to in-fluence the thinking of others. That was the lesson Henry was trying toteach me then. Of course, we can judge the speed with which I learnedthis by the fact that my dissertation, on inhibition of delay, remains tothis day a widely unread paper.

When I got my degree, I took a position at Yale. With Henry at Penn,we communicated only occasionally, although I still recall one occasionwhen I called him while he was briefly in the hospital. I was struck bytwo features of that interaction. First, he wanted to talk psychologyfrom his hospital bed; as I recall he had recently written a paper on getting animals to understand the experimenter’s directions, which of-fered ways for analyzing the structure of an animal’s associative knowl-edge. He insisted on getting my comments on that paper. Second, hecould only spare me a few minutes because so many of his colleaguesfrom the Penn department were in the room visiting. I think that theywere all terrified that something would go wrong with Henry and theywould end up having to teach psych 1.

42 Robert A. Rescorla

It was fifteen years later that I returned to Penn to be a faculty mem-ber and Henry’s colleague. I was unprepared for the stimulating intel-lectual atmosphere that I encountered. At that time it was a place ofsuch a high level of interaction that I had to have an automatic door-closer installed to get any work done. One of the people who fosteredthat interaction, of course, was Henry. He was always ready to talkabout experiments you were doing. He was always concerned to main-tain the sense of community in the department.

A few years after I returned to Penn, I had the opportunity for a rolereversal with Henry. Just as he had been my chairman when I was agraduate student, I became his chairman in 1986. The first lesson hetaught me in that context was that any thoughts of building a depart-ment are secondary to the necessity of keeping the faculty you have. Nosooner was I chair than, much to my horror, he and Lila received an at-tractive offer from another institution. I have always considered it oneof the main unacknowledged achievements of my chairmanship thatthey decided to remain at Penn. But Henry also greatly helped me in thedaily tasks of being a chair. He taught me about my responsibilities tomaintain the quality and atmosphere in the department. He would reg-ularly come into my office to be sure I was aware of problems and issuesthat could adversely affect the department. He taught me the impor-tance of listening to my colleagues with a new ear. I will always begrateful for the wise council he gave me, some of which helped preventme from making terrible mistakes.

Now I am in yet another relationship with Henry, as his dean. Fromthis vantage point, I can see the commitment that he has to the institutionand to its intellectual life. I can see the contribution he makes not only tothe psychology department but also to theater arts and to the institutionas a whole. I can also see that my own decision to accept the position ofdean of the undergraduate college was clearly heavily influenced byHenry’s example as a dedicated teacher and citizen of academia.

So I have seen Henry from many viewpoints: He taught me my firstpsychology course, my first learning course, gave me my first researchjob, my first publication, supervised me in my first teaching position,was a member of my thesis committee, was a member of the faculty thefirst (and only) time I was chair and the first (and assuredly only) time Ihave been dean. From each of these I have learned something aboutHenry and about myself.

But in addition to these general influences on my thinking and career,Henry has had some quite specific influences on my research. I want todescribe briefly two experiments in order to make that point.

The first experiment is one that Ruth Colwill and I performed severalyears ago, Colwill and Rescorla (1985). The issue addressed by the

Multiple Mentorship 43

experiment was the nature of instrumental learning. When, for in-stance, a rat learns to press a bar for food reward, what is it that the ani-mal learns? This is a complex problem with many different pieces, butone issue that stands out has to do with the role of the reward. One cancontrast two classic alternative roles, as indeed Henry pointed out in hislearning seminar those many years ago. The alternatives are that the reward serves as a condition of learning or that it serves as part of thecontent of learning. On the second account, the animal learns that barpressing produces food, what one might describe as a response-outcome association. This is the most obvious thing that he might learn,but not the one proposed by many classical theories. Those theories in-stead saw the food not as part of the content of learning but as a condi-tion of learning. In that view the food serves to stamp in associationsbetween the response and antecedent stimuli. In effect, in that theory,the occurrence of food after a response has been made in the presence ofa stimulus stamps in that S-R association. Or as Henry put it in 1960, thereward serves as a catalyst, helping the animal fuse two other events.

One way to address that issue is to ask about the impact of changes inthe value of the food after the instrumental learning has occurred. If theanimal has learned that lever pressing produces food one would thenexpect learning that food is not valuable would have an adverse effecton lever pressing. On the other hand, if the animal learns an associationbetween some stimulus and the lever press, an association previouslycertified by food, there is no encoding of the food in the learning; conse-quently, subsequent changes in the value of the food should have littleimpact.

Colwill and I used this logic to construct a relatively elaborate exper-iment, the design of which is shown at the top of figure 3.1. We trainedrats to make two responses, lever press and chain pull, each earning a different outcome, a small pellet or liquid sucrose. Naturally the ani-mals made both responses. The question was whether they had the response-outcome associations or the outcomes had simply served ascatalysts helping them learn associations between the responses andantecedent stimuli. To find out, we divided the animals into two groupsand changed the value of one of the outcomes in each group. Ourchange device was the emetic agent LiCl. We simply gave the rats oneparticular outcome (either pellet or sucrose) and then administered LiClso that they felt ill. Such a procedure is well documented to reduce theattractiveness of the food. Then we brought the rats back into the situa-tion and allowed them to make a nonreinforced choice between thelever and chain. If they know what response leads to what outcome andthey know that one outcome is unattractive, then they should be moreenthusiastic about the other response.

44 Robert A. Rescorla

Multiple Mentorship 45

Figure 3.1.Design and results of experiment identifying the presence of associations between re-sponses (R) and outcomes (O). Rats were trained to earn two different outcomes by mak-ing two responses. Then one outcome was devalued by pairing with LiCl and theanimals were given a choice between the responses. Responding is shown during an ex-tinction test after devaluation. (After Colwill and Rescorla, 1985.)

The bottom half of figure 3.1 shows that this is exactly what hap-pened. That figure shows responding over the course of the extinctiontest. The data are separated for response whose outcome was devaluedby LiCl and those for which the outcome was left valuable. Althoughprior to devaluation the responses were made with equal frequency,after devaluation the responses whose outcomes were paired with LiClwere immediately depressed. That finding is important in its ownright—it tells us a good bit about what is learned. It means that the out-come plays a role beyond that of a catalyst and is an actual participantin the learning itself. But the result is more important for the role that itcan play as a valuable analytic tool, allowing us to measure the state ofassociations after a wide range of treatments. We have been successfullyexploiting that tool in the exploration of instrumental learning andPavlovian conditioning for the past decade. For instance, using thistool, one can show that these originally learned associations remain in-tact through such manipulations as extinction produced by rewardomission. So this first experiment has proven to be quite important.

The second experiment is conceptually quite like the first. In this ex-periment, rats were trained to choose between a left and right goal boxin a T-maze, shown in figure 3.2. The rats got the same food whichever

46 Robert A. Rescorla

Figure 3.2.Floor plan of the maze used by Tolman and Gleitman (1949). Rats were trained to entertwo distinctively different goal boxes by making left or right choices. Then one goal boxwas devalued by pairing with shock and the animals were given a choice of turning leftor right.

goal box they entered, but the goal boxes were arranged to be distinc-tively different from each other. So one might say that the left and rightresponses led to distinctively different outcomes, different goal boxes.One can then ask whether the animal has learned those associations be-tween the response and the goal-box outcome. This can be answeredusing the same logic as in the previous experiment, by changing the an-imal’s attitude toward one of the goal boxes. For instance, one mightplace the animal directly in one goal box and apply electric shock. Thenone could bring the animal back into the choice situation and allow himto go left or right. When these treatments were carried out, 22 out of 25of the rats chose the nonshocked side on the first choice trial. Since theycould not see the goal boxes when they were making the choice, thismust mean that they knew which outcome followed which response.Clearly the animals knew which goal boxes followed which responses.

As it happens, this second experiment was not done in my lab in the1980s but instead in Tolman’s lab in 1947. In fact, this experiment wasHenry’s dissertation, published as Tolman and Gleitman (1949). Themethods Henry had at his disposal were more primitive, but the logic isthe same as our experiment from almost forty years later. One might le-gitimately say that our experiment is little more than a refinement ofHenry’s thesis. As usual, he saw the issues clearly and identified how toseparate them. He lacked only the technology.

I am fond of saying that most of the really important ideas are oldideas. For that reason, I routinely advise my students to read the worksof certain major contributors who I think had the best perspective onthe learning process. For my part, I regularly reread the books of Pavlov,Konorski, and Kohler. They are full of ideas that are well worth stealing.But it is now clear that I also reread Gleitman and have greatly profitedby stealing from him.

Henry has not only had a broad impact on my attitudes and my ca-reer, he has also been responsible for my pursing certain specific ideas.For all of this I am deeply grateful.

Acknowledgment

The writing of this chapter was supported National Science Foundationgrants BNS-88-03514 and IBN94-04676.

References

Colwill, R. M. and Rescorla, R. A. (1985). Post-conditioning devaluation of a reinforcer af-fects instrumental responding. Journal of Experimental Psychology: Animal BehaviorProcesses 11:120–132.

Multiple Mentorship 47

Hilgard, E. R. (1956). Theories of Learning, second edition. New York: Appleton-Century-Crofts.

Köhler, W. (1925). The Mentality of Apes. Trans. by E. Winter. New York: Harcort, Brace.Osgood, C. E. (1953). Method and Theory in Experimental Psychology. New York: Oxford

University Press.Tolman, E. C. and Gleitman, H. (1949). Studies in learning and motivation: I. Equal rein-

forcements in both end-boxes, followed by shock in one end-box. Journal ofExperimental Psychology 39:810–819.

48 Robert A. Rescorla

Chapter 4

Some Lessons from Henry and Lila Gleitman

John Sabini

I came to Penn in 1976. Henry was on leave at the time and Lila was inthe School of Education, so neither had anything to do with my beinghired—a fact they have found more than one occasion to remind me ofover the years. It was not long after I was hired that I came to learn a lotabout teaching from Henry and Lila. I wasn’t lucky enough to be agraduate student of theirs, as were many of the other contributors tothis volume, but I was mentored by them nonetheless. Let me tell yousome of the things I have learned first from Henry, then from Lila.

One thing I learned from Henry is that “If a thing isn’t worth doing, itisn’t worth doing well.” (I know; this looks like a typo, but it isn’t.) Thetypical occasion on which Henry utters this is when someone presents aproposal to carry out a very complex and elegant experiment, a well-designed experiment, but an experiment designed to answer a questionof monumental unimportance.

Henry believes that psychology isn’t chess. Our task as psychologistsisn’t to produce appealing designs; that is for fabric-makers. Henry re-minds us that our job is to figure out how it all works, how the mind andeven how the soul works. If an experiment won’t help us do that, thenno matter how beautiful its design, the experiment isn’t worth doing.

Henry has very strong views about teaching as well as about re-search. One thing he thinks (and says, vehemently!) is that people paytoo much attention to the A+ student. We all love to teach the A+ stu-dent, especially if we can win that student over to our preferred kind ofresearch. Indeed, we brag about and feel satisfied about having taughtthat student and having captured him or her for our team. And Henryhas much to be proud of in that line. Most of the contributors to this vol-ume were (graduate) students of Henry’s—won over by him. And, in-deed, just at Penn there are two former undergraduates of his on thefaculty—one the president of a large, ivy-league university. The other isa former associate dean and member of the National Academy ofScience.

But these are not the people, Henry would be the first to tell us, ourteaching should be aimed at. For one thing, people like this barely need

teachers at all; they are perfectly capable of learning on their own. Allthese folk need is a campus map with the library on it (or, maybe, just aconnection to the Internet) and they will find their own way to the truth.No, it is with the C student that we can make a difference. The C student,Henry tells us, is perfectly capable of learning absolutely nothing in acourse. Worse than that, our C students can learn nothing and still per-suade themselves they have learned something, leaving them worse offthan they started. But the C student might also learn a lot in a course ifthe instructor teaches a course for them. It is with the C student, not theA student, then, that the marginal utility of a good teacher is clearest.

Sure, it is more fun to teach the A student. And it is very rewarding toturn that A student’s head, to turn her from the field she was headed fortoward psychology. But, as Henry is fond of asking, “So, there is onemore brilliant psychologist in the world. But there is one fewer brilliantplaywright (or whatever that brilliant psychologist would have becomehad she not become a psychologist). Is the world really better off withone more brilliant psychologist and one fewer brilliant playwright?”(Anyone who thinks Henry would answer that one yes, has never methim.)

These beliefs about teaching are of a piece. Henry has, as one mightnot expect from a native of Leipzig, an essentially Jeffersonian attitudetoward teaching: The point of our educating the C student is to makehim or her a better citizen, a person better able to understand the news-paper, in particular, better able to understand the Science Times article onthe latest discovery about the brain, or identical twins, or social interac-tion. Our educational efforts ought to be aimed at making our studentspsychologically literate. It is not necessary, however, Henry might tellus, to make them psychologically creative.

I have learned a lot in many ways about creativity from Henry.Henry—one of the handful of most creative people I have met—doesnot have an unambivalent attitude toward that aspect of the humanpsyche. He isn’t actually fully opposed to creativity; he’s just suspiciousof it. And most importantly he believes that a scholar’s first obligation,most important obligation, is to master what the past masters havepassed on. Henry knows that the truly creative types, the Newtons,Einsteins, Darwins, and Helmholtzs are, of course, what it’s all about(as he might say). But, sadly, most of us can’t be a Newton, Einstein,Darwin, or Helmholtz. Still, as vital to the life of the mind as those folksare, so too were the medieval monks; the ones who copied the classicaltexts—perhaps without the foggiest understanding of the languagethey were copying. And, more optimistically, we all can do that. And wecan teach our students to do that. The first thing, then, we must teachour students, Henry might tell us, is to love and, because we love, pre-

50 John Sabini

serve our intellectual heritage. Once we and they are fluent in Latin, orGreek, or Hullian learning theory, then maybe, just maybe it is time foran original thought. So Henry does appreciate creativity; he thinks ithas its place, a lofty place, perhaps the loftiest place. But fostering cre-ativity isn’t the only thing a teacher does, or even the most importantthing a teacher does.

Henry and I have both written textbooks. I learned a lot about thatfrom Henry too. First, I learned to take my time. It took me ten years towrite the first edition of my book. I was fortunate, though; my publisherdidn’t grow impatient. After all, they had also published Henry’s text,so they thought my book was a rush job! (Indeed, once they got it, theyheld on to it for a while to let it age, I suppose, before releasing it.) Ilearned from Henry that a textbook is a Renaissance cathedral. Like aRenaissance cathedral it is meant for all comers, all who want to enter,so long as their hearts are pure. It is a place where the greatest theolo-gians can come, but also where the humblest peasants can worship;each should find in its great expanse something to sustain his faith.Indeed, a well-written textbook should educate all who read it, the in-structor as well as the student.

A textbook should also, like a cathedral, be richly and colorfully dec-orated. It should delight the eye as well as the mind. And Henry hasmade sure that his text is a delight to the eye. Henry will also remind us,though, that as important as the art is, it is the last thing that goes intothe cathedral. It is certainly the art that catches the eye of the tourist, butwhat the cathedral really is is a structure. For Henry it is the structure ofa thing that really matters. As it is with a cathedral, the most importantthing about a textbook is its structure.

Henry is, of course, a marvelous stylist—despite not being a nativespeaker (as is his wont to remind us). But the last thing Henry would dois create a book by writing a collection of apercus and then looking for away to arrange them! No, no, no. “A thousand parts architecture beforea single ounce of decoration,” might as well be Henry’s motto. The hun-dreds of conversations we have had about our books were never aboutstyle, and rarely about the exact content—though some were. Mostwere about the structures of our books. Getting that right; that was thething. If you have gotten the structure right then the reader will be ableto grasp the book and the field.

Henry’s text and Henry’s lectures are, of course, thoroughly up-to-date. It is in a way a little odd that they are up-to-date because Henrydoesn’t actually think it is all that important that they be that way. Yousee, the kiddies, as Henry refers to psych 1 students (I will have more tosay about that in a moment), don’t really need to know what is up-to-date. They need to know something else; they need to know the deep

Some Lessons from Henry and Lila Gleitman 51

questions. The deep questions, Henry would tell us, change only veryslowly if at all; the up-to-date answers change every day. By the time thelast make-up exam is given in this year’s psych 1 course, Henry mightpoint out, the up-to-date answers are already out of date anyway.

Now, Henry has nothing against up-to-date answers (just as he hasnothing against creativity). It’s just that he believes we must not be mis-led that we have succeeded in our educational mission if we have ex-posed our students to the up-to-date answers our field offers, if wehaven’t shown them how these answers relate to the deepest questionsour field would like to answer.

This attitude of Henry’s might, one could imagine, lead someone intotrivializing contemporary research. But this is never what Henry does.A positively delightful experience is to have Henry explain your workto someone else. If you are lucky enough to have this happen to you,you will learn that you all along spoke prose—and so wittily, so ele-gantly, so profoundly. Henry will, I promise you, explain how yourwork reaches to the deepest possible issues, how it bears on nothingshort of the nature of the human soul. He has an utterly astonishing ca-pacity to do this. And it is one reason (though only one reason) his stu-dents have been so successful.

As many of the contributors to this volume know well, there is noplace this talent is more on display than at the practice job talks thatHenry and Lila have for each of their fledgling scholars. Practice jobtalks are an institution in which the student about to go on the academicjob market gives his or her dissertation talk to an audience composed ofHenry, Lila, other students, and an assortment of faculty. The facultymembers are each assigned a role to play. One of us might be assignedthe role of, say, parallel distributedprocessing professor, another therole of cognitive neuroscientist, and so on. In any event, the student de-livers his or her job talk to the assembled guests. Then, when the stu-dent is done, the fun begins.

Henry and Lila now give the talk that the student should have given.Now mind you, the talks the students give are usually very good; theseare, after all, very good students working with superb mentors. But thetalks are never as good as they could be. Henry needs to fix them. Andhe fixes them this way: (1) They always need to be simplified. There isalways too much stuff in them. You know, the stuff that goes into theunread parts of journal articles. The stuff that shows off what a good,conscientious scientist one is; the stuff that shows one is, as Henry putsit, house-broken. The audience doesn’t really need to hear that. (2) Thereal point of the talk is often a bit buried and surprisingly often neverstated. It needs to be found and exposed. The listener must never losecontact with what the main point here is. And (3) the souffle that is

52 John Sabini

being baked needs to be lightened. Humor, clever visuals (often in theform of one of Henry’s cartoons), a bit of performance needs to bewhipped into the mix just to pick it up, to make it more graceful. Theseare the ways that Henry the director fixes the play.

I have never seen Henry fix someone’s psychology 1 lecture, but I amsure the process would be identical. This is just as it should be. Henrybelieves, of course, that there are differences between a lecture to a pro-fessional audience and one to a psychology 1 class. He knows that youcan‘t give the same talk to both audiences; but he believes, I think, thatthose differences are much less significant than we usually think.

Erving Goffman pointed out that it is common for people in “serviceindustries” (like teaching) to disparage their customers. It is common inuniversities, for example, to hear faculty complain about students. Andauthors are almost as likely to grumble about their readers as they are tovilify their long-suffering publishers! But this I have never heard Henrydo. I have never heard Henry complain about undergraduates, not oncein twenty years.

Undergraduates are, Henry would remind us, to be respected, notdisparaged. We are not here to judge their motives in taking our coursesor to question their intelligence, sincerity, or integrity. It is, after all, ourcalling to teach them. If we aren’t called to this, we shouldn’t be in uni-versities. It is not good for our students for us to demean them, and itisn’t good for ourselves. If we turn teaching undergraduates into a mat-ter of casting pearls before swine, then what have we turned ourselvesinto? As Goffman didn’t point out, the tendency on the part of serviceprofessionals to denigrate their customers is—at least in the case of undergraduate teaching—as unwise as it is unwarranted; it is twicecursed.

Eventually I became the chairman of the psychology department andas such received many gifts from Henry. One thing I received was thedepartment itself. Henry came to Penn in 1964 as chairman. He camehere in the wake of dramatic changes in the department wrought by thelegendary Robert Bush. Bush moved and shook, but it was Henry whostabilized and solidified the department and its administration. But thispaper is about teaching, not administration. What Henry as a teacherdid for me as a chairman is this: When I took over our department it wasin very high favor with the dean’s office. And we were in the dean’sgood graces in large measure because we had a reputation for being adepartment that took undergraduate teaching seriously. (We had a rep-utation for doing that long before universities discovered their under-graduates and the apparently obscure fact that undergraduate tuitionpays faculty salaries.) Why did we have that reputation? Because ofHenry.

Some Lessons from Henry and Lila Gleitman 53

I am convinced that a necessary condition of a department’s takingundergraduate education seriously is that at least some of its most intel-lectually respected members be highly visible undergraduate teachers.If you have someone who has the respect of his or her colleagues as anintellectual who also teaches highly visible, introductory courses andwho treats doing this as Henry treats it, as a calling, then you have achance of having a department that treats undergraduates seriously.Now certainly in our department Henry has not been the only person toplay this role, but he has been the Olivier of it. Thank you Henry.

Now, what have I gotten from Lila? In my career I have had no morestimulating experiences than coteaching various graduate seminarsover the years with Lila. That will come as a surprise to none of mycoauthors in this volume. Lila, like Henry, is a great teacher. But sinceHenry is also a director it is a lot easier to figure out how he does it thanhow Lila does it. All you have to do in Henry’s case is listen in on the in-structions that Henry the director gives Henry the performer. It isharder to see what Lila is up to. However, I have watched for so longnow that I have a few hints.

First off, any intellectual encounter with Lila is suffused with a par-ticular spirit; that spirit is that you and she are now going to come to un-derstand something (whatever it is you are discussing) together. She ison your side. Always. It is always the same. You and Lila are studentstogether and together you will master this question, whatever it is. Lilacan pull this off because she has a talent all good teachers have, but shehas it in spades.

The talent is this: As you and Lila think about some problem, Lila willconvince you that this is the very first time she has thought the problemthrough. Oh, I don’t mean she claims this, and she would no doubtdeny it if asked point blank. It’s just that she makes you feel it is true. Iswear to you that every time I hear Lila explain why you can’t under-stand the acquisition of nouns just by saying that people point at a rab-bit and say “rabbit” I believe it is the first time she is discussing it withsomeone. Henry is the only director in the family, but he is not the onlyperformer.

How did I wind up coteaching various and sundry seminars withLila, since after all I’m not a psycholinguist? Well, to answer that I musttell you about the relationship between teaching and the rest of Lila’slife. And I have to start with a rule I had to make up for Lila when I waschair of the department. For the rest of the department there was a rulespecifying the minimum number of courses a person was to teach; withLila I had a rule about the maximum. I needed that rule because thenumber of courses Lila teaches is directly connected to the number of

54 John Sabini

good ideas she has, and she has so damned many good ideas! Here’show the ideas and the courses are connected.

If Lila has thought some topic through really well and thinks she hassomething really developed to say about it, then it is something onwhich she wants to do an undergraduate course. As we all know, un-dergraduates test ideas in ways no other audience can, precisely be-cause they haven’t as yet bought into the prescribed professional way oflooking at the world. But, of course, though exposing our ideas to un-dergraduates is useful to us, our students aren’t here to be useful to us.So it is only our well-developed ideas that they need to hear. (After all,as Henry might say, they could be taking a course on Shakespeare in-stead!) So as far as I can tell, then, for Lila undergraduate courses areplaces where she can share her most worked-out good ideas with (veryjunior) colleagues.

There are some of us—Henry, I think, and myself certainly—whocould be gotten to teach all sorts of things that we have no actual inter-est in; we might well view it as an interesting technical problem to fig-ure out how to do it, or something like that. But that is not Lila. Oh, true,Lila could teach things other than the psychology of language. Shecould, for example, teach a course on bridge bidding—though I thinkshe thinks that would be a course on the psychology of language. (Andshe could perhaps be induced to give a course on orchids!) But I cannotsee Lila teaching something in which she had no interest. For her teach-ing undergraduates is too deeply connected to the rest of her intellec-tual life for that. (But neither is Lila of the self-indulgence school ofundergraduate education, the school that thinks that undergraduatecourses are for the edification of the instructor. Nor does she believethat the chance merely to be in the same room as the great genius in-structor is educating enough, that there is no real need to prepare lec-tures!)

Lila’s teaching also goes on, as all of the contributors to this volumeknow, at the Gleitman weekly research seminar. That is where all (or al-most all) of the research programs discussed in this book werelaunched. Since other people have written about them in this volume, Iam sure I needn’t. But perhaps I could say a word about how one comesto be a student of Lila’s.

Of course some students come to Penn specifically to work with Lilabecause of her international reputation. But that is only how some of herstudents become her students. Others come to work on all kinds ofother topics. They might come to work on clinical or social psychology.But, nonetheless, they go for a meeting with Lila because, maybe, theyhave been sent there by the graduate czar or czarina. And Lila says

Some Lessons from Henry and Lila Gleitman 55

something like this to them: “So, you want to work on, say, moral think-ing. Well that’s certainly a nice topic, but, for me, I can’t imagine howanyone could find anything interesting except language acquisition. Iknow other people seem to find other things interesting but . . .” Nowthe poor student asks, “Well, what’s so interesting about language ac-quisition? I mean we all know how that works. An adult says the wordrabbit and points to a rabbit.” At this point the poor student is doomed.Lila will now say, “Yes, yes, of course you are right. It must be just asyou say. It couldn’t be any other way. But there is this one little problemthat Quine pointed out. . . .” And we know what happens after that—anincurable obsession with language follows. It is important to stress thatLila doesn’t just pull the rug out from under the student; she also showsthem in her own work how we can learn about how language is ac-quired. All praise to Quine for pulling the rug out, but that is enough fora philosopher; for Lila the scientist and for her students, there must bemore. The rug must be replaced with the firmer stuff of a research pro-gram. So every student who goes to see Lila is at grave risk of becominga student of language acquisition.

The rest of Lila’s teaching is in the form of graduate seminars that areusually cotaught with one or more colleagues. These seminars come topass for one of two reasons; either Lila decides that there is some topicour graduate students need to know about, and therefore we need toteach them, or Lila has a conversation with you about a topic near tolanguage. If she has had a conversation with you about such a topic,then you are at risk of coteaching a Graduate seminar with her on thattopic. (I think this has been especially true with me because having suchconversations for Lila is very much like teaching a course.)

So now you can, I think, see why there is a maximum number ofcourses for Lila. Lila is constantly engaging others intellectually—aboutlanguage (or bridge, or orchids), and it seems the natural thing that thisintellectual engagement be shared with students and colleagues.Casual conversations evolve into courses.

So what is it that I have learned from Lila? Two things. First, how notto have two careers, an intellectual career and a teaching career. Andsecond, the little linguistics and psycholinguistics I know. Thank youLila.

56 John Sabini

Chapter 5

Gleitology: The Birth of a New Science

Donald S. Lamm

My role here is an unusual one. I appear before an audience of learnedpractitioners of a discipline that in little over a hundred years has devel-oped an immense wingspread. And I address you fully aware that IQcontroversies or no, your intellectual prowess places you way out onthe right tail of the bell curve—right tail, let me emphasize, graphically,not ideologically, speaking.

Humility in such company should be the order of the day for onewhose highest achievement has been to serve as paymaster to thosewith a gift for giving words to ideas. I should listen, not speak. But onthis occasion, something compels me just this once to overrule my in-nate modesty, to play hedgehog among the foxes. Now is the time to re-veal that one big thing I know. For I have been witness to the birth of anew science.

It is a science that until this day has had no name, though hundreds ofthousands have read its laws, axioms, and postulates. Its domainranges from the slime to the sublime; its actors, coursing from A to Z,may be as hard to detect with the unaided eye as the amoeba or thezebra, vasodilating in the African sun. While rooted in scientific method,it partakes deeply of philosophy, literature, drama, sculpture, andpainting.

Until now, I alone in the world have known the new science by its oneand only name. Others watched it evolve over sixteen long years; someeven made significant contributions to its development. But, today, itis appropriate that the name of this new science be made public.

I give you Gleitology . . . and a very short history.The first sightings of Gleitology go back to Greenwich Village, New

York, in the mid-1950s. There, for what in Swarthmore circles passed fora Bacchanale, was the founder himself, Henry Gleitman, wreathed notin grape leaves but in smoke. As the mere escort of a Swarthmore grad-uate, I was entitled to only a brief audience. The founder produced aphoneme that I would later identify as “omm.” I was quickly parceledoff to a Gleitman acolyte from Wesleyan University.

Years were to pass, seven to be exact, before I was to hear that pho-neme again. But this time I was a man with a mission. After numerousdiscussions with George A. Miller, the magic number seven-plus-or-minus-two guru, and, more significantly, an advisory editor to Nortonin psychology, the name Gleitman appeared at the top of a short list ofpotential authors of an introductory text. “I know Gleitman from somearticles,” said Miller, and thereupon produced a Scientific American off-print on “place learning in animals.” The piece read extraordinarilywell, but what, I asked Miller, did this have to do with a psychologytextbook?

Miller instantly replied: “Gleitman’s psych 1 course at Swarthmore isreputed to be the best in the country.” That was enough to send me offon a semi-wild goose chase. Efforts to reach the great Gleitman when Iwas on an editorial trip to Philadelphia were unavailing. No doubt withhis usual prescience, he had decided that the real test is not having themountain come to Mohammed but having the mountain try to findMohammed. He had, in fact, taken a year away from Swarthmore andwas teaching at Cornell. I trekked up to Ithaca to find him.

Our meeting at Cornell is the one Henry Gleitman considers thealpha meeting, describing the Greenwich Village encounter as some-thing from his “primordial past.” It started badly. Within minutesHenry revealed that he had more suitors than Odysseus’s Penelope,publishing suitors that is. And even while disparaging them all for my-opia and assorted mental maladies, he set before me the Gleitmanequivalent of the MMPI. The test consisted of two fairly thick blue note-books, the syllabi for the best psych 1 course in the land. “You see,”Gleitman intoned, “there is so much to psychology that it would requirea two-volume textbook to encompass the whole field.”

I knew that the wrong response would have been to say “impossi-ble,” yet, in publishing terms, that would have been the right response.Should I, perhaps, have invoked the shade of William James? In his longdrawn-out struggle to produce the first psych 1 textbook, James hadrailed intermittently at his publisher, Henry Holt, for not putting atleast some of the work in type while he struggled with what he calledthe “demon.” Most likely fearing that James would insist on any pre-maturely typeset material being printed as volume one of the work,Holt would have none of it: “I will not set a word until I have it all.” Inthe 1890s, as in the 1960s, students were not likely to buy, let alone read,a multivolume text.

The better part of wisdom was to dodge the question. I knew that ateam of psychologists at the University of Michigan that seemed toagree on very little had convinced their publisher to help resolve their

58 Donald S. Lamm

differences by putting out a two-volume edition of their introductorypsychology text. (It turned out to be a colossal failure.) I bought timewith Professor Gleitman by agreeing that in the best of all possibleworlds any textbook he wrote should mirror his course syllabus.

That was not the end of the test. “What do I do about Freud?” the pro-fessor asked. And then, in what I would discover was trademark Gleit-man behavior, he supplied the answer to his own question. Freud wasnot in his syllabus. Even though Freud was recognized as a genius,doubts were widespread in the profession whether he should be admit-ted to the pantheon of great psychologists, whether, indeed, he was apsychologist at all. It turned out that Freud figured in the Gleitmancourse but was, literally, taught under cover of darkness, in two or threeevening classes where attendance was wholly voluntary.

I decided to chime in with an answer of my own. This time I could re-spond without hedging. “Freud must be in the text.” (What else could Isay as an editor at Norton, the publishers of the standard edition of theComplete Psychological Works of the man we irreverently referred to as“Old Whiskers”?) My reply produced a second Gleitman “Omm,” aclear signal that it was time to leave. He would think things over.

I was almost out the door, when Professor Gleitman (still not Henryto me) uttered a sentence that I would hear often again over the next fif-teen years, “Suppose the emperor has no clothes?” Then as now, play-ing dumb comes easily to me. He went on, “Apparently George Millerthinks that I may be the best psych 1 teacher in the country. What hap-pens if I write a textbook that does not live up to my reputation?” I musthave mumbled something semi-witty such as “Isn’t it up to the pub-lisher to dress the emperor?” I received my third “omm” and was gone.

Somewhere, as yet unretrieved, in the Norton archive in ColumbiaUniversity is a letter I wrote to Professor Gleitman after the Ithaca meet-ing. I asked if he would kindly send me a copy of that two-notebook syl-labus to share with George Miller. To the best of my recollection I addedsome airy persiflage about the great loss psychology would suffer if theGleitman lectures resided only in the collective memories of his stu-dents. The notebooks arrived two weeks later, a revealing note at-tached. It said, “Here they are.”

George Miller, who had previously shown signs of indifference to-ward the textbook component of his editorial advisership, never sentme a thorough critique of the Gleitman syllabus. He did remark in ahandwritten note from Oxford University, where he was on leave, thatthe syllabus covered the waterfront, without a psychological pebble un-turned. Was this the sine qua non of a successful text, he wondered.Perhaps it was magic in the classroom that made Gleitman a standout.

Gleitology: The Birth of a New Science 59

A year later Henry Gleitman came to the University of Pennsylvaniaas department chairman. That appointment was not likely to prompt adecision from him to write the text. Still, on a trip to Philadelphia, I de-cided to try my luck again, this time performing that single act that is adistinctive trait of publishers: I invited Professor Gleitman to lunch. Asit happened, he was pressed for time and chose to turn the tables on me,taking me to the faculty club. (I should point out that the psychologydepartment at the University of Pennsylvania had put out a guide toeating in Philadelphia, rating restaurants with letter grades. It had avery brief preface, to wit, “For purposes of comparison, consider theFaculty Club a ‘D.’” Grade inflation was not unknown even in 1965.)

Soon we were on a first-name basis. Over the pièce de résistance,potato chips, we hit on the formula that managed to break Henry’s re-sistance. I’d like to claim that it sprung entire from my brain. In truth, itwas Henry who suggested that, while he had no time to write a text-book, maybe he could do the equivalent for psychology of RichardFeynman’s acclaimed lectures in physics. Now here was something towork with. I remember seizing on the idea and proposing that we bugHenry’s classroom with a recording device. That spurred Henry on; inhis best Cecil B. DeMille manner, he raised the stakes to cameras andfilm. We decided on voice recording as an initial move.

Now came a revelation. It would be a risky and costly business torecord all eighty or so Gleitman lectures and then prepare transcripts ofthem. So, while agreeing in principle to the arrangement, I decided tobuild in a safety factor, selecting a six-foot five-inch, two hundredtwenty pound sales representative who, unbeknownst to Henry, wouldsit in on one of his lectures. This inconspicuous espionage operativecame back with a review, in effect, of a Broadway production: “Boffosmash! Gleitman’s definition of behavior nearly moved me to tears.And the high point of the lecture occurred when, after a tactical pause,he intoned, in his German-accented English, ‘Consider the rat.’ Secondslater, he dashed around the stage of the packed lecture hall, imitating arat navigating a T-maze grid only to get shocked as it neared its goal, thefood powder. The whole class broke out in applause.”

That report was enough to convince my colleagues on the editorialboard. A contract was drawn up on March 29, 1965, committing Nor-ton inter alia to the expenditure of $2,000 to cover the recording andpreparation of transcripts of the Gleitman lectures. This time there wasno hesitation of Henry’s part. He signed the contract. The experimentbegan.

I am unaware that any undergraduates in the psych 1 course duringthe 1965 fall semester knew that almost every word of their dynamic lec-

60 Donald S. Lamm

turer was being recorded for posterity. “Almost” must be stressed, sinceHenry, with his propensity for scampering around the stage, managedoccasionally to venture out of microphone range. The transcript of hislecture on insightful behavior, for example, disintegrated into a numberof sentence fragments, apparently the result of Henry’s strenuous effortsto portray one of Wolfgang Köhler’s chimpanzees on the island ofTenerife using the eminent Prussian psychologist as a “climb-upon-able”in a desperate attempt to grasp an otherwise unreachable banana.

Despite such setbacks, a substantial body of lectures had been tran-scribed by the end of the semester. I must admit that the transcriptswere something of a disappointment. Stunning passages of intellectualdiscourse, entertaining descriptions of experiments, even the occa-sional groan-evoking pun could not mask the discontinuities and di-gressions of extemporaneous speech. Henry acknowledged a new-foundempathy for the sometimes garbled syntax in the transcript of a DwightDavid Eisenhower press conference. Perhaps, he mused, the lectureswere proof merely that good storytelling with a Leipzig accent accountedfor his reputation.

This would not be the only time a touch of Gleitman despondencyclouded the enterprise. We agreed to a pause in the proceedings, sinceaside from the obvious fact that the transcripts would require consider-able doctoring to serve as a textbook, there were other demands onHenry’s time: notably, recruiting psychologists for the University ofPennsylvania to establish his department as one of the best in the nation.

The pause lasted for nearly four years while an onion-skin set of thetranscripts curled and faded on a radiator in my office. Never was theproject abandoned; instead, at a crucial moment between egg rolls andmoo shu chicken at the Mayflower Restaurant in Philadelphia, Henrystated there was simply no alternative: He would have to write the text-book from scratch. Oh, he added, perhaps he might steal an occasionalglance at the transcripts but they would not constitute much more thanelaborated chapter outlines. We agreed that a second experiment shouldbe undertaken, this time a whole chapter, perhaps as much as a sum-mer’s worth of labor, duly compensated for.

A few dining rituals were necessary before Henry actually sat downwith lined pads and typewriter. At one such meal, he observed that thefirst chapters of introductory psychology textbooks carried a lot of bag-gage—clumsy efforts to define psychology, brief synopses of fieldswithin the discipline, listings of careers in and outside academic psy-chology, truncated histories of the science that inevitably opened, ac-cording to Henry, “in the beginning was Wilhelm Wundt.” He had

Gleitology: The Birth of a New Science 61

decided that his would be the first introductory psychology textbookwithout a first chapter. It was an illusion that I had to accept and even tofoster, though murmuring that perhaps in imitation of some mathemat-ics texts there might be a chapter 0. A much happier note was struckwhen Henry said that he was determined to find overarching themesfor his book, themes that would demonstrate to students and colleagueswhat, in fact, made psychology hang together.

That task would not be easy. In 1969, Henry had delivered an addressto division 2 of the American Psychological Association in which hespoke frankly about psychology as a discipline with many perspectives:“In teaching the introductory course we sometimes prefer to blur thedistinctions and sweep the differences under a rug. But surely this dis-torts the subject matter. . . . If psychology is a frontier science, let us pre-sent the frontier as it is, with its brawls and its barrooms and even itsbordellos.”

Nonetheless, Henry persevered and delivered early in 1970 the pre-sumed first chapter of his text, “The Biological Bases of Behavior.” Itwas a mere 170 pages long, tracing the history of investigations into“why,” as he put it, “men and beasts behave as they do.” From Des-cartes to von Helmholtz to Sherrington and, ultimately, contemporaryfigures, Henry spanned the field, pausing en route to deliver livelyasides on such topics as the copulating behavior of the praying mantisas evidence of disinhibition. Length apart, my colleagues and I wereconvinced that Henry’s draft chapter contained most of the ingredientsfor a successful textbook. The academic reviewers confirmed our im-pression. While all the reviewers commented on the extraordinarylength of the chapter, one going so far as to credit Henry with creating“the finest textbook in physiological psychology” he had ever read, itwas Professor Allen Parducci of UCLA who put it best: “Whatever youdo,” he said in a telephone follow-up to his written report, “keep thatman writing.”

The instrument for doing just that was a new contract, drawn up onApril 28, 1970, with considerably more payment for the author upfront—and no reference to tape-recorded lectures. One clause in thecontract stood out from the standard boiler plate: “The publisher will utilize no fewer than 20 academic consultants to review the entire manu-script or portions thereof.” Little did I know that the academic readercount would reach 86 over the ten years Henry worked on the first edition.

With the time of testing behind us, Henry wrote two chapters onlearning in fairly rapid order. The files reveal no serious setbacks toprogress, although Henry did grouse in one letter, “Very deep in Pavlov.

62 Donald S. Lamm

What an unpleasant Russian peasant trying to be an echt Deutscher sci-entist.” Then came a serious bump in the road. One reviewer harshlycriticized the amount of history, what he referred to as “psychology yes-terday,” in the learning chapters. That remark threw the author and, tosome extent his editor, for a loop. For one of the hallmarks of the textwas to be its emphasis on the evolution of psychology, an approach thatHenry would eventually explain by a metaphor in his preface, “a river’swater is much clearer when it is taken from its spring.”

Over many exchanges in phone, letter, and personal visits, a decisionwas reached to modify, not eliminate, the historical component. At nopoint was the thematic structure of the book in danger. But if God is inthe details, then He ordained that some of the coverage of psychologyin its earliest decades would have to give way to recent research. At onepoint in reorienting the project, Henry wrote, “I’m beginning to under-stand that the relation between author and publisher has virtually apsychiatric status. I wonder how Shakespeare ever managed to write asingle play without a kind-hearted Norton editor to cheer him on (orwas there one?).”

For all the stürm und drang set loose by the severest critic of the earli-est chapters, it was a heady discovery that the pioneers of psychologydid not have to be sacrificed en masse, that Henry’s text would still bedistinctive in showing that a science not only builds on the work of itsfounding figures but that it also profits at times from adventures downblind alleys. While it would still be eight years before the manuscriptwas completed, the hallmarks of Gleitology were in place.

Over the long haul through sensation and perception, memory, cog-nitive thinking, personality, intelligence, psychopathology, and more,Henry’s endeavors were constantly supported by close readings of thedeveloping manuscript from first-rate psychologists. Inevitably, revi-sions were called for, and, while Henry had yet to penetrate the myster-ies of word processing, he had a secret weapon in his writing armory:the stapler. Blocks of copy would be moved about with the adroit ma-nipulation of scissors and stapler. (Henry never was one to sink to thelevel of a glue pot.)

Still, there was more to writing and revising than mechanical aids.Two reviewers took on roles that went far beyond encouragement. PaulRozin, a University of Pennsylvania colleague, was an agent provoca-teur almost from day one. During yet another culinary moment, thisone at a Philadelphia restaurant called the Frog (and rather overdedi-cated to its amphibian motif), Professor Rozin unwrapped the key tothe long-postponed opening chapter. The notion was to introduce whatHenry was to call the many faces of psychology through the subject of

Gleitology: The Birth of a New Science 63

dreams, a subject with a rich research component and also a significantappeal to any reader’s experience. Along with Henry’s former student,John Jonides of the University of Michigan, Professor Rozin also devel-oped a study guide that became a key ancillary to the text.

The other reviewer thoroughly dedicated to the enterprise wasProfessor Lila R. Gleitman. Balancing her own career in linguistics withthe raising of two daughters, the planting and pruning of flora in theGleitman greenhouse, and the feeding of visiting fauna in the Gleitmanmanse, Lila became a collaborator in the fullest sense in the develop-ment of the manuscript. Her name appears in the text as coauthor of thechapter on language; her influence was far more pervasive and, whencalled for, subversive. For Lila was the one person I could enlist in peri-odic campaigns to convince Henry that a textbook’s success oftenturned on decisions on what to leave out.

The assistance Henry received from various quarters did not alterone prevailing fact: The book that finally appeared in 1981 was stampedthroughout as a solo accomplishment. Henry was the writer as com-plete impresario, creating unusual, sometimes whimsical schematicdrawings for the artist to render in final form, assisting in the selectionof all the halftone illustrations, and ultimately suggesting that the coverand dust jacket art feature a sculptural masterpiece that, in his eyes atleast, bore a close resemblance to Henry himself: Michelangelo’s David.(By publication date, I myself had come to take on the aspect of a hyper-phagic rat depicted in a halftone in the third chapter of the text.)

Publication was anything but an anticlimax. The book was greetedwith immense acclaim, backed up by well over two hundred adoptionsin its first year. Even though it placed far greater cognitive demands onits readers than most competing textbooks, Gleitman’s Psychology founda home not merely in every Ivy League college but also in state univer-sities, liberal arts colleges, and, occasionally, in community colleges. Itcaused a number of other publishers to commission “high-end” text-books, thereby helping to achieve the ultimate aim of Gleitology—toraise the standards of instruction in the introductory course. And, de-spite all the newly bred competition, Gleitman’s text remained the onlyone to demonstrate that there was cohesion both within psychologyand between psychology and other fields of inquiry.

A successful textbook not only spawns imitators, it takes on an after-life. The first edition of Henry Gleitman’s Psychology was followed ayear later by a briefer edition entitled Basic Psychology, tracing the sametrajectory as William James’s Psychology in which a truncated version(dubbed by the publisher “Jimmy”) appeared shortly after the grandwork (or “James”) itself. While acts of compression had troubled Henry

64 Donald S. Lamm

when writing the original text, no such concern hampered the rapidcompletion of Basic Psychology. And, consistent with the etiology ofPsychology, the crucial decisions as to what and where to cut in order tocreate “Hank” (Norton’s code name for the briefer edition) were madeover a meal in a Chinese restaurant.

Three more editions of both versions of Gleitman’s Psychology havesince appeared. What began as an American textbook has now becomea world textbook, with substantial course use in the United Kingdom,Scandinavia, the Netherlands, Germany, Israel, Australia, and in uni-versities elsewhere around the globe. And while each revision entailssubstantial additions and alterations, the book continues to exhibitthree qualities—passion, power, and elegance—extolled in Henry’sdedication of the first edition:

To three who taught me:Edward Chace Tolman, to cherish intellectual passionHans Wallach, to recognize intellectual powerLila Ruth Gleitman, to admire intellectual elegance

As always with Henry Gleitman, no one could have put it any better.

Gleitology: The Birth of a New Science 65

               

   

Chapter 6

Children’s Categorization of Objects: The Relevanceof Behavior, Surface Appearance, and Insides

Elizabeth F. Shipley

To say I am grateful to the Gleitmans for a major part of my education isan understatement. Henry, as a new but not novice teacher at Swarth-more, revealed challenges, paradoxes, and the sheer fun of psychologyto this former physics major, as he still does today—especially withthose questions that begin “I’m puzzled.” Lila, as a new mother, re-vealed the complexities of language and first language learning, as wellas gaps in received wisdom, to me as a mother of preschoolers, as shestill does today—often with devastating humor. I thank them both.

As, guided by Lila, I looked at young children learning to talk, ques-tions from my undergraduate days resurfaced: Why do we partition theentities in the world as we do? Why are rabbits special in some way butnot things smaller than a breadbox? What are our beliefs about the rela-tions among classes of things? What does it mean that Floppsy is both arabbit and an animal? More generally, I began to wonder how chil-dren’s psychological categories of physical objects develop and whatdetermines for a child which classes of objects are categories and whichare not. I have found possible answers to these questions in NelsonGoodman’s (1955/1983) insights on induction, answers which I willsketch here. See Shipley (1993) for a more extensive discussion.

First, what are psychological categories of physical objects? They areclasses of objects characterized by three psychological properties: (i)Category labels are used for object identification, for instance, as an an-swer to the question “What’s that?” (see, e.g., Anglin 1977; Brown 1958;Shipley, Kuhn, and Madden 1983). (ii) Categories act as the range of in-ductive inferences. When we are told a rabbit has a secum we are morelikely to extend the property of secum-endowment to other rabbits thanto other things smaller than a breadbox (see, e.g., Carey 1985; Gelman1988; Holland, Holyoak, Nisbett, and Thagard 1986). (iii) Categorymembers seem to have a deep resemblance, they belong together andform what Murphy and Medin (1985) called a coherent class, a classthey characterize as “sensible,” “comprehensible,” “informative, use-ful, and efficient” (p. 289). I will use the term category to refer to classesof physical objects with these properties.

A developmental account of categories should provide answers to atleast three interrelated questions:

A. Why do members of a category act as the range of inductive in-ferences?B. What gives coherence to a set of category members?C. What determines whether or not a child considers an object tobe a member of a specific category?

In this chapter I will outline answers to these questions based uponpsychological essentialism (see, e.g., Gelman, Coley, and Gottfried 1994;Gelman and Medin 1993; Medin and Ortony 1989), then consider an-swers to the first two questions derived from Goodman’s (1983) conceptof entrenchment, and finally report two experiments relevant to en-trenchment and the category membership question.

Psychological Essentialism

Current popular answers to these questions invoke psychological es-sentialism. Psychological essentialism must be distinguished from thephilosophical position that categories have real essences. Psychologicalessentialism involves a belief in deep, perhaps unknown, commonproperties possessed by all members of a category. These propertiesconstitute the essence of the category. For example, a belief that an ani-mal’s kind, whether it is a tiger or a lion, is determined by its DNA is apsychological essentialist belief. Psychological essentialism also in-cludes the belief that the essence of category members causally ac-counts for their more obvious properties. The appearance of a tiger andits ability to learn might be attributed to its DNA. Belief in a commonessence is said to underlie inductive inferences over members of a cate-gory and the attribution of coherence to the set of category members(see, e.g., Gelman, Coley, and Gottfried 1994; Gelman and Medin 1993;Medin and Ortony 1989). Induction from a sample to a category is sup-ported by the inherent similarity, the essence, among members of thecategory. Coherence reflects the common underlying essence. Coheren-ce is further enhanced by beliefs in causal relations between the essenceand more obvious properties of category members. Finally, the beliefthat an entity possesses the essence of a specific category accounts for aperson’s assignment of the entity to that category.

What kinds of things have psychological essences? Do only biologicalkinds, such as dogs and roses, the most popular candidates for essence-based categories, have essences (Atran 1995)? Do natural kinds otherthan biological kinds, such as gold and water, have essences (see dis-cussion in Malt 1994)? Can artifacts have essences (see summary in

70 Elizabeth F. Shipley

Malt and Johnson 1992; Bloom 1996)? If induction is necessarily medi-ated by essences, then the fact that preschool children make inductiveinferences as readily over artifact kinds as over biological kinds (Gel-man 1988) suggests artifacts have essences—at least for young chil-dren. Carey (1995) maintains that for the young child everything with alabel has an essence because essentialism “derives from the logicalwork done by nouns” (p. 276). She claims “the child has a default as-sumption that count nouns are substance sortals” (pp. 276–277) andevery substance sortal has identity criteria. Hence, for Carey, it followsthat for every count noun known to a child the child has identity crite-ria that specify the properties that must be unchanged for an entity tomaintain its identity as an instance of a particular substance sortal. ForCarey, these properties constitute the essence of things named by thecount noun for that child. Note that Carey’s definition of essentialismdoes not specify a role for a causal relation between deep propertiesand surface properties, although such could be included among iden-tity criteria.

Work by Smith, Jones, and Landau (e.g., 1996) on the importance ofshape for physical object identification is relevant to Carey’s (1995) po-sition on psychological essentialism. The Smith et al. studies offerstrong evidence that for preschool children label assignment is tied toshape for novel nouns and some novel objects. If all count nouns haveessences then the essence of at least some novel objects with novel la-bels consists of the shape of the object, a conclusion at odds with thespirit of most writings on psychological essentialism. However, theSmith et al. findings are consistent with a less extreme version of essen-tialism provided the child is granted a bias to use shape for label assign-ment when no other information is available. Possibly the shape biascould prompt the child to look for other similarities among entities withthe same label and hence be the key to the discovery of the essence un-derlying a category (Landau 1994). In brief, some versions of psycho-logical essentialism, those that give learning an important role in thecontent of the essence, are consistent with known facts on children’snovel label learning.

Work by Keil (1989; Keil and Batterman 1984) with young children ondiscovery and transformation of properties of physical objects indicatesthat if essence determines category assignment, then it changes as thechild grows older. For instance, children were told about a raccoon thathad a smelly sac surgically implanted and its appearance altered to re-semble a skunk. Then children were shown pictures of a skunk and araccoon and asked what the altered animal was, a raccoon or a skunk.Kindergarten children believed it was now a skunk, 4th-graders be-lieved it was still a raccoon, and 2nd-graders were undecided. As the

Children’s Categorization of Objects 71

child grows older the essence apparently changes from characteristicproperties to something inherent in the animal.

If, for the child, everything labeled with a count noun has an essence,as Carey (1995) proposes, then the theoretical issues of interest are (a)what is the essence of members of a specific category, that is, what prop-erties are tied to identity change and identity persistence, and (b) howand why does the essence change? However, if not all referents of countnouns have an essence then an additional theoretical issue is (c) thecharacterization of those categories whose members have a psychologi-cal essence. Finally, (d) there is the theoretical issue of how a categoryacquires an essence.

For adults, empirical work probing the nature of psychological essen-tialism has cast doubts on the relevance of the essentialist position. Forinstance, Malt and Johnson (1992) found that both function and moresuperficial properties influenced category assignment of artifacts. Forthe natural kind water, Malt (1994) asked college students to judge dif-ferent fluids on four dimensions: percent H2O, typicality as an instanceof water, similarity to other fluids, and acceptability as a type of water.She reports that in addition to the expected essence, chemical composi-tion (H2O), the source, location, and function of a fluid determine theclassification of the fluid as water. Kalish (1995) found adults unwillingto grant that biological considerations could assign entities to a kind ofanimal category absolutely. For instance, sixty percent of his college un-dergraduate subjects thought it possible that an animal could not beproved to be of any specific kind. Such a finding is inconsistent with theessentialist belief that each individual animal has an essence that isunique to its kind. Of course, the weakness of a psychological essential-ist position for adults does not dismiss its possible usefulness in con-ceptualizing children’s understanding of categories.

In brief, psychological essentialism appears to account for the threepsychological properties of categories for children. The possession of acommon essence could account for children’s willingness to make in-ductive inferences over all members of a category based upon informa-tion about a small sample. Further, the possession of a common essencecould account for the coherence of the set of members of the same cate-gory and the assignment of an entity to a category. For most psycholog-ical essentialist positions it is necessary to flesh out the notion ofessentialism with the naive theories, ideas, and beliefs that could consti-tute the basis of the essence in order to account for specific identificationjudgments. That is, for a child to decide if Billy, an animal, is a sheep ora goat the child must marshal her knowledge of sheep and goats andcompare that knowledge with her knowledge of Billy. Such knowledgeof a category is considered central to the essence by most of those ad-

72 Elizabeth F. Shipley

vancing an essentialist account of children’s category development(see, e.g., Gelman et al. 1994; Gelman and Medin 1993; Gelman andWellman 1991; Keil 1989).

Entrenchment

Another way of accounting for a person’s readiness to make inductiveinferences about a category and belief in the coherence of categorymembers is to focus on the history of properties attributed to membersof a category by that person. That is, one might look to the history of in-ductive inferences about the category. Nelson Goodman (1955/1983), inhis analysis of the inductive inferences people actually make, introduceda concept he called “entrenchment.” Classes of objects can have en-trenchment and properties of objects can have entrenchment. Thegreater the entrenchment of a class, the more readily it acts as the rangeof an inductive inference. A newly learned property of a brown dog willmore readily be attributed to other dogs than to other brown animalsbecause the class of dogs has greater entrenchment than the class ofbrown animals.

The greater the entrenchment of a property, the more readily it is ex-tended from a sample to a class exemplified by the sample. Observing amarmot, a kind of animal, sitting on a newspaper and eating grass, weare more willing to attribute eating grass than sitting on newspapers toother marmots because the property of eating grass is better entrenchedthan the property of sitting on newspapers.

For Goodman, classes and properties gain entrenchment from theirinvolvement in inductive inferences. Making the inference Dogs bark en-hances the entrenchment of the class of dogs and the property of bark-ing; making the inference Sheep eat grass enhances the entrenchment ofthe class of sheep and the property of eating grass. A person’s greaterreadiness to make inferences about dogs than about brown animals canbe attributed to the greater number of inductive inferences that personhas made about dogs compared to brown animals. Similarly with theproperties eats grass and sits on newspapers, we have made more infer-ences about eating grass than about sitting on newspapers.

Goodman proposed that the relative entrenchment of one class or oneproperty compared to another depends upon the number of times anactual projection has been made about a specific class or a specific prop-erty. For example, the relative entrenchment of the category dog, com-pared to the class brown animal, depends upon the number of times aperson has made the inductive inference Dogs bark, plus the number oftimes he has made the inference Dogs are loyal, plus the numbers oftimes he has made each of any other inferences about dogs, compared

Children’s Categorization of Objects 73

to the number of inferences, again in the token sense, he has made aboutbrown animals. Goodman’s use of frequency of responses (actual pro-jections) is consistent with the psychology of the 1950s in which re-sponse frequency was the premier parameter.

However, the concept of “more inferences” can be interpreted in sev-eral ways. Given the difficulty of determining when someone actuallymakes an inductive inference, I have suggested that the number of ac-tual projections in the type sense provides a more useful measure of rel-ative entrenchment (Shipley 1993). That is, the number of differentproperties a person has attributed to dogs is the primary determiner ofthe relative entrenchment of the category dog for that person. This pro-posal leaves the contribution of the number of tokens of each projectionunspecified. By this assumption, the greater the number of differentproperties attributed to members of a class, the greater the entrench-ment of the class, and hence the more readily the class serves as therange of an inductive inference. This assumption makes a well-entrenched category correspond to what Markman (1989) has called arichly structured category.

If the projections of properties over a class of individuals were theonly source of entrenchment of the class, then all entrenched classeswould be familiar classes and inductive inferences would not be madeover unfamiliar classes. However, even young children readily makeinductive inferences over some types of unfamiliar classes, such as anovel kind of animal (Davidson and Gelman 1990). Goodman’s pro-posal on the role of entrenchment in induction can account for suchphenomena via inherited entrenchment.

Consider general inferences such as Each kind of animal has a character-istic diet. This kind of general inference is called by Goodman an “over-hypothesis.” It is an inductive inference over such hypotheses as Dogseat meat, Sheep eat grass, and Horses eat hay. Making such a general infer-ence leads to the entrenchment of what might be called a “parent kind”kind of animal and a “parent property” characteristic diet.

A parent kind has kinds as individual members. For instance, the par-ent kind kind of animal has as one member the class of dogs and as an-other member the class of horses. A parent property has as individualinstances specific types of the property; thus the parent property charac-teristic diet has as individual instances individual diets such as eats meatand eats hay. The entrenchment of a parent class is inherited by itsmembers; the entrenchment of kind of animal is inherited by each kindof animal, both by familiar kinds such as the class of dogs as well as bycompletely unfamiliar kinds such as the class of marmots. The en-trenchment of a parent property is inherited by each of its instances; theentrenchment of characteristic diet is inherited by eats meat and eats hay,

74 Elizabeth F. Shipley

as well as by such unfamiliar diets as eats bamboo. The inheritance of en-trenchment means that a novel kind of animal, such as the marmot, isan entrenched class for a person who has projected over-hypothesesabout kinds of animals but knows nothing about marmots except thatthey are a kind of animal. Similarly, novel instances of properties of fa-miliar types, such as an unfamiliar diet, become entrenched propertiesvia inheritance.

I have suggested (Shipley 1993) that the classes of physical objects aperson considers categories are well-entrenched classes for that personin Goodman’s sense of entrenchment. Our well-documented willing-ness to make inductive inferences over categories comes from their en-trenchment. Our belief that members of a category form a coherent classcomes from our readiness to make inductive inferences over the classbecause of its entrenchment. Our use of a category label to identify anobject carries with it the properties that the object possess by virtue ofits category membership; the use of a category label for identification isinformative of past projections (Brown 1958).

How can entrenchment account for the child’s acquisition of categoryknowledge? First, it must be emphasized that the entrenchment posi-tion presupposes that children believe what they are told by othersabout the various labeled classes in their world. So, for instance, chil-dren told Dogs bite people will project the property of biting onto alldogs, even if they have never seen a dog bite a person. Thus the pro-nouncements of authorities enhance the entrenchment of the men-tioned classes and properties.

In addition, the entrenchment account presupposes that the child isbiased to apply a name given to an object to other objects. It is necessaryfor a class to be labeled in some way by a person in order to serve as therange of an inductive inference for that person and to thereby acquireentrenchment. The shape bias literature (e.g., Landau 1994; Smith et al.1996) attests to the existence of this bias for completely novel objects(also Markman 1989).

How might a child acquire an entrenched category? Let us imagine achild encounters an ambiguous-appearing entity such as a sea cucum-ber (an irregular cylinder that resembles animal waste or an industrialby-product) in its natural habitat and is told its label but nothing more.Even if the child is unable to identity it as some kind of plant, animal,substance, or artifact, the child should be willing to apply that label tosimilar appearing entities (Landau 1994), but be reluctant to attributeproperties of this individual sea cucumber to other things called sea cu-cumbers because sea cucumbers have no entrenchment; for example,the child should be unwilling to conclude that sea cucumbers are typi-cally found underwater on the basis of one being observed underwater.

Children’s Categorization of Objects 75

Now suppose the child is told a few properties of sea cucumbers, Sea cu-cumbers are heavy, have potassium inside, are used to make soup, propertiesthe child will project over all sea cucumbers. As a result the class sea cu-cumber will gain entrenchment and the child should be more willing tomake inductive inferences about sea cucumbers, for example, morewilling to conclude that sea cucumbers are typically found underwater.If the child decides sea cucumbers are a kind of animal, whether on thebasis of authority or observation, then the class sea cucumber will inheritentrenchment from the parent class kind of animal and the child’s will-ingness to make inductive inferences about sea cucumbers will increasefurther. Her willingness to make the specific inference Sea cucumbers arefound underwater will be even greater if she has projected over-hypothe-ses about characteristic habitat, such as Each kind of animal lives in a spe-cial place. Thus the child’s history of inductive inferences can account forthe transformation of a class of labeled objects into a category capable ofsupporting inductive inferences. The projection of over-hypotheses canaccount for a greater readiness to support some inductive inferencesrather than others. Finally, the experienced coherence of a category canbe explained as a readiness to make inductive inferences over the cate-gory.

How does a child decide that a novel object belongs in a specific cate-gory? For the entrenchment account, an additional assumption is neces-sary: The properties previously projected over the category are used toassign a novel entity possessing those properties to the category. Thus anovel object would be identified as a sea cucumber if it has the proper-ties previously attributed to sea cucumbers, that is, is heavy, has potas-sium inside, and is used to make soup. Keil’s findings that children’sidentification judgments change with age can be accounted for bychanges in the relative entrenchment of different properties of the testobjects.

The hypothesis that the possession of entrenched properties previ-ously projected over a category determines the identity of an ambigu-ous object as a member of the category is tested in Study 1.

Entrenched Properties and Identification: Study 1

In the section above I attempt to argue that category entrenchment is aplausible alternative to psychological essentialism when induction andcategory coherence are considered. However, the essentialist positionhas strong appeal when the child’s task is to decide the identity, the cat-egory membership, of an object. From an entrenchment perspective Ipropose that those properties that have contributed to the entrench-ment of a category for a child are the most important properties in the

76 Elizabeth F. Shipley

child’s identification of an object as a member of the category. Thisstudy tests this hypothesis with animal stimuli.

First, it is necessary to determine the properties young children at-tribute to different kinds of animals in order to select properties that arelikely to be entrenched for young children. We can learn something of achild’s beliefs about a category by asking. In a preliminary study 12three-year-olds and 12 five-year-olds were asked to explain to a puppet,who claimed to be from another planet, certain “earth words” such as“dog,” “monkey,” and “animal.” The children’s responses primarily con-sisted of mention of properties. These properties were scored as eithersurface appearance properties, apparent in a guide book picture of acategory member (fur, a tail), or behavioral properties, not apparent inevery observation of a category member (barks, eats meat), hence nec-essarily projected properties. The latter are necessarily projected, be-cause not every dog encountered by a child has been observed to barkyet the children’s reports are in the form of generic statements “Dogsbark,” not statements limited to their experience such as “I’ve heardsome dogs bark” or “Sometimes some dogs bark.”

Behavioral properties predominated in the children’s responses, eventhough we were liberal in our counts of surface appearance properties.It should be noted that even three-year-olds can readily supply surfaceappearance properties when explicitly asked, for example, “What dodogs have?” Of the properties mentioned by three-year-olds, 83% werebehavioral, and for five-year-olds, 66% were behavioral. This findingsuggests that young children regard behavioral properties as more im-portant than appearance in determining the nature of an animal. It also

Children’s Categorization of Objects 77

Table 6.1Properties mentioned by child informants.

Informants

3-year-olds 5-year-olds

Properties #Ss Ave. #Ss Ave.

Diet 8 2.2 11 2.7Habitat 8 1.1 10 1.5Locomotion 10 3.4 12 3.8Sound 8 1.9 9 1.3

Note. #Ss is number of subjects out of 12 who mentioned a specific type of property.Ave. is the average number of animals who were said to possess the type of property,given that a property was mentioned.

indicates that behavior is more important than appearance in determin-ing entrenchment, providing we measure entrenchment by the numberof different properties projected over a category. While the reports of be-havioral properties are necessarily inductive inferences, reports of ap-pearance may, or may not, be inductive inferences. They may be meresummaries of past observations; when we look at a dog we see fur.

Certain types of behavioral properties were attributed to animals ofdifferent kinds by the majority of children (table 6.1). For the most fre-quently mentioned types of properties, diet and locomotion, if a childmentioned that type of property for one kind of animal, he or she mentioned it for other kinds of animals as well. In addition to diet andlocomotion, habitat and sound were often mentioned. For instance,three-year-olds reported that dogs bark and cats meow. Pilot work indi-cates young children know these two properties are properties of thesame type. Told “Horses neigh,” “Lions roar,” “Dogs bark,” and thenprompted with “And cats?” the children respond “Cats meow.” Such apattern of responses suggests preschool children have organizedknowledge of kinds of animals and their properties that can be consid-ered over-hypotheses. It should be noted that properties such as these,along with perceptual properties, were included by Keil (1989) in his“discovery studies” as characteristic features and could have guidedthe younger children’s identity judgments.

Using these types of properties, as well as behaviors unique to a spe-cific kind of animal, for example, wags his tail when happy, we selectedthree behavioral and three appearance properties for each of 12 kinds ofanimals and formed 6 pairs of animals. Pretesting with four-year-oldsestablished that each triad of properties, behavioral or appearance,identified the intended animal of the pair.

A puppet who went on a trip and encountered various individual an-imals formed the context for the six specific trials. On each trial the childwas told of an animal that looked like one kind of animal but acted likeanother kind of animal and was asked to identify the animal: “The pup-pet saw an animal that acts like a tiger. It eats meat like a tiger, and itroars like a tiger, and it climbs trees like a tiger. But it looks like a camel.It has humps on its back like a camel, and long eyelashes like a camel,and a long neck like a camel. Remember, it acts like a tiger but it lookslike a camel. What do you think it is? Is it a tiger or a camel?”

Each child considered six ambiguous animals and judged the identityof each one. In addition to camel-tiger, the pairs cat-dog, cow-pig, duck-monkey, chicken-elephant, and horse-snake were used with all subjects.Over subjects, the kind of animal mentioned first (e.g., tiger or camel)and the kind of animal whose appearance was described (e.g., tiger orcamel) were counterbalanced. Within subjects, the order of the six pairs

78 Elizabeth F. Shipley

was randomized. For each subject behavior was mentioned first forthree pairs, and appearance was mentioned first for the other pairs.

We ran three-year-olds and four-year-olds in two different condi-tions: in one condition photographs of the two alternatives were pre-sent on each trial, and in the other no pictures were present.

The children selected consistently on the basis of behavior, not ap-pearance: 65% of the choices were based upon behavior. An ANOVA onthe number of behavior choices with age, sex, and picture conditions asfactors yielded no significant factors and no significant interactions.

No child of the fifty-six who participated selected more frequently onthe basis of appearance than behavior. Seventy percent selected morefrequently on the basis of behavior. Within each of the four groups,three-year-olds and four-year-olds with and without pictures, a signtest on the number of subjects making a majority of behavior choiceswas significant at the 0.01 level or better. Behavior-based choices pre-dominated when only first trials were examined (64%) and for each ofthe six pairs considered separately. We also ran 12 four-year-olds withdifferent pairings of the stimulus animals based upon the frequency ofchoice in the original conditions. (The least frequently selected animals,duck and chicken, were paired, as were the most frequently selected an-imals, etc.) For five of these six new pairs the majority of choices werebased upon behavior. In brief, the identification of ambiguous animalson the basis of behavior rather than appearance is a robust phenome-non.

In sum, Study 1 shows that for preschoolers behavioral propertiesthat can be considered entrenched are sufficient for deciding the iden-tity of an animal when compared to appearance.

The role of appearance versus other kinds of properties in the deter-mination of the identity of an object has been studied more with arti-facts than with natural kinds (see, e.g., Gentner 1978; Keil 1989; Keil andBatterman 1984; Kemler-Nelson 1995; Smith et al. 1996). With artifacts,the question has been which determines identity: appearance or func-tion? The results have not been consistent and two recent carefully con-trolled studies with preschool children report apparently contradictoryresults (Kemler-Nelson 1995; Smith et al. 1996). Smith et al. point out adifference between the two studies that suggests a reconciliation of thefindings by consideration of entrenchment. When the novel object’s func-tion is a property that has been attributed to other objects and hencecould have entrenchment (draws lines, makes sounds), function deter-mines identity (Kemler-Nelson 1995); when the function of the novelobject has no history that could lead to entrenchment (a part’s move-ment forms an arbitrary pattern), appearance determines identity(Smith et al. 1996). Thus work on children’s identification of artifacts

Children’s Categorization of Objects 79

indicates that functional properties determine identity when these areentrenched properties, but not otherwise.

This result suggests that with animal stimuli behavioral propertieswill be more likely to determine identity the better entrenched they are.Preliminary results from a study with unfamiliar animals show en-trenched properties are more effective than nonentrenched properties,consistent with the artifact data.

Projected properties or essence: Study 2

We have shown that young children use behavior rather than appear-ance to determine the identity of an individual animal when the twotypes of information conflict. At least two interpretations of this findingare possible. The first is that children’s projected beliefs about categorymembers are sufficient for identification. The second possibility is thatthe behavioral properties are taken as evidence by the child of the un-derlying essence of the animal, and hence essence determines identifi-cation.

To evaluate this second interpretation of the data, that a commitmentto essentialism underlies the children’s choices, we went on to investi-gate the effect of the experimenter’s identification of the animal’s in-sides upon the child’s identification of the animal. The importance ofthe insides of an animal in the determination of identity from the per-spective of psychological essentialism has been argued by Gelman andWellman (1991) who found that the hypothetical removal of the insidesof an animal changed the child’s identification.

We first examined the effects of insides upon identification when in-formation on insides conflicted with behavior. Twelve four-year-oldsparticipated in this condition. The same context for the task, a puppeton a trip, and the same pairs of animals were used as in the originalstudy.

Each question specified that the animal had the insides of one kind ofanimal and listed three internal constituents of the animal kind. Thespecific internal parts were pretested to ensure that four-year-olds be-lieve them to be inside rather than on the outside of animals, and thatthey believe them to be inside the specific kind of animal they were at-tributed to.

As in Study 1, the child was asked to judge the identity of an animalreported by a traveling puppet: “The puppet saw an animal that actslike a tiger. It eats meat like a tiger, and roars like a tiger, and climbs treeslike a tiger. But it has the insides of a camel. It has the brain of a camel,and the lungs of a camel, and the bones of a camel. Remember, it acts

80 Elizabeth F. Shipley

like a tiger but has the insides of a camel. What do you think it is? Is it atiger or a camel?”

Again, which animal of a pair had behavioral properties mentionedand which animal was mentioned first were both counterbalanced oversubjects. For each subject, behavior was mentioned first on three trials,and insides were mentioned first on the other three trials.

Behavior was more important than the insides of an animal in the de-termination of identity. Sixty-five percent of the children’s choices werebased upon behavior, not internal parts (difference from chance p<0.02).This suggests that the child’s use of behavior to determine identity isnot due to a belief that behavior reflects the interior nature of the ani-mal, which in turn determines identity. Thus this pattern of responses isinconsistent with the child’s being an essentialist who believes theessence resides in the insides of animals.

In addition, we examined the child’s identifications when appear-ance and insides were in conflict. Twelve different children participated.We found that appearance was also more effective than insides in thedetermination of identity. Seventy-two percent of the choices werebased upon appearance (difference from chance p<0.002). This findingis apparently contrary to the results of Gelman and Wellman (1991) and,again, contrary to the position that children locate the essence in the in-sides of an animal.

The failure to identify on the basis of insides rather than appearancehas several possible interpretations. One is that children consider allbones, all blood, etc., as equivalent, so that the blood of a camel and theblood of a tiger are the same. Being told that an animal looks like a tigerbut has the blood of a camel provides no reason to identify the animal asa camel. A second possible reason for a lack of effect of insides is that thepreschool child is generally ignorant of insides.

Children do have some knowledge of insides. However, most studiesof preschool children’s notions about insides indicate children believedifferent ontological kinds, such as animals (e.g., a sheep) and artifacts(e.g., a machine), have different insides. However, there is little datasuggesting children believe instances of different categories of the sameontological kind, such as two animals—a camel and a tiger—have dif-ferent insides. For instance, R. Gelman (1990) asked preschool childrenwhat was inside various animates, such as a mouse and a person, andinanimates, such as a rock and a doll. Similar answers—blood, bones,various organs—were given to each of the animates. Inanimates of dif-ferent kinds also elicited similar answers, answers that were very differ-ent from the answers concerning animates. Simons and Keil (1995) alsoprovide data suggesting young children believe different ontological

Children’s Categorization of Objects 81

kinds, such as animals and artifacts, have different insides, although thechildren do not know what is inside the different ontological kinds.

Gelman and Wellman (1991) focused on specific kinds in their studiesof insides and essences. Preschool children considered individual enti-ties before and after their insides were removed and were asked if theentity with insides removed was the same kind and had the same prop-erties as the preoperated entity, for example, “Was it still a dog?” and“Could it still bark?” The children answered in the negative. However,this study may have revealed sensitivity to ontological distinctionsrather than kind distinctions. A dog with its insides removed seems moreakin to a collapsed balloon or an empty costume, a two-dimensional ob-ject, than to a three-dimensional living animal. From this perspective,the indifference of our subjects to insides does not conflict with theGelman and Wellman data although it casts doubts on their conclusionthat the insides of an animal contain its essence.

Discussion and Summary

I began with the commonly recognized characteristics of psychologicalcategories of physical objects: serving as the range of inductive infer-ences, possessing psychological coherence, and having a label that isused for identification. I sketched how the psychological essentialistsaccount for these aspects of categories. Then I indicated that Good-man’s (1983) concept of entrenchment provides an alternative accountof induction and coherence. To explain what determines children’sidentification judgments I suggested that they are based upon the bestentrenched properties attributed to members of a category. Study 1 sup-ports this hypothesis and Study 2 discounts a possible essentialist ex-planation of the Study 1 results.

Given that the entrenchment position and the essentialist positionseem comparable in their ability to account for these three characteris-tics of categories, what are the meaningful differences between the twopositions?

An essentialist account, which physically locates the essence, for in-stance, in the insides of biological kinds (Gelman and Wellman 1991),differs from an entrenchment account, which focuses on the history ofinductions, which may or may not concern insides. As Study 2 indi-cates, children apparently do not locate the essence of different kinds ofanimals in their insides. Of course, psychological essentialists couldclaim either that the insides are not the locus of the essence of an animalor that it was a mistake to assume that children assign any location tothe essence of an animal, and hence Study 2 is not a test of the generalessentialist position.

82 Elizabeth F. Shipley

An essentialist account, which relies upon an innate essence to unifya category (Atran 1995), is different from an entrenchment account,which relies upon the projection of beliefs. In the former case one wouldonly need to know that a class is a kind of animal to grant that class anessence. However, a history of inductive inferences, both inferencesabout specific kinds and inferences about the parent kind kind of animal,is required for a novel kind of animal to be an entrenched category.

But what of psychological essentialist accounts, which tie essence toknowledge that is not necessarily innate? It may be that the essential-ists’ noninnate knowledge that is relevant to a category’s ability to con-strain induction and account for identification is equivalent to projectedhypotheses about classes of individuals and projected over-hypothesesabout classes of classes. To support such a conjecture requires analysisof the types of information deemed relevant to an essentialist accountand examination of what classes gain entrenchment from various be-liefs, and perhaps experimental work as well. This analysis has not beendone, nor is this the place to attempt it.

The entrenchment account has four apparent advantages over essen-tialist accounts. The first advantage is that it subsumes both categoriesand properties under the same theoretical approach. Entrenchment ac-counts for which category most readily acts as the range of an inductiveinference, as can essentialism, but it also accounts for which propertiesare most readily projected, for instance the example given above of eatsgrass rather than sits on newspapers. The second advantage is that it pro-vides an explanation of the development of categories via the acquisi-tion of entrenchment through inductive inferences and inheritance. Thethird advantage of the entrenchment position is that it uses hierarchiesof parent kinds and individual kinds to make categories of novel classes.Thus a novel kind of animal such as the gerenuk will be considered aninduction-supporting category by someone who knows nothing of thegerenuk except that it is a kind of animal. The fourth advantage of theentrenchment position is the power it gives to hierarchies with parentkinds, together with over-hypotheses, to support inductive inferencesabout specific types of properties from minimal information. For in-stance, the diet of gerenuks can be inferred from observation of onegerenuk eating leaves.

Given the comparability of psychological essentialism and entrench-ment in accounting for the three primary characteristics of categories ofphysical objects and the apparent advantage of the entrenchment posi-tion in some additional respects, at this time the entrenchment positionwould seem to merit further elaboration as an alternative to psycholog-ical essentialism.

Children’s Categorization of Objects 83

Acknowledgments

This work was supported in part by NSF Grant BNS-8310009 and NSFGrant SBR-9414091. I am grateful to the children, teachers, and parentsof Trinity Cooperative Day Nursery, and Swarthmore Friends School.Thanks are due Barbara Shepperson for thoughtful assistance. Cor-respondence should be addressed to the author at Department of Psy-chology, University of Pennsylvania, 3815 Walnut St., Philadelphia, PA19104 or liz@cattell.psych.upenn.edu. Part of this work was presentedat the 1996 meeting of the Psychonomic Society in Chicago.

References

Atran, S. (1995) Causal constraints on categories and categorical constraints on biologicalreasoning across cultures. In Causal Cognition, ed. D. Sperber, D. Premack, and A. J.Premack. Oxford: Clarendon Press.

Anglin, J. M. (1977) Word, Object, and Conceptual Development. New York: Norton.Bloom, P. (1996) Intention, history, and artifact concepts. Cognition 60:1–29.Brown, R. (1958) How shall a thing be called? Psychological Review 65:14–21.Carey, S. (1985) Conceptual Change in Childhood. Cambridge, MA: MIT Press.Carey, S. (1995) On the origin of causal understanding. In Causal Cognition, ed. D. Sperber,

D. Premack, and A. J. Premack. Oxford: Clarendon Press.Davidson, N. S. and Gelman, S. A. (1990) Inductions from novel categories: The role of

language and conceptual structure. Cognitive Development 5:121–152.Gelman, R. (1990) First principles organize attention to and learning about relevant data:

Number and the animate-inanimate distinction as examples. Cognitive Science14:79–106.

Gelman, S. A. (1988) The development of induction within natural kind and artifact cate-gories. Cognitive Psychology 20:65–95.

Gelman, S. A., Coley, J. D., and Gottfried, G. M. (1994) Essentialist beliefs in children: Theacquisition of concepts and theories. In Mapping the Mind, ed. L. A. Hirschfeld andS. A. Gelman. Cambridge: Cambridge University Press.

Gelman, S. A. and Medin, D. L. (1993) What’s so essential about essentialism? A differentperspective on the interaction of perception, language, and conceptual knowledge.Cognitive Development 8:157–168.

Gelman, S. A. and Wellman, H. M. (1991) Insides and essences: Early understanding ofthe non-obvious. Cognition 38:213–244.

Gentner, D. (1978) What looks like a jiggy but acts like a zimbo? A study of early wordmeaning using artificial objects. Paper presented at the Stanford Child LanguageResearch Forum, April 1978.

Goodman, N. (1955/1983) Fact, Fiction, and Forecast. New York: Bobbs-Merrill Company,Inc.

Holland, J. H., Holyoak, K. J., Nisbett, R. E., and Thagard, P. R. (1986) Induction: Processesof Inference, Learning, and Discovery. Cambridge, MA: MIT Press.

Kalish, C. W. (1995) Essentialism and graded membership in animal and artifact cate-gories. Memory and Cognition 23:335–353.

Keil, F. C. (1989) Concepts, Kinds, and Cognitive Development. Cambridge, MA: MIT Press.Keil, F. C. and Batterman, N. (1984) A characteristic-to-defining shift in the development

of word meaning. Journal of Verbal Learning and Verbal Behavior 23:221–236.

84 Elizabeth F. Shipley

Kemler-Nelson, D. G. (1995) Principle-based inferences in young children’s categoriza-tion: Revisiting the impact of function on the naming of artifacts. CognitiveDevelopment 10:347–380.

Landau, B. (1994) Object shape, object name, and object kind: Representation and devel-opment. In The Psychology of Learning and Motivation, volume 31, ed. D. L. Medin.Orlando:Academic Press.

Malt, B. C. (1994) Water is not H2O. Cognitive Psychology 27:41–80.Malt, B. C. and Johnson, E. C. (1992) Do artifact concepts have cores? Journal of Memory and

Language 31:195–217.Markman, E. M. (1989) Categorization and Naming in Children. Cambridge, MA: MIT Press.Medin, D. L. and Ortony, A. (1989) Psychological essentialism. In Similarity and Analogical

Reasoning, ed. S. Vosniadou and A. Ortony. New York: Cambridge UniversityPress.

Murphy, G. L. and Medin, D. L. (1985) The role of theories in conceptual coherence.Psychological Review 92:289–316.

Shipley, E. F. (1993) Categories, hierarchies, and induction. In The Psychology of Learningand Motivation, volume 30, ed. D. L. Medin. Orlando:Academic Press.

Shipley, E. F., Kuhn, I. F., and Madden, E. C. (1983) Mothers’ use of superordinate cate-gory terms. Journal of Child Language 10:571–588.

Simons, D. J. and Keil, F. C. (1995) An abstract to concrete shift in the development of bio-logical thought: the insides story. Cognition 56:129–163.

Smith, L. B., Jones, S. S., and Landau, B. (1996) Naming in young children: A dumb atten-tional mechanism? Cognition 60:143–171.

Children’s Categorization of Objects 85

Chapter 7Mechanisms of Verbal Working Memory Revealedby Neuroimaging Studies

John JonidesIn 1970, when I had the privilege of beginning my work with HenryGleitman, the study of cognition was in the midst of a vital period. Thereasons for this vitality were many, but among them was the vision thatperception, memory, language, and thinking could be understood bydecomposing cognitive processes into their essential elements. Thetools for this decomposition were chronometric analysis and the analy-sis of patterns of errors that subjects made in various tasks, both ofwhich were being applied to a host of problems in the study of cogni-tion. Henry and I embraced these developments and used these tech-niques to study processes of item-recognition and short-term memoryin a series of studies that we conducted under the watchful eye ofHenry and Lila Gleitman’s Thursday evening (and often late-night) re-search seminar.

The strategy of understanding complex cognitive processes by de-composing them into their simpler parts continues to be widely taughtand learned as an essential skill for students of cognition. At the time ofmy graduate education, implementing this strategy most often in-volved the careful design of behavioral experiments that would revealunderlying processes in the chronometric or errorful data of subjects. Inthe hands of clever psychologists such as Henry and Lila, this strategyyielded insights into a wide variety of cognitive processes. In the ensu-ing years, the strategy has been broadened while still maintaining itsessence. The first significant broadening involved the widespread useof computer models as ways of further defining and refining concep-tions of elementary processes and how these combine to yield cogni-tion. The second significant broadening is the one that is the currentfocus of much interest among students of cognition: the use of neu-roimaging techniques to reveal brain processes that mediate cognitiveprocesses. These techniques allow us to go well beyond merely localiz-ing processes to various brain areas; in combination with behavioraldata on normal and brain-injured humans and in combination with suit-able data from invasive studies of animals, neuroimaging studies can ac-

complish the same goal that we had in 1970: to decompose cognitiveprocesses into their elementary components. In short, neuroimagingtechniques provide another modality of data for understanding cogni-tion.

To illustrate how the use of neuroimaging has enriched the study ofcognition, I shall briefly review the results of a program of research con-ducted by Ed Smith, me, and our colleagues in recent years. We havebeen investigating what is now called working memory. The reason forfocusing on working memory is simple: Various lines of evidence makeit clear that working memory is an essential component of cognition inthat it participates in such skills as problem solving, reasoning, catego-rization, and language comprehension. Indeed, an individual’s work-ing-memory capacity (if measured in the right way) has been found tocorrelate with a large variety of complex cognitive skills (see, e.g.,Carpenter, Just, and Shell 1990; Daneman and Merikle 1996; Kyllonenand Christal 1990). Furthermore, it has been shown repeatedly that adecline in working memory with normal aging and with various brainpathologies is an important predictor of declines in performance inproblem solving and reasoning, showing again that working memory iscritical to cognitive life (see, e.g., Salthouse 1993). In short, the phenom-ena of working memory are examples of the sort of “big effects” that theGleitmans advocated as worthy topics for intense research.

We define working memory in a canonical way. It is the memory sys-tem that keeps active a limited amount of information for as long as onecontinues to work with that information. The information in workingmemory is easily and readily accessible, and it is subject to frequent up-dating or substitution by new information. Finally, the feature that dis-criminates the concept of working memory from earlier conceptions of“short-term” or “primary” memory is that information held in workingmemory is subject to processing in various ways by what Baddeley(1986, 1992) and others have called executive processes. It is in thissense that working memory goes beyond mere storage; it involves themanipulation of information that is stored as well.

Our research group has devoted itself to several issues concernedwith working memory; here we review research concerned with two ofthese. One concerns the architecture of the working memory system it-self and whether this system consists of several subcomponents con-cerned with storage, rehearsal, and manipulation of information that isstored. On this issue, much of our research has concentrated on work-ing memory for verbal information. A second issue is whether workingmemory is a single system or whether it consists of several subsystemseach tied to the processing of different sorts of information.

88 John Jonides

The Architecture of Verbal Working Memory

In his influential model of working memory, Alan Baddeley (1986, 1992)proposed that working memory for verbal information consists of threecomponents. One is a buffer responsible for the temporary storage ofverbal codes. A second is a rehearsal mechanism that recirculates infor-mation in the verbal buffer for the purpose of preventing decay or inter-ference. A third is a set of processing mechanisms, collectively called thecentral executive, that are capable of manipulating the informationstored in the buffer (including rehearsing it). Evidence for a dissociationbetween storage and rehearsal processes has come largely from behav-ioral studies of normal and brain-injured subjects. Although this evi-dence is compelling, it is not decisive (see Jonides et al. 1996 for areview), and so we have conducted experiments using neuroimagingtechniques to try to identify the subcomponent processes of workingmemory.

One experiment makes use of a paradigm that recruits all the com-ponents of working memory (storage, rehearsal, and executive), the“n-back” task. In this task, subjects are presented a series of single let-ters; as each is presented, subjects must decide whether it matches theone that was presented n items back in the series. In an experiment fromour laboratory reported by Awh, Jonides, Smith, Schumacher, Koeppe,and Katz (1996), n was set to 2. A schematic of the memory condition ofthe task is shown in the top panel of figure 7.1. The panel illustrates thatletters were presented for 0.5 sec. each with 2.5 sec. intervening betweensuccessive letters; subjects engaged in the memory task for a continuousperiod of approximately 60 seconds while they reclined in a PET scan-ner. Note that successful performance in this task requires storing inmemory a constantly changing set of at least two letters, the “oldest” ofwhich must be compared with the currently presented letter. Thus thetask requires both storage of verbal information as well as processesthat must update this information continuously.

Of course, performance in this memory condition also includes pro-cesses of perception and response, processes that were not the targets ofour interest in this experiment. The typical strategy for eliminating theeffects of such processes on brain activations in a neuroimaging experi-ment is to test subjects in a second condition that includes only these an-cillary processes, by hypothesis, and then to subtract the brainactivations from such a control condition from the condition of interest.We recognize that this “subtraction” strategy has earned some well-deserved criticisms, and we address this issue below. Nevertheless, wefollowed this strategy by also testing subjects in a memory controlcondition, shown in the second panel of figure 7.1. In this condition,

Mechanisms of Verbal Working Memory 89

subjects were presented with a series of letters, but they responded pos-itively only if each letter matched a fixed target letter for which theysearched on that trial (say, “P”). Activations from this control conditionwere subtracted from those in the memory condition. The results re-vealed activation in anterior parts of the cortex: in Broca’s, supplemen-tary motor, and premotor areas. In addition, there was activation inposterior parietal cortex in the superior parietal lobule and supramar-ginal gyrus.

The anterior sites that were activated in this task have been claimedto be part of a circuit (including the cerebellum, which also showed ac-tivation) that is responsible for rehearsal (Paulesu, Frith, and Frack-owiak 1993); indeed, this is quite sensible in that some or all of thesesites are involved in the production of explicit speech as well. The pos-terior sites have been implicated in storage and selective attentionprocesses required in this task. How can one confirm these putativefunctions? Paulesu et al. (1993) provided some evidence for the re-hearsal function of the anterior sites by showing that these sites were

90 John Jonides

Figure 7.1.Schematics of the three tasks used in the experiment by Awh et al. (1996).

also activated in another task in which judgments of rhyming were required, judgments that presumably also engage internal language-production mechanisms (see the chapter by Reisberg in this volume foradditional evidence about such judgments). Furthermore, the site wefound activated in the supramarginal gyrus of parietal cortex is themost common site of damage in patients who suffer deficits in verbalmemory span (McCarthy and Warrington 1990), suggesting that it maybe involved in storage of verbal material. Also, the site in the superiorparietal lobule has been implicated as a region involved in shifts of at-tention from one item of material to another in many tasks, as shown bya parallel between this site and sites that are activated in tasks in whichexplicit shifts of attention are required (see Awh and Jonides 1999 for areview). So there is circumstantial evidence for the functions that weclaim.

Beyond this, Awh et al. (1996) included a second rehearsal controlcondition in their experiment that allowed them to test for the rehearsalfunction of the anterior sites. It is schematized in the third panel of fig-ure 7.1. In this condition, subjects were again shown a sequence of indi-vidual letters, and they were instructed to rehearse each one silentlyuntil the next one appeared; then they were to rehearse the next one,and so on. This condition requires internal speech, and so subtractingthe activation due to this condition from that due to the memory condi-tion should diminish activation in anterior sites but not in posteriorones if the anterior ones in the Memory condition are, indeed, responsi-ble for rehearsal. In fact, this is just what Awh et al. (1996) reported.Thus it appears that experiments using neuroimaging technology canprovide corroborating and converging evidence for claims rooted in be-havioral data as well as provide information on the brain areas that arethe substrates of working memory.

Note that the experimental rationale underlying the application ofneuroimaging techniques owes a great debt to the experimental ratio-nale first articulated by Donders (1868) for the study of reaction times, arationale discussed and debated at great length in our research semi-nars some one hundred years after Donders described it. Donders ar-gued that if one could construct two tasks such that the first containedall the processing components of the second plus an additional one,then the difference in reaction time between the two tasks should be arelatively pure measure of the time required for the additional processrequired by the first task. In a similar fashion, much neuroimaging re-search has relied on subtracting the activation of one or more controltasks from an experimental task of interest, as illustrated above, to re-veal activation due to processes of interest that are required by the ex-perimental task. Just as the rationale due to Donders has been called

Mechanisms of Verbal Working Memory 91

into question, so also can one raise questions about the validity of as-suming that the activations due to selective processes can be subtractedout of a neuroimaging experiment without affecting activations due toother processes (see Jonides et al. 1997). The issue here is much the sameone that occupied us in the 1970s: finding a way to isolate in data the ef-fects of certain variables that have an impact on experimental per-formance. In the 1970s the data that concerned us on this score were reaction times and errors; in neuroimaging research they are patterns ofbrain activation that result from some task.

One experimental strategy that goes beyond the subtraction methodrelies on parametric variation of some variable of interest (in someways, mimicking the strategy first applied by Sternberg (1966, 1969) tothe analysis of chronometric data). We have implemented this strategyfor the study of the components of verbal working memory in severalexperiments (Jonides et al. 1997; Braver et al. 1997; Cohen et al. 1997).Our work is based on the paradigm illustrated in figure 7.2. The figurereveals various versions of the n-back task, in which n is varied system-atically. In the most demanding version, shown at the top of the figure,subjects must decide whether each letter matches in identity the onethat appeared 3-back in the series. We also included 2-back and 1-backversions in other conditions, and we included a 0-back condition simi-lar to the condition used by Awh et al. (1996) as a control; in the 0-backcondition, subjects were provided a single letter-target at the beginningof a trial, and they were to respond positively any time that letter ap-peared anywhere in the series that was presented on that trial. The de-sign shown in figure 7.2, therefore, implements a systematic variation oftask load in a working memory task. We collected PET scans for each ofthe conditions shown in figure 7.2, and we also collected PET data for abaseline control condition not shown in the figure in which a series ofletters was presented, and subjects simply responded with a keypresswhen each letter was shown; there was no memory requirement at all.This baseline control condition served as a way to subtract from each ofthe memory conditions the activation that was due to idiosyncratic dif-ferences in brain activity among subjects. The main comparisons in theexperiment, however, did not rely on subtraction methodology; ratherthey relied on a comparison of activation in various regions across vari-ations in task load.

Figure 7.3 reveals that, as expected, increases in task load produce adecrement in performance. This decrement appears as an increase in re-sponse time as well as a decrease in accuracy. The decrement in perfor-mance is accompanied by strikingly parallel changes in brain activationwith task load at each of several sites, as shown in figure 7.4. The data inthis figure were accumulated by taking regions of activation and deacti-

92 John Jonides

vation (that is, where the control condition shows less activation thanthe experimental condition and where it shows more activation, respec-tively) that had been previously identified with verbal working mem-ory tasks (from the studies of Awh et al. 1996 and Schumacher et al.1996). We then found levels of brain activation in the present data foreach of these regions. The average activation in each of these regionswas then plotted as a function of task load, and this is what is displayedin figure 7.4. The most striking feature of the data in this figure is thatthere is an overwhelming tendency for brain activation to increase

Mechanisms of Verbal Working Memory 93

Figure 7.2.Schematics of the tasks used by Jonides et al. (in press).

monotonically as task load increases, and for deactivation to decreasemonotonically with task load (confirmed by statistical analysis: Jonideset al. 1997). Thus there is reason to suspect that the activations and de-activations that are shown in figure 7.4 are systematically related to thememory requirements of the task. Note that these changes occur inmany regions, as the functions in figure 7.4 reveal. The presumption isthat increases in activation reflect increased brain activity that is re-quired by a task; likewise, decreases in brain activation in selected re-gions reflect requirements for inhibition of brain activity in thoseregions. The specific account of which regions increase and which de-crease is beyond the scope of this review. Here we merely highlight thatbrain activation in many regions is systematically related to the require-ments of the task.

Of course, one might argue that the variation in brain activation withtask load shown in figure 7.4 is merely a reflection of increased overalleffort as task load increases, and not selectively related to processeshaving to do with the memory requirements of the tasks per se. This ar-gument is laid to rest by examining other areas of the brain that should

94 John Jonides

Figure 7.3.Reaction time and error data from the experiment of Jonides et al. (in press) plotted as afunction of memory load.

not be recruited by memory processes: occipital areas that are involvedin visual processing, somatosensory areas that are not relevant to thetask, and primary motor areas whose activation should not vary withtask load. The activations for these regions as a function of task load aredisplayed in figure 7.5, which shows that there is no systematic varia-tion in brain activation in these regions as the memory task increases indifficulty. The contrast between the functions in figure 7.5 and those infigure 7.4 suggests that the areas identified in figure 7.4 do, indeed, re-flect memory-sensitive processes that are active during these tasks.

Note that the outcomes of parametric studies of working memoryvalidate findings reported using the subtraction methodology based onthe logic of Donders. That is, the various regions of activation shown infigure 7.4 and in other studies using parametric techniques are just theones, by and large, that show significant activation in subtraction para-digms such as the one described above from our laboratory (Awh et al.

Mechanisms of Verbal Working Memory 95

Figure 7.4.Brain activations (above the horizontal line) and deactivations (below the line) for memory-related areas plotted as a function of memory load from the experiment ofJonides et al. (in press). Each function corresponds to one region of activation (labeled byits number according to the system of Brodmann or by anatomical structure). The identi-fied regions are those that showed significant activation in a previous study using a sim-ilar task (see text).

1996). Thus, although one must exercise caution in using subtractionlogic and in interpreting the results of such experiments, the outcomesof experiments on working memory that have used this logic seem to bereplicated in experiments with parametric experimental manipulationof relevant variables.

Another contribution of the parametric method is that it provides anopportunity to examine details of the “dose-response” curves that re-sult, such as those shown in figure 7.4. We have exploited this propertyof parametric designs both in the experiment described above and inexperiments using this paradigm with functional MRI as the imagingmodality (Braver et al. 1997; Cohen et al. 1997). Functional MRI mea-surements permit a more detailed examination of brain activation pat-terns because they provide somewhat greater spatial resolution as wellas an opportunity to examine the temporal dynamics of processeswithin a single experimental trial.

Consider, for example, an experiment in which we varied memoryload in an n-back task while stretching out the retention interval on each

96 John Jonides

Figure 7.5.Brain activations for motor, visual, and somatosensory areas plotted as a function ofmemory load from the experiment of Jonides et al. (in press). The areas shown in the fig-ure were identified by placing regions-of-interest on primary visual, motor, and so-matosensory areas of the brain and calculating the activations in those regions.

trial so that we could examine the dynamics of activation in variousbrain regions during that retention interval (Cohen et al. 1997). The de-sign is much like that displayed in figure 7.2, but the delay interval be-tween successive letters in each condition was increased to 10 seconds.This permitted us to collect four scans of the entire brain, each one oc-cupying a 2.5-second interval during the time between letters. Thuslevel of brain activation could be assessed four times during the intervalwhen subjects were engaged in the working memory task.

The value of this technique is revealed by examining the results of theexperiment. Examine first figure 7.6, which shows the activations thatwere obtained in a region of extrastriate occipital cortex. The four func-tions in the figure correspond to the four conditions of memory load (0-,1-, 2-, and 3-back). Each function has four points, each corresponding toan activation for one of the four recording intervals during the retentionperiod. Note the four functions lie fairly close to one another and thatthere is little systematic effect of memory load. However, the four func-tions are also all noticeably bowed. This can be taken to mean that astime passed during the retention interval, the amount of activation in-creased in this region of occipital cortex and then declined. Why shouldthis be so? The most reasonable interpretation is that these functions allreveal the activation that was caused by presentation of a stimulus let-ter. The activation rises over the course of the first 7.5 seconds of the

Mechanisms of Verbal Working Memory 97

Figure 7.6.Activations from a region of extrastriate occipital cortex as a function of the time ofrecording during retention interval (from the experiment of Cohen et al., in press).

retention interval because the hemodynamic signal that is recorded byfMRI (corresponding to the neural signal that is tied to the increase inblood flow) is delayed relative to the neural event that causes it. By thisinterpretation, the functions in figure 7.6 reveal that fMRI recordingscan be sensitive to early processes of encoding, and that these processesare not particularly sensitive to the memory load of the task.

Examine now the activations in figure 7.7. This figure shows func-tions analogous to those of figure 7.6, but for activations recorded in aregion of dorsolateral prefrontal cortex in the right hemisphere. Notethat the pattern of data in this figure differs markedly from that in figure7.6. First of all, the functions are not at all bowed in shape. That is, acti-vations for all levels of memory load appear to be quite steady through-out the retention interval, not transient, as they are in occipital cortex.Notice also that in prefrontal cortex there is a dramatic effect of memoryload on activation level. As the figure indicates, the 2-back and 3-backtasks resulted in substantially more activation throughout the retentioninterval than did the 0-back and 1-back tasks. The sustained nature ofthis activation and its sensitivity to memory load suggest that this areaof prefrontal cortex is somehow involved in the maintenance of repre-sentations in working memory. This may be via direct storage of in-formation or via some more indirect role. For example, regions in

98 John Jonides

Figure 7.7.Activations from a region of dorsolateral prefrontal cortex as a function of time of record-ing during the retention interval (from the experiment of Cohen et al., in press).

prefrontal cortex may serve as pointers to other regions of the brain(possibly posterior regions) that are the sites of information storage.Whatever the specific role of these prefrontal structures, it is clear fromdata such as those of figure 7.7 that the activation of prefrontal cortexduring the n-back task is not transient in nature. It might have been so ifprefrontal cortex were involved strictly in executive functions such asupdating the contents of working memory or temporally tagging lettersas they are presented. In both of these cases, one would have expectedthe activation to be transient in nature, quite different from the data thatwere obtained.

These data indicate the potential that neuroimaging studies have forextending our knowledge of cognitive mechanisms. In connection withthe sort of behavioral data that one can gather from working memorystudies, neuroimaging data are proving helpful in specifying the vari-ous component mechanisms that contribute to complex cognitive phe-nomena.

Differing Subsystems of Working Memory

A central issue in the study of working memory has been whether itconsists of a unitary processing system or whether it is composed ofseveral subsystems. One line of evidence that has been illuminatingabout this issue comes from studies of brain-injured patients. There isevidence, reviewed in detail by Jonides et al. (1996), among others, thatthere are multiple subsystems of working memory defined by the typeof information that is maintained. For example, there are reports of pa-tients who have deficits in verbal working memory with no deficits inworking memory for spatial information; by complement, there is a re-port of a patient who has a deficit in spatial but not verbal workingmemory.

In support of this distinction, there is also evidence from strictly be-havioral studies of a dissociation between subsystems of workingmemory defined by the type of information that is processed. The be-havioral technique that has been used to provide evidence for this claiminvolves experiments in which subjects engage in one or another work-ing memory task while a second (presumably interfering) task is per-formed. The logic of these experiments is this: If secondary tasks can befound that require the use of one or another internal code for informa-tion that is processed, then they should selectively interfere with a pri-mary working memory task to the extent that that primary task makesuse of the same code. For example, a secondary task that engages aphonological code should interfere with verbal working memory if ver-bal working memory also requires the use of a phonological code; but it

Mechanisms of Verbal Working Memory 99

should not interfere with spatial working memory if that subsystemuses a code that is not phonological. Likewise, a secondary task thatmakes use of a spatial code should interfere with spatial working mem-ory but not verbal working memory. Various experiments have imple-mented this rationale (see, e.g., Meudell 1972; Salthouse 1974; Logie,Zucco, and Baddeley 1990; Logie 1995), and they have led to the viewthat working memory is, indeed, composed of several subsystems de-fined by the type of information that is processed.

These lines of neuropsychological and behavioral evidence are notimmune to criticism, however. Experiments on neuropsychologicalpopulations about the dissociation of different working memory sub-systems are limited to precious few patients, who are often tested ontasks that may not purely recruit one or another working memory sub-system. As for selective interference experiments testing normal sub-jects, it is often difficult to justify the assumption that a secondary taskhas a truly selective interfering effect on a primary task. There may beseveral sites of interference (see Jonides et al. 1996 for a detailed discus-sion). Consequently, we sought a line of evidence from neuroimagingstudies of working memory that might converge with the behavioralevidence to address the issue of whether working memory is best con-ceptualized as a set of subsystems rather than a single system of infor-mation processing.

The details of our experiments are reported elsewhere (Jonides et al.1993; Smith and Jonides 1995), but it is instructive to examine one ex-ample to see how neuroimaging evidence can strengthen the case forseparable working memory subsystems (Smith, Jonides, and Koeppe1996). Consider the pair of tasks illustrated in figure 7.8. Each panelshows schematics of the events on typical trials of the spatial and verbalmemory conditions respectively; both conditions involve a 3-back task,with the nature of the memorandum differing between conditions. Inthe spatial memory condition, subjects saw a stream of letters with theletters appeared in varying locations on the screen. The subjects’ taskwas to answer positively (via a button-response) if a letter’s locationmatched the location of the one that appeared three back in the se-quence; if not, they were to answer negatively (via another button-response). Similarly, the verbal memory condition also required sub-jects to match the current stimulus to the one that was 3-back in the se-quence; however, in this case they were to match letters on theiridentities regardless of their spatial locations. Thus, in both tasks, sub-jects had to keep in working memory information about several previ-ous stimuli, they had to match the current stimulus against the one thatappeared 3-back in the sequence, and they had to update the contents oftheir memories with each succeeding stimulus presentation. The major

100 John Jonides

difference between the conditions was in the sort of information thatwas stored in memory: In the spatial memory condition it was loca-tional information, and in the verbal memory condition it was identityinformation (storing a visual code for each letter would not suffice inthe verbal memory condition because the case of the letters haphaz-ardly varied from upper to lower).

We collected data for each of these two memory conditions whilesubjects reclined in a PET scanner. The scanning recorded all the brainareas that were activated during any portions of these tasks. To focus onthe processes specifically involving memory, we also tested subjects incontrol conditions whose activations were then subtracted from thoseof the memory conditions. In the spatial control condition, subjectswere shown three locations on the screen prior to a sequence of stimulisuch as those shown in the top panel of figure 7.8, and they were to re-spond positively anytime any letter appeared in any of these positions;otherwise, they responded negatively. Likewise, in the verbal controlcondition, they were shown three letters, and they responded positivelyanytime any of these appeared.

The brain activations that resulted from the subtractions of the con-trol from the memory conditions revealed both substantial overlap in

Mechanisms of Verbal Working Memory 101

Figure 7.8.Schematics of the memory tasks used by Smith, Jonides, and Koeppe (1996).

activations between conditions and substantial differences. Generallyspeaking, there was bilateral activation in both spatial and verbal tasksin both anterior and posterior regions of the brain. Beyond this,though, there was evidence that the spatial task activated some struc-tures in the right hemisphere more than in the left; in a complementaryway, the verbal task activated regions in the left hemisphere more thanin the right.

Other than this noticeable difference between the tasks, there wasconsiderable similarity in the regions activated. There was clear activa-tion in two regions of posterior parietal cortex, one more lateral than theother, similar to some of the results described above. Also, there was ac-tivation in dorsolateral prefrontal cortex, also similar to results pre-sented above. And there was evidence of activation in inferior frontalgyrus in the left hemisphere in the verbal task, an indication of the in-volvement of verbal rehearsal in this task.

For present purposes, our interest in this experiment is in its demon-stration of different patterns of activation as a function of whether thematerial to be retained was verbal or spatial in nature. Although theoverall circuitry revealed in these two conditions was similar, therewas, as noted above, a dissociable pattern of activation with spatial ma-terial engaging the right hemisphere more than the left and verbal ma-terial engaging the left more than the right. Note that this pattern wasobtained in an experiment in which the physical stimuli were nearlyidentical in the two memory conditions, rendering it unlikely that thedifferent patterns of activation were a function of perceptual process-ing. We conclude that the different patterns are instead a reflection ofdifferent underlying circuitry for spatial than verbal working memory.This result, then, confirms and extends the data from neuropsychologi-cal and behavioral experiments and adds currency to the hypothesisthat working memory is best characterized as a set of subsystems, eachresponsible for the processing of different sorts of information.

A Closing Reflection

Experiments such as those I have described are in the forefront of thenews in cognitive psychology these days. They are there because neu-roimaging techniques make it seem as if the often vague and ephemeralconstructs of psychological theory can now be displayed in neural tis-sue. There is a certain excitement in being able to palpate somethingthat was previously only imaginable, to see a functioning process on acomputer display of brain activation where previously that process wasonly inferred from patterns of reaction times and errors. This is a livelydevelopment in the science. However, there are reasons beyond this to

102 John Jonides

value data from neuroimaging laboratories, and these reasons are mereextensions of the ones that guided the discussions in the Gleitmans’Thursday evening research seminars (which the clock suggested wereendless, but which ended all too soon each week). Neuroimaging tech-niques can be applied, as our research suggests, to the identification ofcomponents of cognition and to the detailed description of the archi-tecture of these components. In this way, it is valuable to conceive ofneuroimaging data as an additional modality of insight into the phe-nomena of cognition, one that can supplement and enhance the behav-ioral study of normal and brain-injured humans.

Acknowledgment

This research was supported by a grant from the Office of NavalResearch and by a grant from the National Institute on Aging.

ReferencesAwh, E. and Jonides, J. (1999) Spatial selective attention and spatial working memory. In

The Attentive Brain, ed. R. Parasuraman. Cambridge, MA: MIT Press.Awh, E., Jonides, J., Smith, E.E., Schumacher, E.H., Koeppe, R.A., and Katz, S. (1996)

Dissociation of storage and rehearsal in verbal working memory: Evidence fromPET. Psychological Science 7:25–31.

Baddeley, A. D. (1986) Working Memory. Oxford: Oxford University Press.Baddeley, A. D. (1992) Working memory. Science 255:556–559.Braver, T. S., Cohen, J. D., Nystrom, L. E., Jonides, J., Smith, E. E., and Noll, D. C. (1997) A

parametric study of prefrontal cortex involvement in human working memory.NeuroImage 54:49–62.

Carpenter, P. A., Just, M. A., and Shell, P. (1990) What one intelligence test measures: Atheoretical account of the processing in the Raven Progressive Matrices Test.Psychological Review 97:404–431.

Cohen, J. D., Perlstein, W. M., Braver, T. S., Nystrom, L. E., Noll, D. C., Jonides, J., andSmith, E. E. (1997) Temporal dynamics of brain activation during a working mem-ory task. Nature 386:604–608.

Daneman, M. and Merikle, P. M. (1996) Working memory and language comprehension:A Meta-Analysis. Psychonomic Bulletin and Review 3:422–433.

Donders, R. C. (1868) Over de snelheid van psychische processen. Onderzoekingen gedaanin het psyiologish laboratorium der Utrechtsche Hoogeschool: Tweede Reeks 2: 92–120.

Jonides, J., Reuter-Lorenz, P., Smith, E. E., Awh, E., Barnes, L., Drain, M., Glass, J., Lauber,E., Patalano, A., and Schumacher, E. H. (1996) Verbal and spatial working mem-ory. In The Psychology of Learning and Motivation, ed. D. Medin.

Jonides, J., Schumacher, E. H., Smith, E. E., Lauber, E. J., Awh, E., Minoshima, S., andKoeppe, R. A. (1997) Verbal-working-memory load affects regional brain activa-tion as measured by PET. Journal of Cognitive Neuroscience 9:462–475.

Jonides, J., Smith, E. E., Koeppe, R. A., Awh, E. Minoshima, S., and Mintun, M. A. (1993)Spatial working memory in humans as revealed by PET. Nature 363:623–625.

Kyllonen, P. C. and Christal, R. E. (1990) Reasoning ability is (little more than) working-memory capacity?! Intelligence 14:389–433.

Mechanisms of Verbal Working Memory 103

Logie, R. H. (1995) Visuo-spatial Working Memory. Hillsdale, NJ: Lawrence ErlbaumAssociates.

Logie, R. H., Zucco, G. M., and Baddeley, A. D. (1990) Interference with visual short-termmemory. Acta Psychologica 75:55–84.

McCarthy, R. A. and Warrington, E. K. (1990) Cognitive Neuropsychology: A ClinicalIntroduction. San Diego: Academic Press.

Meudell, P. R. (1972) Comparative effects of two types of distraction on the recall of visu-ally presented verbal and nonverbal material. Journal of Experimental Psychology94:244–247.

Paulesu, E., Frith, C. D., and Frackowiak, R. S. J. (1993) The neural correlates of the verbalcomponent of working memory. Nature 362:342–344.

Salthouse, T. A. (1974) Using selective interference to investigate spatial memory repre-sentations. Memory and Cognition 2:749–857.

Salthouse, T. A. (1993) Influence of working memory on adult age differences in matrixreasoning. British Journal of Psychology 84:171–199.

Schumacher, E. H., Lauber, E., Awh, E., Jonides, J., Smith, E. E., and Koeppe, R. A. (1996)PET evidence for an amodal verbal working memory system. NeuroImage 3:79–88.

Smith, E. E. and Jonides, J. (1995) Working memory in humans: Neuropsychological evi-dence. In The Cognitive Neurosciences, ed. M. Gazzaniga. Cambridge, MA: MITPress, pp. 1009–1020.

Smith, E. E., Jonides, J., and Koeppe, R. A. (1996) Dissociating verbal and spatial workingmemory using PET. Cerebral Cortex 6:11–20.

Sternberg, S. (1966) High-speed scanning in human memory. Science 153:652–654.Sternberg, S. (1969) Memory-scanning: Mental processes revealed by reaction-time exper-

iments. American Scientist 57:421–457.

104 John Jonides

Chapter 8

A Nativist’s View of Learning: How to Combinethe Gleitmans in a Theory of Language Acquisition

Elissa L. Newport

An extremely prominent fact of my graduate training is that I workedwith both Henry and Lila Gleitman. The three of us collaborated, duringmy years at Penn and beyond, on a series of studies of mothers’ speechto young children (christened “Motherese” by Henry), and the Gleit-mans signed my dissertation as my joint advisors. Together, along witha number of other honored teachers at Penn, they forged the concepts Istill hold dear about honor, integrity, citizenship, and scholarship inacademic life and in my department. They have remained my dearfriends and mentors for more than twenty-five years. I look forward toand demand at least another twenty-five with them.

Most pertinent for writing this chapter, as I look back over the re-search I have done since graduate school, I discover that I have spentthe ensuing years thinking about my research like both Lila and Henry,and have conducted on a long-term basis two apparently quite differentlines of research, one which derives most clearly from the thinking Ibegan with Lila and the other which derives most clearly from thethinking I began with Henry. As a product of the two of them, I ofcourse think these lines of work are related; but I do confess that theyare styles of research typically performed by investigators on differentsides of the field and with different views of language acquisition. In the present chapter I will present them as the “Henry” part of my workand then the “Lila” part of my work. I hope by the end it will becomeclear that they are indeed relatable and related, and that the Gleitmanshave also succeeded in teaching me something about integrating theirviews.

The Problem of Language Acquisition, and Two Views on Its Solution

As we all know, the problem of language acquisition is as follows(Chomsky 1965): Natural languages are large (and hierarchically orga-nized) combinatorial systems. The learner’s task is to figure out thebasic elements, and then to learn which combinations of these elementsare permitted. The difficulty, of course, is how one learns, by observing

a subset of the permitted combinatorial, which parts of the unobservedcombinatorial space are permitted and which parts of the unobservedcombinatorial space are not.

The usual accounts of the solution, typically offered from benches onopposing sides of the field, are these:

Claim A (stated in current wording, but nonetheless an old view): Thereare rich statistical regularities in natural languages, and humans absorbthese input statistics remarkably readily.

Claim B (likewise an old view stated in current wording): There are toomany, and also too few, regularities present in the input to explainlearning. Learners bring to the task a strong set of advance biases, lead-ing them to “acquire” certain types of combinatorial patterns evenwhen they are not present in input, and to fail to acquire other types ofcombinatorial patterns even when they are present in input.

Because these accounts are usually maintained by investigators withopposing points of view, they are typically offered as though they werein conflict with one another. Alternatively, however, they might both betrue, and might therefore demand an integration with one another.Over the last twnety-five years I have been attempting to study each ofthese processes, in separate lines of work: one a long-term program ofexperimental research using an artificial language learning methodol-ogy in the lab, and the other a long-term program of research on criticalperiods and creolization in natural sign language acquisition. In thepresent chapter I will overview our most recent findings in each of theselines of research, and then attempt to indicate how these findings mightultimately be integrated with one another in a coherent picture of lan-guage acquisition.

Statistical Learning

Most models of language acquisition have approached the problem byconsidering how a learner might form a rule from a single sentence at atime. As we all know, given a single example string from the language,any open-minded learner could hypothesize a huge (indeed, infinite)number of potential rules (Chomsky 1965; Gold 1967). The usual solu-tion to this problem is to propose that there is relatively light learningfrom each example, and relatively heavy advance constraints on whatthe rules or principles of the language might be. One example is the no-tion of “triggering” of the setting of linguistic principles.

However, given a large and representative corpus of relevant sen-tences from the language (like that used by a good field linguist), theproblem changes somewhat (cf. Harris 1951; Maratsos and Chalkley

106 Elissa L. Newport

1980). Distributional information involves patterns of syllable, mor-pheme, and word sequencing—sometimes quite complex patterns exhibited over different parts of the corpus—which linguists have tra-ditionally used to identify the structures of a language. Recently thishas been called statistical information (cf. Charniak 1993; Saffran,Newport, and Aslin 1996)—for example, calculations of which se-quences of sounds appear together recurrently, or which linguistic con-texts form a recurring set of alternatives for the same items. Even with adistributional corpus, the learning problem does not change in prin-ciple: There are still infinitely many generalizations consistent with alinguistic corpus, just as there are with a single instance. But such infor-mation does provide a potentially richer input base for a learner suit-ably predisposed to use it. Given a corpus, a learner who is innatelyendowed with a constrained and structured set of analytic techniquesmight be able to use that corpus to reduce the alternatives greatly.

There are two important reasons this approach to language acquisi-tion has not been extensively pursued until the last few years (see, e.g.,numerous critiques of Maratsos and Chalkley 1980). First, there has notbeen an adequate theory of how to constrain such a learner, or even anadequate argument showing that such a theory could be constructedwithout looking essentially identical to the theory required for the single-sentence learner. (Unfortunately, the present chapter will not re-solve this problem.) Second, it has seemed on the face of it somewhatimplausible to imagine that a very young language learner could store alarge linguistic corpus in a relatively raw form, preserving it for use invarious types of distributional analysis during acquisition, or couldcompute the rather complex statistics from this corpus that would be re-quired to reduce the acquisition problem. However, in the last few yearsmy colleagues Richard Aslin, Jenny Saffran, Toby Mintz, and I, as wellas a number of other researchers in the fields of language acquisitionand computational linguistics, have begun to show that human learn-ers—even eight-month old infants—can perform surprisingly complexstatistical analyses of language data, even from brief exposures in thelab, and also that they are extremely selective in the types of analysesthey perform. In this section I will overview our results on a first distrib-utional problem, that of word segmentation. However, a great deal morework is required before we can determine how far an infant is capable ofprogressing through acquisition using an approach of this kind.

Statistical Learning and Word Segmentation

In a series of studies (Saffran, Newport, and Aslin 1996; Saffran, Aslin,and Newport 1996; Saffran et al. 1997; Aslin et al. 1998, 1999), Jenny

A Nativist’s View of Learning 107

Saffran, Richard Aslin, and I began to study this question by focusingon a problem in language acquisition that clearly involved learning,that innate knowledge could not solve, and for which a distributional orstatistical analysis could, at least in principle, provide an important con-tribution, if learners were capable of performing it. These studies haveexamined the problem of word segmentation: how does the learner de-termine, from the apparently continuous stream of speech, which se-quences of sounds form the words of the language? Part of the answerinvolves the use of prosodic and rhythmic information, as well as si-lence at the ends of utterances (Aslin, Woodward, LaMendola, andBever 1996; Brent and Cartwright 1996; Christiansen, Allen, and Seiden-berg 1998; Jusczyk, Cutler, and Redanz 1993; Mehler, Dupoux, andSegui 1990; Morgan and Saffran 1995). However, such cues are not al-ways available for use in initial segmentation (Aslin et al. 1996).

Several investigators (Chomsky 1955; Harris 1955; Hayes and Clark1970; Goodsitt, Morgan, and Kuhl 1993) have noted that this problemmight be solved by keeping track of relative consistency in the soundsequences. This observation can be converted into a statistical form:Learners might compute the conditional probabilities between sequen-tial syllables (called transitional probabilities; cf. Miller and Selfridge 1950;Goodsitt, Morgan, and Kuhl 1993; Christophe, Dupoux, Bertoncini, andMehler 1994; Saffran, Newport, and Aslin 1996). Over a speech corpus,those sequences with relatively high conditional probabilities are likelyto be inside words, and those with relatively low probabilities are likelyto be the accidental juxtapositions of sounds at word boundaries.Following an important study by Hayes and Clark (1970), we askedwhether human learners were capable of performing such computa-tions.

Word segmentation studiesOur initial study involved presenting adults with an artificial language(Saffran, Newport, and Aslin 1996; see also Hayes and Clark 1970). Thelanguage consisted of trisyllabic “words,” concatenated in randomorder and spoken by a speech synthesizer with no prosodic or acousticmarkers of word boundaries, to create an unbroken twenty-one-minutecorpus. Although transitional probabilities varied both within and be-tween words, the transitional probabilities inside words were relativelyhigh, while those spanning a word boundary were relatively low, as isthe case for real languages. After exposure to the corpus, subjects weregiven a series of two-alternative forced choice items, each containing aword from the language and either a nonword or a part-word (depend-ing on the experimental condition). Nonwords were three-syllable se-quences made of the same syllables used in the language, but in an

108 Elissa L. Newport

order that did not occur in the exposure corpus. Part-words were three-syllable sequences consisting of two syllables in the correct positionsand order, and a third syllable that did not occur in that word in the cor-pus. Subjects were to choose which of the alternatives in each itemsounded more familiar. Subjects in both the nonword and the part-wordconditions performed significantly and substantially above chance, sug-gesting that adults not only can acquire syllable order, but can also seg-ment a stream of syllables into groups based on the distributionalcharacteristics of the corpus.

In a second study we asked whether five- to six-year old childrencould perform the same task (Saffran, Newport, Aslin, Tunick, andBarrueco 1997). To prevent them from getting bored during familiariza-tion, we asked them to color on the computer, using a program calledKidPix, and we merely played the speech stream in the background,with no instructions to learn or even listen to the sounds. For compari-son, adults were given the same exposure. After one or two twenty-one-minute coloring sessions, both the adults and the children performedsignificantly above chance on a word-nonword forced-choice task.Thus children can also segment words from fluent speech based solelyon statistical information from a continuous corpus. Moreover, thisprocess can apparently proceed implicitly, without subjects’ attentiondirected at the speech stream or the analytic process.

We have also conducted a series of three studies on statistical learningin eight-month-old infants (Saffran, Aslin, and Newport 1996; Aslin,Saffran, and Newport 1998), chosen because this is the age at whichword segmentation in natural language acquisition is underway. In ourfirst study, infants were exposed to a simplified corpus of trisyllabicnonsense words (with transitional probabilities inside words of 1.0 and those across word boundaries of 0.33), presented continuously foronly two minutes. Then, using the preferential listening methodology(Jusczyk and Aslin 1995), we tested each infant with two words fromthe language and two nonwords (made up of familiarization syllablesin a novel order). Our results showed that infants listened differentiallyto words versus nonwords, indicating that they could discriminate be-tween them. Because the individual syllables in words and nonwordsoccurred with equal frequency in the familiarization corpus, the resultscannot be due to subjects’ discriminating the frequency of individualsyllables, but rather must be due to their discriminating syllable order:Infants must be noting that the syllables in nonwords never occurred inthat order in the familiarization corpus. In our second study, we askedwhether eight-month-olds could perform the more difficult task of dis-criminating words from part-words. In this study, the familiarizationcorpus was like that in the first study, but the test items consisted of

A Nativist’s View of Learning 109

words and part-words. Moreover, part-words in this study were moredifficult to discriminate from words than in our previous adult work.Here, part-words consisted of the final syllable of one word and the firsttwo syllables of another word. Thus these part-words had in fact oc-curred in the familiarization corpus. They differed from words in havingtransitional probabilities of 0.33 and 1.0 (as compared with 1.0 and 1.0for the words). Infants in this second study also listened differentially tothe part-words as compared to the words. Thus eight-month-olds do notmerely note whether a syllable sequence occurred or not, but apparentlycan perform an analysis of the statistics of the language corpus.

This second infant study does not, however, demonstrate preciselywhat statistic the infants are computing, and whether in particular theyare capable of computing conditional probabilities among sequentialsyllables. Because the words of the corpus were each presented withequal frequency, the part-words formed by their junctures were all lessfrequent than the words. Infants therefore could have been respondingto trisyllabic frequencies, rather than trisyllabic conditional probabili-ties. (Either of these would be quite impressive, but conditional proba-bilities would be more structurally informative in real languagelearning.) To pursue this issue further, we conducted a third study ofeight-month-olds (Aslin, Saffran, and Newport 1998), in which the testwords and part-words were matched in frequency of presentation dur-ing familiarization, and differed only in conditional probabilities. Weachieved this by creating a corpus in which two of the four words pre-sented in the familiarization corpus were more frequent than the othertwo. This resulted in a corpus in which the two part-words (formed bythe juncture between the high-frequency words) occurred with thesame frequency as the two low-frequency words. Nonetheless, the tran-sitional probabilities within words were still higher (1.0 and 1.0) thanthe transitional probabilities within the part-words (0.50 and 1.0). Ourresults showed that infants continued to discriminate between thewords and the part-words, demonstrating clearly that they can com-pute transitional probabilities and can use them to segment multisyl-labic words from fluent speech.

Segmentation in other modalities and domainsThe problem of segmenting elementary units out of a complex and apparently continuous array is a problem that occurs not only in lan-guage, but in other domains as well. Moreover, statistical solutions—examining which notes recurrently occur together to form a melody, orwhich parts of a visual scene move together against a background—might also be applicable in these domains. My colleagues and I havetherefore devised nonlinguistic materials, structured with the same sta-

110 Elissa L. Newport

tistics as our speech streams, but composed of quite different basic ele-ments, and presented them to adults and infants, to see whether thesame types of computations can be performed as readily in these do-mains, or whether these extraordinary computational abilities are in-stead restricted to or are specialized for speech. Our results show quitecomparable outcomes, in both adults and infants, using nonlinguistictone sequences (Saffran, Johnson, Aslin, and Newport 1999), and visualand visuomotor sequences (Asaad 1998; Hunt and Aslin 1998).

In sum, then, human learners show quite surprising abilities to com-pute and keep track of detailed aspects of sequential materials. Theseabilities appear not only in adults and children, but even in young in-fants of language-learning age; and they appear for rapidly presentedlinguistic materials, as well as for musical and visual sequences.

Expanding Statistical Learning to Other Aspects of Language

We are just beginning to ask how this statistical approach might be ex-tended to other aspects of language. The particular computations wehave examined thus far—transitional probabilities between adjacentsyllables—would be extremely helpful in word segmentation, butwould not be adequate for learning other aspects of natural languages(Chomsky 1955, 1957; Newport and Aslin, in preparation; Saffran,Aslin, and Newport 1997). Our subsequent question is therefore notwhether this precise approach can be extended, but rather whetherhuman learners can also compute other complex aspects of linguistic se-quences that might be relevant to the acquisition of syntax and mor-phology—for example, statistics concerning the formation of wordclasses (Mintz, Newport, and Bever 1995, and under review), long-distance dependencies (Newport, Calandra, and Aslin, in progress),and hierarchical structure. Much further research is required before wecan say whether the approach we have taken in our word segmentationwork is limited to such low-level, early parts of the acquisition process,or rather whether it can be extended as well as to other higher-levelproblems. Moreover, as one expands this approach, it is critical that it beintegrated with appropriate constraints, so that one can explain whylearners do not always learn the regularities of their input, as well aswhy they sometimes do. This leads directly to considering the secondline of research I have been conducting.

Non-Statistical Learning?

Although statistical, or distributional, information may be extremelyhelpful to at least certain parts of language acquisition, it is not the case

A Nativist’s View of Learning 111

that learners acquire statistical information in a simple or slavish way.First, as noted earlier, there is an infinite number of statistical propertiesthat might be computed or learned from input, and no learner could ac-quire all of these. Second, even for those properties highly pertinent tothe structure of a language, children do not merely reproduce the statis-tics of their input. In some cases, children use these statistics to build arulelike, nonstatistically organized output. In other cases they produceregularities not present in their input at all. Examples of these discrep-ancies between input and output come from a second line of research,on “creolization” in the acquisition of reduced and inconsistent signlanguage input.

Studying Language Acquisition Using the Natural Experiment

The previous line of work I have described descends in fact from bothLila and Henry. Lila, a student of Zellig Harris, began working in lin-guistics as a collaborator on the first computational linguistics project,and has transmitted to me a love of distributional analyses and of mech-anisms of acquisition. But designing miniature language studies andthinking about learning theories is a love I acquired primarily fromHenry. On the other hand, my second long-term research program initi-ates most clearly with Lila. As many have noted in this volume, a numberof us began, in our graduate work with Lila, to pursue our nature-nurture questions by seeking unusual “natural experiments” of acquisi-tion—natural deviations of input or internal state that might shed speciallight on how these variables affect the course of acquisition.

This is a method I have continued in my own work, for example, inseeking subjects who have been exposed to their primary language atvarying ages (Newport 1990), or who have learned their language in in-fancy but only from an unusual source (Newport in press). In perhapssurprising contrast to what I have argued above, these studies have al-ways suggested that internal state, and not linguistic input, is the dom-inant controller of the course of acquisition. Examined in more detail,these studies provide an important modification of, and integrationwith, the studies of statistical learning described above.

Natural Experiments of Linguistic Input

Children virtually always acquire their primary language from speak-ers who are fully fluent in the language. This means that their input ishighly regular and systematic: Even though it may be remarkably diffi-cult for a learning theory to explain how the regularities could beuniquely reconstructed from this input, there are rules and patterns un-

112 Elissa L. Newport

derlying linguistic strings to which the learner is exposed. A statisticalapproach to this type of learning, then, is an approach that hypothesizesthat the rules may be helpfully revealed by computing some input sta-tistics.

In contrast, in ongoing work we are observing children who are ac-quiring their primary language entirely from speakers who are them-selves neither fluent nor native users of the language. In some cases theinput they provide to their children is thus truly statistical: Morpho-logical rules of the language are used only probabilistically, and manyinconsistent errors are made. In other cases the input omits certain nat-ural language properties. The outcomes of acquisition in these circum-stances show quite clearly that, although children may use inputstatistics to learn parts of their language, they do not reproduce theinput statistics in their own output. Rather, the architecture of their out-put grammar is sharpened and systematized.

Our subjects are congenitally and profoundly deaf children who areacquiring American Sign Language (ASL) as their primary language.All of them are exposed to some form of ASL from birth or shortly after.However, because their families vary in their proficiency in ASL (andbecause we study children who have little or no input to sign languageoutside of their families and teachers, whom we study), the children’sinput to ASL may be very reduced and inconsistent. Our findings todate concern two particular types of input variation. In one line of workwe have observed the acquisition of the morphemes of ASL verbs ofmotion. All of the parents use these morphemes to some degree, butvary in the consistency with which they use morphemes in their requiredcontexts. When they err, they either omit the required morphemes or replace them with ungrammatical forms. Studies of the children’s ac-quisition of this morphology allow us to see the effects of input incon-sistency on the acquisition of these language-specific structures. In asecond line of work we are observing the acquisition of syntactic andmorphological rules that are not just specific details of ASL but formpart of the universal patterns and principles of all languages. Here, as ithappens, the parents vary not only in the consistency with which theyuse the structures, but also in whether they themselves exemplify or vio-late these linguistic universals. This work allows us to ask whether chil-dren must be exposed to these structural principles at all in order toobserve them in their own productions.

Inconsistent input to language-specific morphologyOur first work on this topic has been a case study of a deaf child, whomwe call Simon, acquiring ASL as his native language from his parents(Newport 1999; Ross and Newport 1996; Singleton 1989; Singleton and

A Nativist’s View of Learning 113

Newport 1994 and under review). Simon is the only congenitally deafson of two deaf parents; both parents were first exposed to ASL in theirlate teens and now use it as their primary language, with each other andSimon. Simon attends a school where none of the teachers or other stu-dents knows ASL; the school uses a form of Signed English, which doesnot contain the morphology or syntax of ASL that we have studied inSimon, and all other students in the school have hearing parents whodo not know ASL. Simon’s parents’ friends are also nonnative learnersof ASL. In short, Simon’s only input to ASL is from his parents. We havefilmed this family’s signing since Simon was two years old, but our firstanalyses focused on Simon’s performance, compared with that of hisparents, at a time when he should have completed his acquisition ofASL, at age 7:11.

Simon, his mother, and his father were each tested for their produc-tion of the morphemes of ASL verbs of motion. Simon’s performancewas also compared with that of deaf children of his age who have nativesigning parents, and his parents’ performances were compared withadult native signers and late learners of ASL. In native ASL, verbs ofmotion involve producing a large number of morphemes in combina-tion, and these verbs are therefore difficult for both late learners andyoung children to acquire. Each of the morphemes does, however, havea set of obligatory contexts, and is produced by native signers in ahighly regular and systematic way.

Simon’s parents sign like other late learners: They use virtually all ofthe obligatory ASL morphemes, but only at middling levels of consis-tency. On relatively simple morphemes (the movement morphemes ofASL), they average 65–85% correct usage. In contrast, Simon uses thesemorphemes much more consistently (about 90% correct), fully equal tochildren whose parents are native ASL signers. Thus, when input isquite inconsistent, Simon is nonetheless able to regularize the languageand surpass his input models. On more difficult morphemes (the hand-shape classifiers of ASL), where his parents were extremely inconsistent(about 45% correct), Simon did not perform at native levels by age 7; buteven here he did surpass his parents.

Ross and Newport (1996) examined Simon’s development over timeon the same morphology studied at age 7:11 by Singleton and Newport.This analysis examined Simon’s acquisition of the morphology of verbsof motion from age 2:6 through 9:1 (both earlier and later than Singletonand Newport’s analyses), again compared with children of the sameages receiving fully native input. For movement morphemes (whereSimon’s input was moderately consistent), Simon matched children receiving native input throughout development; his use of these mor-phemes exhibited no developmental delay, and no reduction in consis-tency or complexity. For handshape classifiers (where input consistency

114 Elissa L. Newport

was lower), Simon began his acquisition process normally, but reachedan asymptote, well below normal usage, at age 4:6 that continued un-changed until 9:1. Ross (in preparation) has suggested that Simon inthis portion of ASL may be forming his own system, somewhat differ-ent than ASL but nonetheless systematic.

To examine a greater range of inconsistency in linguistic input, Rossand Newport (in progress) have begun to study deaf children acquiringtheir sign language from hearing parents. These parents have learned tosign only slightly before their child, and their fluency in the language isoften extremely limited. In one study we have compared these childrenand their parents to native signing families, on the same morphology aswas studied in Simon. The full range of subjects we have studied thusranges from native input (for control subjects) through moderately con-sistent input (for subjects with deaf late-learning parents, like Simon,and also some with hearing late-learning parents) to extremely incon-sistent (for example, one child, Sarah, receives input from her hearingmother which is only 15% consistent on movement morphemes and 8%consistent on handshape classifiers). A summary picture of the data isshown in figure 8.1. As can be seen there, all of these children perform at native or near-native levels on movement morphemes. On the moredifficult handshape classifiers (where parents are often extremely in-consistent), the children are not fully native, but their consistency sub-stantially exceeds that of their input.

What is the mechanism by which children overcome such high de-grees of inconsistency in their input? Two hypotheses seem possible.One hypothesis is that children know, innately, that natural languagemorphology is deterministic, not probabilistic, and acquire this mor-phology in accord with this knowledge. An alternative hypothesis is re-lated to the distribution of mappings between form and meaning in theinput data. Whereas, say, only 65% of the verbs referring to “falling

A Nativist’s View of Learning 115

Figure 8.1.

events” use the FALL morpheme, the other 35% use a scattering of otherforms, with no one of these used with substantial frequency. This distri-bution may be one in which learners will acquire only the major form,and will fail to acquire the very low frequency forms. Moreover, this se-lectivity of learning may be particularly true for children, with limitedability to learn complex data. We are in the process of testing the secondhypothesis with experimental studies in the lab (Hudson and Newport,in progress).

Building structure with no input: Universal architectural principles of grammarWe have also studied other linguistic structures, whose input exhibitsdifferent types of reduction. In some studies we examine not individ-ual morphemes, but rather the way constructions combine over thelanguage. In contrast to language-specific individual forms, such com-binations across the language concern the architecture of the grammar,and are the domain of linguistic universals. What input do late-learningparents provide to such patterns, and what do their children do whenlearning from this input?

Singleton and Newport (in preparation; Newport 1999) analyzed twosuch arenas in Simon at 9:1 and his late-learning deaf parents. One con-struction concerned inflections for number and aspect, and their combi-nation. Simon’s parents used each of the inflections with middlingconsistency, but they never combined these inflections; rather, they de-scribed complex events by using only one inflection at a time, express-ing the remaining part of the meaning in a periphrastic expression (aphrase accompanying the verb). They thus provided Simon with noinput concerning how these inflections might be combined to expresscomplex meanings, and in fact provided input that suggested that twoinflections cannot be combined. This noncombinatorial pattern, com-mon in late learners, is quite uncharacteristic of natural languages; nat-ural languages have rules that apply independently (either in sequenceor simultaneously), and thus rules can freely combine. (For example, ifthe event concerns plurality and possession, both plural and possessivemorphemes will be used; one does not block or exclude the other.)Despite his input but in line with other linguistic systems, Simon com-bined inflections perfectly (on our elicitation tasks, Simon scored 100%correct, compared with 25% and 0% correct for his mother and father).This outcome suggests that he did not learn these architectural princi-ples from his input, but imposed them on his language.

Singleton and Newport also examined Simon’s comprehension oftopicalized structures in ASL. Topicalization is the movement of aphrase to the beginning of the sentence, marking it as the topic of thatsentence; in ASL, topics are marked with a special facial expression as

116 Elissa L. Newport

well as a special word order. Syntactic movement, such as topicaliza-tion, universally follows a principle called structure dependence: Onlyproper units of structure, such as phrases, can be moved; other stringsof adjacent words which do not form a phrase cannot undergo topical-ization (Chomsky 1975). However, Simon’s parents (and other latelearners of ASL), tested on their comprehension of topicalized struc-tures, did not exemplify or observe structure dependence. In their ownsigning, Simon’s parents did not produce the full range of topicalizedstructures; they only topicalized subjects (for example, JOHN HITBALL, where JOHN is topic-marked). In comprehension, they scoredbelow chance on sentences in which other phrases were topicalized. Inshort, their input to Simon exhibited a highly reduced range of exam-ples for this principle, and their comprehension consistently violatedthe principle. Nonetheless, Simon comprehended topic-marked sen-tences 100% correctly, indicating that his comprehension was fully inaccord with structure dependence. Simon presumably acquired thebasic phrase structure of ASL from his input; but his organization ofmovement rules in accord with structure dependence appears not tohave come from his input. In both these cases—combining inflectionalrules, and moving words to new positions—Simon appears to havegone well beyond his input, imposing universal principles of rule archi-tecture that his parents’ usage did not illustrate. Presumably the con-straints underlying these principles are part of the child’s internalbiases about how language must be structured.

Conclusions

As I noted at the beginning of this chapter, the apparent contrasts inthese lines of work—one showing remarkable statistical learning, theother showing remarkable reshaping and restructuring of input—mightsound like they come from opposing views of acquisition. However, ifcombined, they offer an appropriately rich picture of acquisition, inwhich there is both learning and constraints within which this learningoccurs.

One relatively simple integration within these findings concerns therelation between statistical learning and Simon’s ability to surpass hisinconsistent linguistic input. When faced with probabilistically usedmorphology, Simon acquires the most consistent portions of his input,and fails to learn the inconsistencies. By selectively learning parts of hisinput, then, he turns statistics into rules (Singleton and Newport, underreview).

In other parts of our research a more complex integration and furtherinvestigation are needed. Where do learners sharpen the probabilitiesprovided by the input, where do they ignore their input, and where do

A Nativist’s View of Learning 117

they add structure that is not present? Are the occasions for theseprocesses created by particular distributional contexts in the input itself, or (as is more traditionally claimed) by innate knowledge ofgrammatical architectures? I hope it is not surprising that I don’t haveanswers to these questions. But one more lengthy meeting of theGleitman research seminar, with Henry and Lila putting their remark-able minds together over “my” research, would really help.

Acknowledgments

All of the research described in this chapter was done with importantcollaborators, whom I gratefully acknowledge: Richard Aslin and JennySaffran, my collaborators on statistical learning in word segmentation;and Jenny Singleton, Danielle Ross, and Ted Supalla, my collaboratorson studies of language acquisition from imperfect input. This researchwas supported in part by NIH grant DC00167 to E. Newport and T.Supalla, and in part by NSF grant KDI–9873477 to R. Aslin, E. Newport,R. Jacobs, and M. Hauser.

References

Asaad, P. (1998) Statistical learning of sequential visual patterns. Honors thesis,Department of Brain and Cognitive Sciences, University of Rochester.

Aslin, R. N, Saffran, J. R., and Newport, E. L. (1998) Computation of conditional probabil-ity statistics by 8-month-old infants. Psychological Science 9:321–324.

Aslin, R. N, Saffran, J. R., and Newport, E. L. (1999) Statistical learning in linguistic andnon-linguistic domains. In B. MacWhinney, ed., Emergentist Approaches to Lang-uage. Mahwah, NJ: Lawrence Erlbaum Associates.

Aslin, R. N., Woodward, J. Z., LaMendola, N. P., and Bever, T. G. (1996) Models of wordsegmentation in fluent maternal speech to infants. In J. L. Morgan and K. Demuth,eds., Signal to Syntax: Bootstrapping from Speech to Grammar in Early Acquisition.Mahwah, NJ: Lawrence Erlbaum Associates.

Brent, M. R. and Cartwright, T. A. (1996) Distributional regularity and phonotactic con-straints are useful for segmentation. Cognition 61:93–120.

Charniak, E. (1993). Statistical Language Learning. Cambridge, MA: MIT Press.Chomsky, N. (1955/1975) The Logical Structure of Linguistic Theory. New York: Plenum

Press.Chomsky, N. (1957) Syntactic Structures. The Hague: Mouton.Chomsky, N. (1965) Aspects of the Theory of Syntax. Cambridge, MA: MIT Press.Chomsky, N. (1975) Reflections on Language. New York.Christiansen, M. H., Allen, J., and Seidenberg, M. S. (1998) Learning to segment speech

using multiple cues: A connectionist model. Language and Cognitive Processes.Christophe, A., Dupoux, E., Bertoncini, J., and Mehler, J. (1994) Do infants perceive word

boundaries? An empirical study of the bootstrapping of lexical acquisition. Journalof the Acoustical Society of America 95:1570–1580.

Gold, E. M. (1967) Language identification in the limit. Information and Control 16:447–474.Goodsitt, J. V., Morgan, J. L., and Kuhl, P. K. (1993) Perceptual strategies in prelingual

speech segmentation. Journal of Child Language 20:229–252.Harris, Z. S. (1951) Methods in Structural Linguistics. Chicago: University of Chicago Press.

118 Elissa L. Newport

Harris, Z. S. (1955) From phoneme to morpheme. Language 31:190–222.Hayes, J. R. and Clark, H. H. (1970) Experiments in the segmentation of an artificial

speech analog. In J. R. Hayes, ed., Cognition and the Development of Language. NewYork: Wiley.

Hunt, R. H. and Aslin, R. N. (1998) Statistical learning of visuomotor sequences: Implicitacquisition of sub-patterns. Proceedings of the Twentieth Annual Conference of theCognitive Science Society. Mahwah, NJ: Lawrence Erlbaum Associates.

Jusczyk, P. W. and Aslin, R. N. (1995) Infants’ detection of the sound patterns of words influent speech. Cognitive Psychology 29:1–23.

Jusczyk, P. W., Cutler, A., and Redanz, N. J. (1993) Infants’ preference for the predomi-nant stress patterns of English words. Child Development 64:675–687.

Maratsos, M. and Chalkley, M. A. (1980) The internal language of children’s syntax: Theontogenesis and representation of syntactic categories. In K. Nelson, ed., Children’sLanguage, vol. 2. New York: Gardner Press.

Mehler, J., Dupoux, E., and Segui, J. (1990) Constraining models of lexical access: Theonset of word recognition. In G. Altmann, ed., Cognitive Models of Speech Processing:Psycholinguistic and Computational Perspectives. Cambridge, MA: MIT Press.

Miller, G. A. and Selfridge, J. A. (1950) Verbal context and the recall of meaningful mater-ial. American Journal of Psychology 63:176–185.

Mintz, T. H., Newport, E. L., and Bever, T. G. (1995) Distributional regularities of formclass in speech to young children. In J. N. Beckman, ed., Proceedings of NELS 25 (vol.2, pp. 43–54). Amherst, MA: Graduate Linguistic Student Association.

Morgan, J. L. and Saffran, J. R. (1995) Emerging integration of sequential and supraseg-mental information in preverbal speech segmentation. Child Development66:911–936.

Newport, E. L. (1990) Maturational constraints on language learning. Cognitive Science14:11–28.

Newport, E. L. (1999) Reduced input in the acquisition of signed languages: Con-tributions to the study of creolization. In M. DeGraff, ed., Language Creation andLanguage Change: Creolization, Diachrony, and Development. Cambridge, MA: MITPress.

Ross, D. S. and Newport, E. L. (1996) The development of language from non-native lin-guistic input. In A. Stringfellow, D. Cahana-Amitay, E. Hughes, and A. Zukowski,eds., Proceedings of the 20th Annual Boston University Conference on LanguageDevelopment, vol. 2. Somerville, MA: Cascadilla Press.

Saffran, J. R., Aslin, R. N., and Newport, E. L. (1996) Statistical learning by 8-month old in-fants. Science 274:1926–1928.

Saffran, J. R., Aslin, R. N., and Newport, E. L. (1997) Reply to five letters to the editor onthe topic of “acquiring language.” Science 276:1177–1181, 1276.

Saffran, J. R., Johnson, E. K., Aslin, R. N., and Newport, E. L. (1999) Statistical learning oftonal sequences by human infants and adults. Cognition 70:27–52.

Saffran, J. R., Newport, E. L., and Aslin, R. N. (1996) Word segmentation: The role of dis-tributional cues. Journal of Memory and Language 35:606–621.

Saffran, J. R., Newport, E. L., Aslin, R. N., Tunick, R. A., and Barrueco, S. (1997) Incidentallanguage learning: Listening (and learning) out of the corner of your ear. Psycho-logical Science 8:101–105.

Singleton, J. L. (1989) Restructuring of language from impoverished input: Evidence forlinguistic compensation. Doctoral dissertation, University of Illinois.

Singleton, J. L. and Newport, E. L. (1994) When learners surpass their models. Unpub-lished manuscript, University of Illinois.

Singleton, J. L. and Newport, E. L. (under review) When learners surpass their models:The acquisition of American Sign Language from inconsistent input.

A Nativist’s View of Learning 119

Chapter 9

Learning with and without a Helping Hand

Susan Goldin-Meadow

Lila was at Swarthmore College when I first met her. I was a prospectivegraduate student at Penn and anxious to meet the “language person”affiliated with the department. I took the train out to Swarthmore wherewe had our meeting in her office. I proudly described to this famousperson, whose New York accent instantly made me feel at home, thework that I had begun as an undergraduate during my junior yearabroad in Geneva, a project conducted with a fellow student (AnnetteKarmiloff-Smith) under the guidance of the “language person” at thePiagetian Institute (Mimi Sinclair). I carefully explained the findings onchildren’s comprehension of embedded and nonembedded relativeclauses and was doing fine until Lila asked me why I did the research.What was interesting about it? We were both, at that point, embar-rassed—Lila for having asked me a question to which I clearly did notknow the answer, and me for being so completely naive as to havenever even thought about the question. I have since learned manythings from Lila—the importance of an elegantly turned phrase, of agood pun in the title of a manuscript, of having food at a seminar—butnothing more essential than that lesson, reinforced by example manytimes over the years. It is vital not only to know what you’re doing butto know why you’re doing it. Equally important, the “why” must bestatable in terms that are accessible to your grandmother, or at the leastto all members of a psychology or linguistics department.

My most salient memory of Henry, other than the late-night conver-sations after the seminars around the Gleitman kitchen table (I havelearned that kitchen tables are where all important conversations takeplace, and I have a great fondness for that particular table), was his rolein preparing us to present our work to the world, and in particular, tothe departments interested in interviewing us for a job. The job-talk rit-ual at Penn was an incredible learning experience and it revolvedaround Henry. You would give a talk, including slides (no overheadsthen) and videotapes (reel-to-reel, no cassettes) to a subset of the de-partment, always including Henry and always in A29 (the basement

room with no windows). The faculty then ripped the talk to shreds. Atthe end, you were told that this was great work, it just needed to be putin the right light. The next step was to take the shredded talk to Henry.What Henry did to the talk was part theater, but only part. His real con-tribution was to listen—really listen—to what you thought was impor-tant about your work and to help you say just that. And the outcomewas not only a polished job talk but a much deeper understanding ofyour own work. As I watched student after student go through thisprocess (the difference between the first version of the job talk and thefinal product presented to the entire department was always large andstriking), I developed a deep appreciation for the intelligence and skillthat led to each transformation. When it came time for my own practicejob talk, I was not disappointed. The reformulation that Henry workedout with me the night we repaired my shredded talk has stayed with meand influenced my research for twenty years—both as a model for howto do good research and as a model for how to be a good teacher.

While Henry helped me to understand, and be articulate about, myown ideas about language learning, it was Lila whose remarkable cre-ativity helped generate those ideas in the first place. Lila nurtured thefascination with language that I brought to Penn, giving me an appreci-ation for the complexities and elegance of linguistic structure and forhow hard the language-learning problem is, for both the child and theexperimenter. Language learning was a frequent topic of conversationat the weekly evening seminars at the Gleitman household. Indeed, itwas there that the strategy of exploring language learning by varyingits parameters was born. The experimentalist in Henry demanded ma-nipulation—to understand a phenomenon fully, one must be able tovary the factors thought to be important to it and observe their effect onit. But, for obvious ethical reasons, the language-learning situation isnot easily amenable to experimental manipulation.

Resourceful as always, Henry and Lila arrived at the idea of makinguse of experiments of nature to explore the language-learning problem,a strategy that has resulted in a large number of profitable studies con-verging on a coherent picture of language learning (Gleitman and New-port 1995). First came the “Motherese” work—did the natural variationsin how mothers spoke to their children result in differences in howthose children learned language (Newport, Gleitman, and Gleitman1977)? Next came attempts to extend the range of natural variation.Would children lacking access to linguistic input develop languagenonetheless (Feldman, Goldin-Meadow and Gleitman 1978)? Wouldchildren lacking access to vision be able to align their sightless worldswith the linguistic inputs they receive so that they too could learn lan-guage (Landau and Gleitman 1985)? Would children whose mental

122 Susan Goldin-Meadow

development is delayed be able to acquire language following the sametrajectory as more cognitively able children (Fowler, Gelman, andGleitman 1994)? An intellectually exciting research program grew outof those meetings in the Gleitman living room, one that was devoted toexploring how children learn language, not just documenting the stageschildren pass through as they learn language.

This question—how do children arrive at language given their inher-ent endowments and the inputs they receive—has guided my researchfor two decades. My research program grew out of the work I beganwith Lila and Heidi Feldman, exploring learning over ontogenetic timewith a focus on tapping the aspects of language learning that are devel-opmentally stable or “resilient.” I have studied the gesture systems cre-ated by deaf children whose hearing losses prevent them from learningthe spoken language that surounds them, and whose hearing parentshave not yet exposed them to a conventional sign language. The lin-guistic properties of these gesture systems, by virtue of the fact that theyare developed under language-learning circumstances that vary dra-matically from the typical, must be robust in children. To the extent thatthe outcome of the language development process is the same when somany of the input parameters are changed, we learn about how imper-vious the process is to variations in the environment (“learning withouta helping hand”).

In addition, I have over the last ten years begun another research pro-gram (linked superficially with the first by virtue of its focus on gesture)that explores learning over much shorter periods of time. Here, in nor-mally hearing children, gesture is not called upon to fulfill the functionsof a linguistic sytem (speech does that admirably well), and indeed ges-ture does not. Instead, gesture takes on a different role, reflecting thethoughts, sometimes inexpressible in speech, that learners have as theygo from a less adequate to a more adequate understanding of a task. Itis, in fact, the mismatch between the information conveyed in gestureand in speech that signals a learner’s readiness to make this transition.Occurring as it does in naturalistic conversation, gesture can serve as anobservable index of one’s readiness to learn and can therefore providean additional medium of instruction and communication for bothlearners and teachers (“learning with a helping hand”).

Learning without a Helping Hand: Gesture as a Testament to the Resilience ofLanguage

There may be no greater proof of the resilience of language in humansthan the fact that, when deprived of a language model entirely, thehuman child will invent one nonetheless. Deaf children whose severe

Learning with and without a Helping Hand 123

hearing losses prevent them from learning spoken language and whosehearing parents have not exposed them to a sign language might be ex-pected to fail to communicate—or to communicate in nonlanguage-likeways. But, in fact, deaf children in these circumstances do communicatewith those around them and they use gesture to do so. Although itwould certainly be possible to convey information in a mimelike fash-ion (e.g., elaborately enacting a scene in which the child gets and eats acookie to request one), the children don’t behave like mimes. Rather,they produce gestures according to a segmented and combinatorial for-mat akin to the format that characterizes all natural languages, be theysigned or spoken (e.g., the child points to the cookie and then jabs herhand several times at her mouth, using two gestures in sequence to con-vey “cookie EAT”1). In the next sections, I briefly describe the propertiesof the deaf child’s gesture system, properties that constitute resilientproperties of language learning.

Resilient Properties of Language-Learning

Sentence-level structureThe gesture strings that the deaf children generate can be described interms of very simple “rules.” The rules predict, on a probabilistic basis,which semantic elements are likely to be gestured and where in the ges-ture string those elements are likely to be produced (Feldman et al.1978; Goldin-Meadow and Feldman 1977; Goldin-Meadow and My-lander 1984, 1990). Thus, for example, the children were likely to pro-duce a gesture for the patient, as opposed to the actor, in a sentenceabout eating (e.g., a gesture for the cheese rather than the mouse) andwere likely to place that gesture in the first position of their two-gesturesentences (e.g., “cheese EAT” rather than the reverse).

Furthermore, each of the children’s gesture sentences, although fre-quently incomplete at the surface level, was associated with a completepredicate frame at the underlying level (Goldin-Meadow 1985). For ex-ample, there was evidence in the children’s gestures themselves for apredicate frame consisting of three elements—the actor, the action, andthe patient—underlying sentences about eating.

Finally, the children’s gesture sentences were characterized by recur-sion, the concatenation of two or more one-proposition predicate framesinto a single, complex sentence (Goldin-Meadow 1982); for example,“GIVE palm EAT,” a sentence requesting that the experimenter put inthe child’s palm a toy grape (proposition 1) that could be eaten (propo-sition 2). In addition, the surface form of the deaf children’s complexsentences was characterized by the systematic reduction of redundant

124 Susan Goldin-Meadow

elements (Goldin-Meadow 1987), devices that are fittingly reminiscent ofthose described by Lila in her early work on conjunction (Gleitman 1965).

Word-level structureIn addition to structure at the sentence level, each deaf child’s gesturesalso had structure at the word level (Goldin-Meadow, Mylander, andButcher 1995). Each gesture was composed of a handshape component(e.g., an O-handshape representing the roundness of a penny) and amotion component (e.g., a short arc motion representing a putting-down action). The meaning of the gesture as a whole was then de-termined by the meanings of each of these parts in combination (“putting-down-roundness”).

Although similar in many respects, the morphological systems offour deaf children studied thus far were sufficiently different to suggestthat the children had introduced relatively arbitrary—albeit stilliconic—distinctions into their systems (Goldin-Meadow et al. 1995). Forexample, two children used a C-shaped hand to represent objects two tothree inches in width (e.g., a cup or a box), while two other childrenused the same handshape to represent objects that were slightly smaller,one to two inches in width (e.g., a banana or a toy soldier). The fact thatthere were differences in the ways the children defined a particularmorpheme suggests that there were choices to be made (although all ofthe choices still were transparent with respect to their referents). More-over, the choices that a given child made could not be determined with-out knowing that child’s individual system. In other words, one cannotpredict the precise boundaries of a child’s morphemes without know-ing that child’s individual system. It is in this sense that the deaf chil-dren’s gesture systems can be said to be arbitrary.

Grammatical categoriesIn addition to combining components to create the stem of a gesture,one deaf child also altered the internal parts of his gestures to mark thegrammatical function of those gestures (Goldin-Meadow, Butcher,Mylander, and Dodge 1994). In particular, the child tended to abbrevi-ate a form when it played a noun role but not when it played a verb role.In contrast, the child would alter the placement of a form when itplayed a verb role but not when it played a noun role. For example,when used to mean “jar” (noun), the TWIST form would be producedwith a single turn and in neutral space, but when used to mean “twist”(verb), it would be produced with several turns and extended towardthe intended patient.

In addition to marking grammatical categories morphologically, thesame child also marked the categories syntactically (Goldin-Meadow et

Learning with and without a Helping Hand 125

126 Susan Goldin-Meadow

al. 1994). The child placed a form in the initial position of a two-gesturesentence when it played a noun role (“TWIST jar,” used to identify thejar) but in second position when it played a verb role (“jar TWIST,” usedto request that the jar be opened).

Interestingly, as in many natural languages (cf. Thompson 1988), adjec-tives in this deaf child’s gesture system were marked like nouns morpho-logically but like verbs syntactically (Goldin-Meadow et al. 1994). Forexample, when used as an adjective to mean “broken,” the BREAK ges-ture was produced in neutral space and abbreviated, like a noun (i.e., twofists held side-by-side in the chest area, separated from each other onlyonce) but it was placed in second position, like a verb (i.e., “toy BREAK”).

Language useThe deaf children did not invent this structural complexity to serve asingle function. Rather, they used their gestures for a wide variety offunctions typically served by language—to convey information aboutcurrent, past, and future events, and to manipulate the world aroundthem (Butcher, Mylander, and Goldin-Meadow 1991). For example, todescribe a visit to Santa Claus, one of the deaf children first pointed athimself, indicated Santa via a LAUGH gesture and a MOUSTACHEgesture, pointed at his own knee to indicate that he sat on Santa’s lap,produced a FIRETRUCK gesture to indicate that he requested this toyfrom Santa, produced an EAT gesture to indicate that he ate a pretzel,and then finished off the sequence with a palm hand arcing away fromhis body (his nonpresent marker) and a final point at himself (Morfordand Goldin-Meadow 1997).

In addition to the major function of communicating with others, onedeaf child used gesture when no one was paying attention, as though“talking” to himself (Goldin-Meadow 1993). Once when the child wastrying to copy a configuration of blocks off of a model, he made anARCED gesture in the air thus indicating the block he needed next;when the experimenter offered a block that fit this description, the childignored her, making it clear that his gesture was not directed at her butwas for his use only. The same child also used gesture to refer to his owngestures (Goldin-Meadow 1993), and to comment on (indeed criticize)the gestures of his hearing sister (Singleton, Morford, and Goldin-Meadow 1993).

The Environmental Conditions That Support the Development of a GestureSystem

These then are properties of language that arise when a child develops acommunication system without benefit of conventional linguistic input.

What does this list have to do with language learning that takes placeunder normal circumstances? Perhaps nothing, although it is an ac-cepted fact that children, even when given access to a language model,routinely go beyond that input—at the least, children hear sentences,but learn rules. What is striking about the deaf children is not just thatthey are creating a language with little environmental support, but thattheir product has properties in common with the languages learned bychildren exposed to conventional languages even though they havevery different materials to work with. To better understand the relation-ship between the product and the materials the deaf children have attheir disposal, we explored the environmental conditions under whichthe children developed their gesture systems.

Input from the gestures of hearing individualsWe first observed the spontaneous gestures that the children’s hearingparents produced when they communicated with their children. Wefound that the structure evident at the sentence and word levels in eachof the deaf children’s gesture systems could not be traced back to theirmothers’ spontaneous gestures (Goldin-Meadow and Mylander 1983,1984; Goldin-Meadow et al. 1995), nor could their grammatical cate-gories (Goldin-Meadow et al. 1994) or many of their communicativefunctions (Butcher et al. 1991; Morford and Goldin-Meadow 1997).Indeed, the gestures the parents produced appeared to be no differentfrom the gestures that any hearing individual uses along with speech(Goldin-Meadow, McNeill and Singleton 1996) and, as such, are globaland synthetic in form with structure quite different from the structureof natural language (McNeill 1992). The surprising result is that the chil-dren’s gestures are structured so much like natural language eventhough their parents’ gestures, which are likely to serve as input tothose gestures, are not.

Parental responsiveness to the deaf child’s gesturesIt is possible, however, that the structure in the children’s gesture sys-tems came from other nonlinguistic aspects of their environment. Forexample, by responding with either comprehension or noncomprehen-sion to their children’s gestures, the hearing parents of the deaf childrenmight have (perhaps inadvertently) shaped the structure of those ges-tures. However, we found that the mothers responded with compre-hension to approximately half of each child’s gesture strings—whetheror not those strings followed the child’s preferred orders. In otherwords, the mothers were just as likely to understand and act on the chil-dren’s ill-formed strings as their well-formed strings, suggesting thatthese particular patterns of parental responsivity did not shape the

Learning with and without a Helping Hand 127

orders that the children developed in their gesture systems (Goldin-Meadow and Mylander 1983, 1984).

There seems little doubt that comprehensibility determined the formof the deaf children’s gestures at a general level—the children’s ges-tures were iconic, with gesture forms transparently related to the in-tended meanings. Indeed, the overall iconicity of the children’s gesturesmay have contributed to the fact that variations in gesture order had lit-tle effect on the parents’ comprehension—a mother could easily figureout that her child was describing apple eating whether the childpointed at the apple before producing an EAT gesture, or produced theEAT gesture before pointing at the apple. Thus, although the children’sgestures were quite comprehensible to the hearing individuals aroundthem, there was no evidence that the structural details of each child’sgesture system were shaped by the way in which the mothers respondedto those gestures—we found no evidence that the child was given ahelping hand by the mother.

Parent-child interaction and its effect on the deaf child’s gestures: A look acrossculturesNevertheless, there may be other, more subtle ways in which parent-child interaction affects child communication. For example, Bruner(1974/1975) has suggested that the structure of joint activity betweenmother and child exerts a powerful influence on the structure of thechild’s communication. To determine the extent to which the structurein the deaf children’s gestures is a product of the way in which mothersand children jointly interact in their culture—and in so doing, develop amore stringent test of the resilience of the deaf children’s gesture sys-tems—we have begun a study of deaf children of hearing parents in asecond culture, a Chinese culture. The literature on socialization (Miller,Mintz, and Fung 1991; Young 1972), task-oriented activities (Smith andFreedman 1982), and academic achievement (Chen and Uttal 1988;Stevenson, Lee, Chen, Stigler, Hsu, and Kitamura 1990) suggests thatpatterns of mother-child interaction in Chinese culture differ greatlyfrom those in American culture, and we have replicated these differ-ences in our own studies of interaction between hearing mothers andtheir deaf children in Chinese and American families (Wang 1992;Wang, Mylander, and Goldin-Meadow 1995).

The salient differences between Chinese and American maternal in-teraction patterns provide us with an excellent opportunity to examinethe role that mother-child interaction plays in the development of thegestural communication systems of deaf children. If, as our currentwork suggests (Goldin-Meadow and Mylander 1998), there are similar-ities between the spontaneous gestural systems developed by deaf chil-

128 Susan Goldin-Meadow

dren in Chinese culture and deaf children in American culture, an in-creasingly powerful argument can be made for the noneffects ofmother-child interaction patterns on the development of these gesturalsystems—that is, we will have increasingly compelling evidence for theresilience of the linguistic properties found in the deaf children’s ges-tural systems. Conversely, to the extent that the gestural systems of theChinese deaf children are consistently different from the American deafchildren’s gestural systems, an equally compelling argument can bemade for the effects of cultural variation—as instantiated in mother-child interaction patterns—on the spontaneous gestural systems of deafchildren.

Resilience in the Face of External and Internal Variability: Equifinality

The phenomenon of gesture creation suggests that language develop-ment is resilient across environmental conditions that vary dramaticallyfrom the typical. However, language is resilient not only in the face ofexternal variation,but also in the face of organic variation. For example,the acquisition of grammar in the earliest stages has been found to pro-ceed in a relatively normal manner and at a normal rate even in the faceof unilateral brain injury (Feldman 1994). As a second example, chil-dren with Down’s syndrome have numerous intrinsic deficiencies thatcomplicate the process of language acquisition; nevertheless, most ofthese children acquire some basic language reflecting the fundamentalgrammatical organization of the language they are exposed to (theamount of language that is acquired is in general proportion to theircognitive capabilities, Rondal 1988; see also Fowler et al. 1994).

Thus human language appears to naturally assume a certain form,and that form can be reached through a wide range of developmentalpaths, some varying from the norm in terms of external factors, some interms of internal factors. In other words, language development is char-acterized by “equifinality”—a term coined by the embryologist Driesch(1908, as reported in Gottlieb 1995) to describe a process by which a sys-tem reaches the same outcome despite widely differing input condi-tions. Are there any implications for mechanisms of development thatwe can now draw having identified language learning as equifinal? Atleast two types of systems seem possible:

(1) A system characterized by equifinality can rely on a single devel-opmental mechanism that not only can make effective use of a widerange of inputs (both external and internal) but will not veer off track inresponse to that variability; that is, on a mechanism that is not sensitiveto large differences in input. The image that comes to mind here is asausage machine that takes inputs of all sorts and, regardless of the

Learning with and without a Helping Hand 129

130 Susan Goldin-Meadow

type and quality of that input, creates the same (at least on one level)product.

(2) A system characterized by equifinality can rely on multiple devel-opmental mechanisms, each activated by different conditions but con-strained in some way to lead to the same endproduct (cf. Miller,Hicinbothom, and Blaich 1990). The analogy here is to four distinct machines, each one designed to operate only when activated by a par-ticular type of input (e.g., a chicken, pig, cow, or turkey); despite the dif-ferent processes that characterize the dismembering operations of eachmachine, the machines result in the same sausage product. At firstglance, it may seem improbable that a variety of developmental mecha-nisms would be constrained to arrive at precisely the same outcome.However, it is relatively easy to imagine that the function served by themechanisms—a function that all of the developmental trajectorieswould share, such as communicating via symbols with other humans(cf. Goldin-Meadow et al. 1996)—might have been sufficient, over time,to constrain each of the mechanisms to produce the same product.

The findings that we have assembled thus far on the gesture systemscreated by deaf children do not allow us to distinguish between thesetwo hypothetical mechanisms—not yet. Nevertheless, it is certain thatby continuing to compare the process of language learning in typicaland atypical circumstances—the strategy born in the Gleitman livingroom—we will approach a more complete understanding of how chil-dren learn language.

Learning with a Helping Hand: Gesture’s Role in the Learning Process

One of the questions that plagued me for quite some time in my workon the deaf children’s gesture systems was this: If the linguistic proper-ties listed above are so resilient, why don’t they appear in the gesturesthat the deaf children’s hearing parents use? I have since come to realizethe answer—the parents’ gestures were not “free” to assume the lan-guage-like structure found in their children’s gestures simply becausethe parents always produced their gestures while talking. Gesture andspeech in hearing individuals form a single integrated system—the twomodalities work together to convey the speaker’s intended message,with speech assuming the segmented and combinatorial form that char-acterizes natural languages, and gesture assuming an imagistic andholistic form (McNeill 1992; when speech is absent, as in sign languagesand the deaf child’s gesture system, it is the manual modality that as-sumes the segmented and combinatorial form that characterizes naturallanguages). My second research program explores the integrated ges-

ture-speech system in its own right, particularly in relation to learningover the short-term.

The Relationship between Gesture and Speech as an Index of Readiness-to-Learn

My students and I have examined the gestures that hearing childrenspontaneously produce when explaining their solutions to a task. Wecoded gesture and speech independently and made an interesting dis-covery—at times, the children conveyed one message in speech and an-other in gesture. For example, a young child asked to solve a liquidquantity conservation task says that the transformed object is differentfrom the original because “this one is taller than this one” but, in thesame response, produces a gesture reflecting an awareness of the widthsof the objects; specifically, she indicates with her hands the skinny di-ameter of the original object and the wider diameter of the transformedobject, thus revealing knowledge of the widths of the task objects thatwas not evident in her speech (Church and Goldin-Meadow 1986). Wehave labeled instances in which gesture and speech convey different in-formation in a problem-solving situation “mismatches.”

Some children produce many gesture-speech mismatches on a giventask while others produce few. Moreover, children who produce a rela-tively large number of mismatches in their explanations of a particulartask (e.g., a conservation task or a mathematical equivalence task) aremore likely to benefit from instruction in that task than children whoproduce few mismatches (Alibali and Goldin-Meadow 1993; Churchand Goldin-Meadow 1986; Goldin-Meadow, Alibali, and Church 1993;Perry, Church, and Goldin-Meadow 1988, 1992). Thus gesture-speechmismatch signals that the child is in a transitional state with respect to atask, and is therefore ready to make progress on that task if given ap-propriate input.

Why does gesture-speech mismatch index readiness-to-learn? Whena child produces a mismatch, that child is, by definition, activating twonotions on the same problem—one displayed in speech and one in ges-ture. We suggest that the activation of two ideas on the same problem is in-deed what characterizes the transitional state and what may destabilizethe learner so that input can have an effect. If children who produce alarge number of mismatches on a task are, in fact, activating two no-tions every time they solve a task of that type, they should be expendingmore effort in reaching their incorrect solutions on the task than chil-dren who produce few mismatches (who also solve the problems incor-rectly but should do so more efficiently). Evidence from a cognitive load

Learning with and without a Helping Hand 131

task shows that children who produce many mismatches in their expla-nations of a math task, when later asked to solve the task but not explainit, expend more effort on the task (as gauged by their performance on asimultaneously performed word recall task) than do children who pro-duce few mismatches (Goldin-Meadow, Nusbaum, Garber, and Church1993).

Gesture’s Effect on the Learner and the Learning Situation

A difference—or mismatch—between the information conveyed in ges-ture and the information conveyed in speech appears to reflect the acti-vation of two notions on the same problem and, as a result, to signalreadiness for cognitive growth. It is an open question whether the ac-tual production of gesture-speech mismatch contributes to change—that is, does the act of producing two different pieces of informationacross modalities but within a single communicative act improve thelearner’s ability to transpose that knowledge to a new level and thusproduce those pieces of information within a single modality? Some ev-idence suggests that it might (gesturers were somewhat more likely toimprove after instruction on a math task than nongesturers; Alibali andGoldin-Meadow 1993) but more work is needed to determine whetherthe act of producing mismatches itself facilitates transition. We aren’tyet sure whether “sitting on your hands” is something more than ametaphor.

Even if it turns out that the production of gesture-speech mismatchesdoes not affect learners themselves and thus does not contribute to thelearning process directly, it may still play an indirect role in the processby shaping the learning environment. The information conveyed in thegestured component of a child’s mismatch is frequently not found any-where in that child’s speech (Alibali and Goldin-Meadow 1993). For ex-ample, a child who indicates the width of the containers in the gesturedcomponent of her mismatch is not likely to demonstrate an understand-ing of width in her speech on any of the liquid quantity trials (interest-ingly, the reverse does not hold—the information conveyed in thespoken component of a mismatch is quite likely to be conveyed by thatchild in gesture on other trials; Garber, Alibali and Goldin-Meadow1998). In other words, gesture often conveys information about a child’sunderstanding that is not conveyed anywhere in that child’s spokenrepertoire. As such, gesture provides a unique window into the thoughtsof the learner.

If those who interact with children in learning situations were able to“read” a child’s gestures, they would gain access to the unique informa-tion conveyed in gesture. Knowledge about a child’s understanding of

132 Susan Goldin-Meadow

a task gained through that child’s gesture might then affect how theadult interacts with the child on the task. For example, an adult who has“read” a child’s width gesture on a conservation task might then treatthat child as though she understood the importance of width in judgingquantity, an instructional stance that could promote in the child an ex-plicit understanding of width. We have taken several steps in exploringthis hypothesis. We have shown that adults can accurately interpret thegestures that children produce when those gestures are replayed onvideotape (Alibali, Flevares, and Goldin-Meadow 1997; Goldin-Meadow,Wein, and Chang 1992) and when they are observed “live” in fleetingreal time (Goldin-Meadow and Sandhofer 1999). In fact, even childrenare able to read the gestures produced by other children (Kelly andChurch 1997).

Our final step—to discover whether the information adults gain froma child’s gestures affects the way they interact with that child—is still inthe “in progress” stage. We have, however, shown that the gesturesspontaneously produced in a teaching situation can affect the partici-pants—in this case, the teacher’s gestures affected the learner’s responses(Fernandez, Flevares, Goldin-Meadow, and Kurtzberg 1996). Whenasked to instruct children individually in a mathematical equivalenceproblem, each of two teachers produced at least one gesture-speechmismatch with each child (it is not clear what prompted the mis-matches; one hopes uncertainty about how best to teach the concept,rather than teacher uncertainty about the concept itself). What was in-teresting was that the children were much less likely to merely repeatwhat the teacher said when responding to a teacher mismatch thanwhen responding to a teacher match or no-gesture turn. We have sincereplicated this pattern with eight additional teachers, each individuallyinstructing a series of children (Goldin-Meadow, Kim, and Singer 1999).These findings are the first to suggest that the spontaneous gesturesproduced by the participants in a learning situation can alter the courseof instruction—that learning may be influenced by the helping hand.

Summary

To summarize, I have explored the gestures of children in two very dif-ferent learning contexts—in deaf children whose only form of effectivecommunication is gesture, and in hearing children whose gestures rou-tinely and naturally accompany speech. My program of research on thedeaf children, launched in the Gleitman living room, has shown that notonly does gesture in this context take on the functions language serves,it also takes on its forms—and it does so without benefit of an explicitmodel for those forms, that is, without a helping hand. In contrast, inhearing children, gesture coexists with speech and assumes neither the

Learning with and without a Helping Hand 133

functions nor the forms of language. Rather, gesture assumes a formatthat is less analytic and more imagistic than speech (and than the deafchild’s gestures)—a format that complements the segmented and com-binatorial structures found in speech. In this context, gesture is free toexpress ideas that are not easily incorporated into speech and, as such,has the potential to lend a helping hand to the learning process.

In closing, I’d like to say what a privilege it is to have been mentoredby Lila and Henry. Even more than learning what good research is andhow it should be done, I have learned from Lila and Henry how to be agood teacher. By example and by explicit instruction, I learned how toteach in large classes, in small discussions, and in interactions with in-dividual students, and I learned that good teaching is something to bevalued highly and worked toward. I fear that I may have set standardsthat are just too high when I ask myself to be as superb a mentor to myown students as Lila and Henry were to all of us. But I figure I might aswell aim high, and I thank Lila and Henry from the bottom of my heartfor showing me how good it can get.

Acknowledgment

The work described here was supported by grants from the March ofDimes, the Spencer Foundation, the National Institutes of Child Healthand Human Development (R01 HD18617), and the National Instituteon Deafness and other Communication Disorders (RO1 DC00491).

Notes

1. “Cookie EAT” is a sentence consisting of two gestures. Deictic pointing gestures aredisplayed in lower case letters, iconic gestures in capital letters. The boundary of a ges-ture sentence is determined by motoric criteria. If the hand is relaxed or returned toneutral position (chest level) prior to the onset of the next gesture, each of the two ges-tures is considered a separate unit. If there is no relaxation of the hand between the twogestures, the two are considered part of a single gesture sentence.

References

Alibali, M. W. and Goldin-Meadow, S. (1993) Gesture-speech mismatch and mechanismsof learning: What the hands reveal about a child’s state of mind. CognitivePsychology 25:468–523.

Alibali, M., Flevares, L., and Goldin-Meadow, S. (1997) Assessing knowledge conveyed ingesture: Do teachers have the upper hand? Journal of Educational Psychology89:183–193.

Bruner, J. (1974/1975) From communication to language: A psychological perspective.Cognition 3:255–287.

Butcher, C., Mylander, C., and Goldin-Meadow, S. (1991) Displaced communication in aself-styled gesture system: Pointing at the non-present. Cognitive Development6:315–342.

134 Susan Goldin-Meadow

Chen, C. and Uttal, D. H. (1988) Cultural values, parents’ beliefs, and children’s achieve-ment in the United States and China. Human Development 31:351–358.

Church, R. B. and Goldin-Meadow, S. (1986) The mismatch between gesture and speechas an index of transitional knowledge. Cognition 23:43–81.

Driesch, H. (1908/1929) The Science and Philosophy of the Organism. London: A. and C.Black.

Feldman, H. M. (1994) Language development after early unilateral brain injury: A repli-cation study. In Constraints on language acquisition: Studies of atypical children, ed. H.Tager-Flusberg. Hillsdale, NJ: Erlbaum Associates, 75–90.

Feldman, H., Goldin-Meadow, S., and Gleitman, L. (1978) Beyond Herodotus: The cre-ation of a language by linguistically deprived deaf children. In Action, Symbol, andGesture: The Emergence of Language, ed. A. Lock. New York: Academic Press,351–414.

Fernandez, E., Flevares, L., Goldin-Meadow, S., and Kurtzberg, T. (1996) The role of thehand in teacher-student interaction. Paper presented at the annual meeting ofAERA, New York, April 1996.

Fowler, A. E., Gelman, R., and Gleitman, L. R. (1994) The course of language learning inchildren with Down’s syndrome: Longitudinal and language level comparisonswith young normally developing children. In Constraints on language acquisition:Studies of atypical children, ed. H. Tager-Flusberg. Hillsdale, NJ: Erlbaum Asso-ciates, 91–140.

Garber, P., Alibali, M. W., and Goldin-Meadow, S. (1998) Knowledge conveyed in gestureis not tied to the hands. Child Development 69:75–84.

Gleitman, L. R. (1965) Coordinating conjunctions in English. Language 41:260–293.Gleitman. L. R. and Newport, E. L. (1995) The invention of language by children:

Environmental and biological influences on the acquisition of language. InLanguage, Volume 1, Invitation to Cognitive Science Series, ed. L. R. Gleitman and M.Liberman. Cambridge, MA: MIT Press, 1–24.

Goldin-Meadow, S. (1982) The resilience of recursion: A study of a communication sys-tem developed without a conventional language model. In Language Acquisition:The State of the Art, ed. E. Wanner and L. R. Gleitman. New York: CambridgeUniversity Press, 51–87.

Goldin-Meadow, S. (1985) Language development under atypical learning conditions:Replication and implications of a study of deaf children of hearing parents. InChildren’s Language, volume 5, ed. K. Nelson. Hillsdale, NJ: Lawrence Erlbaum andAssociates, 197–245.

Goldin-Meadow, S. (1987) Underlying redundancy and its reduction in a language devel-oped without a language model: The importance of conventional linguistic input.In Studies in the acquisition of anaphora, Volume II: Applying the constraints, ed. B.Lust. Boston: D. Reidel Publishing Company, 105–133.

Goldin-Meadow, S. (1993) When does gesture become language? A study of gesture usedas a primary communication system by deaf children of hearing parents. In Tools,Language, and Cognition in Human Evolution, ed. K. R. Gibson and T. Ingold. NewYork: Cambridge, 63–85.

Goldin-Meadow, S., Alibali, M. W., and Church, R. B. (1993) Transitions in concept acqui-sition: Using the hand to read the mind. Psychological Review 100:279–297.

Goldin-Meadow, S., Butcher, C., Mylander, C., and Dodge, M. (1994) Nouns and verbs ina self-styled gesture system: What’s in a name? Cognitive Psychology 27:259–319.

Goldin-Meadow, S. and Feldman, H. (1977) The development of language-like communi-cation without a language model. Science 197:401–403.

Goldin-Meadow, S., Kim, S., and Singer, M. (1999) What the teacher’s hands tell the stu-dent’s mind about math. Journal of Educational Psychology 91:720–730.

Learning with and without a Helping Hand 135

Goldin-Meadow, S., McNeill, D., and Singleton, J. (1996) Silence is liberating: Removingthe handcuffs on grammatical expression in the manual modality. PsychologicalReview 103:34–55.

Goldin-Meadow, S. and Mylander, C. (1983) Gestural communication in deaf children:Non-effect of parental input on language development. Science 221:372–374.

Goldin-Meadow, S. and Mylander, C. (1984) Gestural communication in deaf children:The effects and non-effects of parental input on early language development.Monographs of the Society for Research in Child Development 49, no. 207.

Goldin-Meadow, S. and Mylander, C. (1990) Beyond the input given: The child’s role inthe acquisition of language. Language 66:323–355.

Goldin-Meadow, S. and Mylander, C. (1998) Spontaneous sign systems created by deafchildren in two cultures. Nature 391:279–281.

Goldin-Meadow, S., Mylander, C., and Butcher, C. (1995) The resilience of combinatorialstructure at the word level: Morphology in self-styled gesture systems. Cognition56:195–262.

Goldin-Meadow, S., Nusbaum, H., Garber, P., and Church, R. B. (1993) Transitions inlearning: Evidence for simultaneously activated strategies. Journal of ExperimentalPsychology: Human Perception and Performance 19:92–107.

Goldin-Meadow, S. and Sandhofer, C. M. (1999) Gesture conveys substantive informationabout a child’s thoughts to ordinary listeners. Developmental Science 2:67–74.

Goldin-Meadow, S., Wein, D., and Chang, C. (1992) Assessing knowledge through ges-ture: Using children’s hands to read their minds. Cognition and Instruction9:201–219.

Gottlieb, G. (1996) A systems view of psychbiological development. In The lifespan devel-opment of individuals: behavioral, neurobiological, and psychosocial perspectives (pp.76–103). ed. D. Magnusson. New York: Cambridge University Press.

Kelly, S. D. and Church, R. B. (1997) Children’s ability to detect nonverbal behaviors fromother children. Cognition and Instruction 15:107–134.

Landau, B., and Gleitman, L. R. (1985) Language and Experience: Evidence from the BlindChild. Cambridge, MA: Harvard University Press.

McNeill, D. (1992) Hand and Mind: What Gestures Reveal about Thought. Chicago:University of Chicago Press.

Miller, D. B., Hicinbothom, G., and Blaich, C. F. (1990) Alarm call responsivity of mallardducklings: Multiple pathways in behavioural development. Animal Behavior39:1207–1212.

Miller, P. J., Mintz, J., and Fung, H. (1991) Creating children’s selves: An American andChinese comparison of mothers’ stories about their toddlers. Paper presented atthe biennial meeting of the Society for Psychological Anthropology, October 1991.

Morford, J. P. and Goldin-Meadow, S. (1997) From here to there and now to then: The de-velopment of displaced reference in homesign and English. Child Development68:420–435.

Newport, E. L., Gleitman, H., and Gleitman, L. R. (1977) Mother I’d rather do it myself:Some effects and non-effects of maternal speech style. In Talking to children, ed. C.E. Snow and C. A. Ferguson. New York: Cambridge University Press, 109–150.

Perry, M., Church, R. B. and Goldin-Meadow, S. (1988) Transitional knowledge in the ac-quisition of concepts. Cognitive Development 3:359–400.

Perry, M., Church, R. B., and Goldin-Meadow, S. (1992) Is gesture-speech mismatch ageneral index of transitional knowledge? Cognitive Development 7:109–122.

Rondal, J. A. (1988) Down’s syndrome. In Language Development in ExceptionalCircumstances, ed. D. Bishop and K. Mogford. New York: Churchill Livingstone,165–176.

136 Susan Goldin-Meadow

Singleton, J. L., Morford, J. P., and Goldin-Meadow, S. (1993) Once is not enough:Standards of well-formedness in manual communication created over three differ-ent timespans. Language 69:683–815.

Smith, S. and Freedman, D. (1982) Mother-toddler interaction and maternal perception ofchild development in two ethnic groups: Chinese-American and European-American. Paper presented at the annual meeting of the Society for Research inChild Development, Detroit.

Stevenson, H. W., Lee, S.-L., Chen, C., Stigler, J. W., Hsu, C.-C., and Kitamura, S. (1990)Contexts of achievement. Monographs of the Society for Research in Child Development55, no. 221.

Thompson, S. A. (1988) A discourse approach to the cross-linguistic category “adjective.”In Explaining Language Universals, ed. J. A. Hawkins. Cambridge, MA: BasilBlackwell, 167–185.

Wang, X-L. (1992) Resilience and fragility in language acquisition: A comparative studyof the gestural communication systems of Chinese and American deaf children.Unpublished Ph.D. dissertation, University of Chicago.

Wang, X-L., Mylander, C., and Goldin-Meadow, S. (1995) The resilience of language:Mother-child interaction and its effect on the gesture systems of Chinese andAmerican deaf children. In Language, Gesture, and Space, ed. K. Emmorey and J.Reilly. Hillsdale, NJ: Erlbaum Associates, 411–433.

Learning with and without a Helping Hand 137

Chapter 10

The Detachment Gain: The Advantage of ThinkingOut Loud

Daniel Reisberg

The weekly meetings of the Gleitman research seminar were for manyof us an extraordinary opportunity to learn psychology—its substance,its methods, and its history. In that seminar, we gained from others’comments and criticisms of our work, and we also gained simply fromthe opportunity to articulate our ideas before this sophisticated and de-manding audience. Likewise, Henry and Lila Gleitman were (in thisseminar as in all other contexts) eloquent and effective, articulate andpersuasive. They were, perhaps, merely “thinking out loud,” but it wasthinking of a remarkable sort, one that benefited all of us enormously.

Two decades later, however, I am led to ask: What does it mean to“think out loud”? Does one merely articulate one’s thoughts, or (asseems more likely) must one “translate” the thoughts into some ex-pressible form? If the latter, what gains and what losses derive from thistranslation?

These questions are tied to a number of research problems on the cur-rent scene (Crutcher 1994; Ericsson and Simon 1980; Payne 1994;Schooler, Ohlsson, and Brooks 1993; Wilson 1994), but these questionsare of course also tied to a much older theoretical framework: Early inthis century, John Watson took a strong stand on the relation between“thinking” and “talking,” arguing that the former was merely a covertversion of the latter; “thinking out loud,” therefore, was no trick at all,but merely the overt voicing of representations (for Watson: “responsesequences”) already in linguistic form.

Watson intended that this argument be understood quite literally. Infact, Henry Gleitman is fond of quoting a quip from Watson, emphasiz-ing this point: What’s the difference, Watson asked, between someonewith laryngitis and a congenital moron? As Henry tells it, Watson’s an-swer was simple: Someone with laryngitis will get better.

Modern psychologists tend to scoff at these claims. Thought is not,we all know, covert speech. But one might still wonder whether there isan element of truth in Watson’s observations. Could there be some in-tellectual impairment associated with laryngitis? Or to put this more

realistically, what is it that happens when we think out loud (or writeout our thoughts, or express our thoughts in sign language)? Is thereany functional advantage in putting thoughts into these “externalized”forms—judgments we can make or problems we can solve that wecouldn’t solve with silent thought?

As a related question, Roger Shepard made this observation someyears back: Imagine that you’ve been asked how many windows thereare in your home or apartment. Many people tackle this task by closingtheir eyes and drawing the perimeter of their home in the air with anindex finger. Why this peculiar behavior? Is there any advantage to ex-ternalizing the representation in this (motoric) fashion?

To anticipate the data, it turns out that we do benefit in many tasksfrom this sort of externalization—be it talking out loud or just subvocal-izing, and likewise whether it’s drawing on paper or merely drawing inthe air. However, not all tasks show this “externalization benefit,” andthis invites the crucial question: Why is it that some tasks can be doneperfectly well based only on the information contained in a mental rep-resentation, whereas other tasks benefit from externalization?

As a way of entering this discussion, consider the following proce-dure. Subjects were shown a number of letter and number strings, suchas D-2-R, or N-C-Q-R. The subjects’ task was to discern what word orphrase these strings would produce, if pronounced aloud. (In the twoexamples just given, the solutions would be “detour” and “insecure.”)

Once subjects had learned to play this game, they were given a seriesof such strings, presented visually, and asked to write down their re-sponses. In one condition, subjects did this with no further require-ments beyond the stipulation that they make no noises while workingon the puzzles. They were forced, in other words, to form images ofthese sounds and to make their judgments based on the images. In asecond condition, subjects were blocked from talking to themselveswhile doing the task. This was achieved by requiring subjects to say “ta-ta-ta,” aloud, over and over, while working on the puzzles. This manip-ulation severely impaired performance. Apparently, subjects need tosubvocalize to do this task. In a third condition, subjects did the puzzlesin a noisy environment, and this too impaired performance. Thus sub-jects need to hear themselves thinking to do this task. Finally, in a fourthcondition, subjects had to say “ta-ta-ta” aloud and were in a noisy envi-ronment. This too was disruptive, but no more so than either of the in-dividual manipulations of covert speech or hearing by themselves.

One might worry that these effects simply reflect the distracting effectof noise, such that performance is better in a quiet environment than anoisy one. In two of the conditions just described, the noise is presentedby the experimenter; in two other conditions, the noise is produced by

140 Daniel Reisberg

the subjects’ own pronunciation of ta-ta-ta. In either case, the manipula-tions do produce sound, and this may lead to distraction. To check onthis possibility, a further study compared three conditions. The first wasa control condition in which subjects simply solved the puzzles alreadydescribed. In another condition, subjects did this task while saying “ta-ta-ta” aloud, over and over. In the final condition, subjects were askedto solve these puzzles while gritting their teeth together, pushing theirtongue up against the roof of their mouths, and pushing their lips to-gether. This “clamping” is silent, but occupies the muscles needed forsubvocalization, and also preempts central circuits needed to controlthis subvocalization. And clamping was indistinguishable in its effectsfrom the ta-ta-ta manipulation; both were reliably worse than the con-trol procedure. It would seem, therefore, that it is the loss of subvocal-ization, and not simply the presence of noise, that produces theseinterference effects.

A number of other tasks show a similar data pattern. For example, inanother study, subjects were given the titles of highly familiar songs(“Happy Birthday to You”; “London Bridge”), and were asked whetherthe melody of each song rises or falls from the second note to the third(rises for “Birthday,” falls for “Bridge”). Again, subjects did this in oneof four conditions: In two of the conditions, covert speech was blocked(by the requirement of concurrent articulation); in two, hearing wasblocked (by the presentation of task-irrelevant noises). The data, as be-fore, show good performance in the control condition, and impairedperformance with either concurrent articulation or ambient noise. Thedouble-interference condition (concurrent articulation and ambient noise)produced performance levels similar to those in the single-interferenceconditions.1

These data make it clear that auditory imagery does in some cases de-pend on subvocalized support. When this support is denied, subjectsare unable to use their images for some purposes. More, this sub-vocalization seems to create a quasi-perceptual representation: If thechannels of audition are unavailable (because of background noise),performance again suffers. Perhaps, then, Watson was not altogethermistaken: Thought sometimes does require enactment. This enactment,even if covert, produces a stimulus to which we are then able to re-spond.

However, we need to put these data into a broader context and, inparticular, the context provided by other findings showing that tasks re-quiring auditory imagery do not always rely on subvocalization. Forexample, Baddeley and Andrade (1995) simply asked subjects to reflecton the subjective vividness of their auditory imagery while doing theta-ta-ta task. Their subjects reported a slight decrease in image vivid-

The Detachment Gain 141

ness under these circumstances, but, according to their self-report, theywere still able to maintain their (slightly impoverished) images. Itwould seem, then, that concurrent articulation need not devastate audi-tory imagery. Put differently, we can sometimes create auditory repre-sentations without subvocalized support.

In addition, common sense strongly argues that some auditory im-ages do not rely on subvocalization. After all, most of us can easilyimagine sounds outside of the human vocal repertoire—the sounds ofglass breaking or of brakes squealing. We cannot vocalize these soundsand so presumably we cannot subvocalize them either. Therefore, forthese sounds, imagery cannot depend on subvocalization.

This last claim, however, like the Baddeley and Andrade result, relieson an introspective observation. Is there some way to corroborate thisclaim? One line of evidence comes from Crowder and Pitt (1991) and fo-cuses on subjects’ ability to imagine the timbres of various orchestral in-struments. The key assumption here is that these timbres are outside ofsubjects’ vocal repertoires, and so, if subjects can imagine these sounds,this would demonstrate the ability to imagine sounds without subvo-calized support.

Crowder and Pitt’s data on this issue are instructive, but, in the inter-ests of brevity, I note simply that the procedure they use provides only acoarse portrait of auditory imagery. Concretely, their data seem to showus that subjects’ auditory image of, say, a flute is more like a flute than it islike a tuba. But whether the flute image really “sounds like” a flute is an-other question. Thus Crowder and Pitt’s findings tell us little about theactual fidelity of these images, and, in particular, tell us little aboutwhether subjects can, to some criterion, accurately imagine these sounds.

Concerns such as this led us to seek a new paradigm to explore these is-sues. In our studies, subjects were asked to imagine two instrumentsounds, say, the sound of a bassoon and the sound of an oboe. They werethen asked to assess how similar these images were to each other on aseven-point similarity scale. Then they did the same with two other in-struments, say, a clarinet and a violin, and so on for a number of otherpairs. Once subjects had compared their images in this way, we were ableto do a multidimensional scaling of their data in order to ask what the na-ture of the similarity space is for describing these various auditory images.

We ran a parallel procedure with subjects actually hearing thesesounds, rather than imaging them, and making the same pair-wise sim-ilarity judgments on these perceived sounds. Thus we can also assessthe multidimensional similarity space for perception.

Our hope was that the multidimensional space for imagined soundswould match reasonably well with the multidimensional space for per-ceived sounds. If this came to pass, it would be an indication that sub-jects’ images preserve all of the pair-wise similarities available in actual

142 Daniel Reisberg

perception. And, if the full pattern of pair-wise relationships is pre-served, then it seems fair to argue that subjects’ images are reproducingthese sounds with some fidelity.

Our data were very much in line with these predictions. Subjects inall conditions first heard “reminders” of these instrument sounds,played via computer; subjects could repeat this exposure sequence asmany times as they wished until they felt certain that they knew thesesounds well. (All stimuli presented in this experiment were digitizedversions of actual instrument sounds.) Subjects in the imagery condi-tion were then given a series of trials in which only the instrument’sname was supplied, and they were instructed to practice forming im-ages of these timbres. Subjects then received the test trials, with two instrument names presented on each trial and subjects making a judg-ment about the similarity between the sounds of the two instrumentsnamed. Subjects in the perception condition ran through roughly thesame procedure, but made their judgments based on actual sounds,rather than images.

The data were averaged across subjects and then run through a stan-dard multidimensional scaling algorithm, seeking the best reduction ofthe pair-wise relationships. Figure 10.1 shows the data from our im-agery condition. The data are quite orderly, with the y-axis correspond-ing roughly to rise-time (abrupt starts are high, gradual starts are low)and with the x-axis corresponding to the pattern of harmonics (withbrighter sounds to the right and duller sounds to the left). The patternalso agrees reasonably well with that reported by Iverson and Krum-hansl (1993), using actually perceived sounds, with the main aberrationlying in the flute, which one would expect, based on their data, to belower and to the left in this plot.

Figure 10.2 shows the data from our perception condition, and thematch to the imagery data is reasonably good, but not perfect. For ex-ample, the bells, piano, and vibes cluster together as one group in bothplots; in both plots, the cello and tenor saxophone are reasonably closeto each other. But contrasts between the two plots are also easy to find:The bassoon, for example, appears in the lower right quadrant for theimagery plot, but in the upper left for the perception data.

These visual comparisons of the two plots, however, are difficult tointerpret, largely because the axes in this (or any) multidimensionalscaling are arbitrary. To compare these two plots, therefore, we need amore sophisticated axis-independent assessment. One way to do this issuggested by Besag and Diggle (1977). This analysis begins by comput-ing the point-to-point distances for every pair of points in the percep-tion plot, and the same in the imagery plot. Next, we calculate thecorrelation between these two sets of distances. (That is, one data pair is the flute-to-clarinet distance in the imagery data, and the flute-to-

The Detachment Gain 143

144 Daniel Reisberg

Figure 10.1.The two dimensional space summarizing subjects’ pairwise similarity ratings of their men-tal images of the sounds produced by various orchestral instruments. The y axis seems tocorrespond roughly to rise-time (with abrupt starts high, gradual starts low); the x axisseems to reflect the pattern of harmonics (with brighter sounds to the right and dullersounds to the left). See Figure 10.2 for corresponding data obtained with perceived sounds.

The Detachment Gain 145

Figure 10.2.The two dimensional space summarizing subjects’ pairwise similarity ratings of their per-ceptions of the sounds of various orchestral instruments. As in Figure 10.1, the y axisseems to correspond roughly to rise-time and the x axis seems to reflect the pattern ofharmonics (with brighter sounds to the right and duller sounds to the left).

clarinet distance in the perception data. A second data pair is the clar-inet-to-oboe distance in the imagery data, and the corresponding dis-tance in the perception pair. And so on.) The correlation calculated inthis way is 0.455 (p<0.001), indicating that there is a reasonable concor-dance between these two sets of distances, and thus reasonable align-ment between these two multidimensional plots.2

Let us be careful, though, not to overinterpret these findings. As wehave already acknowledged, the correspondence between figures 10.1and 10.2 is not perfect. Perhaps this merely reflects the difficulty of mak-ing comparisons between two multidimensional spaces. Perhaps it re-flects the difficulty, for our subjects, of remembering the full set ofinstrument sounds. (What does an English horn sound like, as opposedto an oboe?) Or perhaps it does reflect some limitations in the quality ofsubjects’ auditory imagery.

Even with these cautions in view, the fact remains that there is a rea-sonable and highly reliable correspondence in this task between subjects’perceptual and imagery judgments. This seems to be pushing us, there-fore, toward an inelegant conception of auditory imagery, along the linesdepicted in figure 10.3. In this figure, there are two different pathsthrough which an auditory image can be created. In some cases, the im-age is created through subvocalization; this pathway seems crucial, forexample, in the D-2-R task. In other cases, the image is created with-out subvocalization; this seems to be the case in our study of instrumenttimbres.

Is this model necessary? Do we really need two pathways, throughwhich images can be created? Or can the data be explained in more par-simonious terms? One could argue, perhaps, that one’s image of (say) aflute sound, or bassoon, does depend on subvocalization. In this case,one could maintain the claim that all auditory images depend on sub-vocalized support, eliminating the top pathway shown in the figure.

There are several ways this claim might be developed and defended(cf. Baddeley and Logie 1992; Smith, Wilson, and Reisberg 1992), but itis a claim, in any case, easily addressed empirically. Two further groupsof subjects were run through our instrument timbres procedure. Onegroup of subjects was asked to compare their images of the various mu-sical instruments, but was also blocked from subvocalizing. Thus eachtrial began with the instruction that subjects repeat “ta-ta-ta” aloud, andthen, a moment later, the test instruments for that trial were named, andsubjects made their similarity assessment. Subjects were then allowedto cease the articulation until the start of the next trial, when they beganagain. A second group of subjects did the same, but with actually per-ceived stimuli rather than imagined ones.

Again, the data were averaged across subjects and assessed via a mul-tidimensional scaling algorithm. If subjects’ images in this procedure de-

146 Daniel Reisberg

pend on subvocalized support, then these results should contrast withthe multidimensional spaces obtained without the requirement for con-current articulation. But that is not what the data show. Instead, themultidimensional spaces obtained in this newer study resemble eachother quite closely: A comparison of the imagery and perception datayields r = 0.704, p < 0.001. There is also a reasonable concordance be-tween these new imagery data (with concurrent articulation) and theimagery data shown in figure 10.1 (r = 0.679, p < 0.001). There is aweaker correspondence, but still a reliable correlation, between theseperception data (with concurrent articulation) and the perception datashown in figure 10.2 (r = 0.227, p < 0.001).

Thus we obtain roughly the same results if covert speech is permittedand available for use, or if covert speech is denied. Apparently, there-fore, the auditory images for this task are not being supported throughsubvocalization and so we can safely conclude that subjects can imag-ine sounds, with reasonable accuracy, without subvocalized support.

It truly seems, then, that we are driven toward the unparsimoniousmodel shown in figure 10.3. Some images are created and maintainedby means of covert speech; I will refer to those as “enacted images.”Other images are created without this enactment; those I will call “pure

The Detachment Gain 147

Figure 10.3.The data suggest that there are two ways to produce an auditory image. The top pathwayindicates “pure imagery,” created without subvocalized support. The bottom pathwayindicates “enacted” imagery, dependent on covert speech (or covert singing).

images.” (The terminology was first suggested by Reisberg, Smith,Baxter, and Sonenshine 1989.)

A broadly similar point can be made based on studies of working-memory rehearsal. It has long been known that this rehearsal is dis-rupted by concurrent articulation, and this has led many researchers tosuggest that this rehearsal relies on subvocalized speech. However, anumber of authors have raised questions about the interpretation of theworking-memory evidence (e.g., Gupta and MacWhinney 1995). Thesequestions led us to further experiments, and, as we will see, these ex-periments again highlight the “duplex” nature of auditory images, withsome images created via subvocalization and some created in a more di-rect fashion, with no need for covert speech.

To understand the debate over working-memory rehearsal, consideragain the two pathways shown in figure 10.3. Both pathways, I havesuggested, create a quasi-perceptual representation, a representationthat resides in a buffer within the auditory system. Use of this bufferwill be denied, however, if there are other noises present during the ex-perimental procedure. These noises will also land in the auditory buffer,perhaps displacing, perhaps simply mixing with, the representationneeded for imagery. In either of these cases, the incoming noise will dis-rupt the representation.

In contrast, imagine what will happen if subjects are blocked fromsubvocalizing during an experimental treatment. Let’s say that subjectsare forced to bite firmly on a pencil during the procedure, or to chew agreat big wad of gum. These activities produce little noise, and so theywill not disrupt the auditory buffer. But these activities do demand thearticulators, and so these articulators won’t be available for subvocal-ization. Thus these sorts of activities will disrupt enacted imagery, butnot pure. They will, in short, disrupt the bottom path in figure 10.3, butnot the top path.3

Finally, imagine that subjects are asked to say “ta-ta-ta” aloud duringan experimental task. This manipulation takes over the articulators andalso produces noise, and therefore should have the effects of disruptingpure imagery (because the activity is noisy) and also enacted imagery(because the activity is noisy and because it blocks covert speech).

The idea, in short, is that concurrent articulation—such as saying “ta-ta-ta” aloud—has a double effect, and this renders ambiguous the effectsof this manipulation. Put differently, if a task is disrupted by concurrentarticulation, this is not enough, by itself, to allow the conclusion that thetask relies on subvocalization. Instead, the task might rely on pure im-agery, and be disrupted by the noisiness of saying “ta-ta-ta” aloud.

We earlier confronted this same point in considering our D-2-R task,and there we showed that silent clamping of the articulators had the

148 Daniel Reisberg

same effect as concurrent articulation. We concluded that the D-2-Rpuzzles do indeed require subvocalized support. But the same issue,and the same ambiguity, emerges with memory rehearsal. Prior studiesmake it clear that concurrent articulation disrupts this rehearsal, butwhy is this? Is it because concurrent articulation preempts covertspeech? Or is it because concurrent articulation is noisy?

A number of researchers have reported evidence pertinent to thisissue, but, for our purposes, the most interesting studies are those thatprovide side-by-side comparisons between memory rehearsal and vari-ous imagery tasks. Several of our own procedures have been designedto yield these comparisons, but the clearest data on this issue come froma study by Dorothy Bishop at the Applied Psychology Unit in Cam-bridge. In half of the trials in her design, subjects were given a conven-tional digit-span task (with nine digits as the to-be-rememberedmaterial; spoken presentation; written recall). In the other half of the tri-als, the same subjects were tested for their ability to detect “verbaltransformations” with imagined stimuli. Ordinarily, verbal transforma-tions are perceived when an actual (not imagined) auditory stimulus isrepeated over and over and over. Thus, for example, the word “stress,”repeated several times aloud, will soon be perceived as (repetitions of)“dress,” or, for some subjects, “rest.” The issue is whether these sametransformations will be detected when the repetitions are imaginedrather than perceived.

The data are shown in table 10.1. In the top row, it is clear that con-current articulation (“ta-ta-ta aloud”) is disruptive of memory-spanperformance, as one would expect, dropping performance from 65%correct to 57%. But silent clamping has no effect, and yields perfor-mance essentially equivalent to that in the control condition. The patternis different, however, in the bottom row. For the verbal transformations,both interference manipulations yield performance reliably worse thanthat observed in the control condition (also see Reisberg et al. 1989;Smith et al. 1994). It also seems that, in this case, concurrent articulationis more disruptive than clamping—perhaps because the former isnoise-producing, perhaps because it is more distracting (a dynamic ac-tivity, rather than the static state required for silent clamping). One wayor the other, the key point remains: Silent clamping disrupts verbaltransformations, and seems not to disrupt memory rehearsal.

We do urge some caution in interpreting these data, since other studiesin this series have yielded slightly different patterns, and the reasons forthis are not yet clear. (For other data on this issue, and other complica-tions, see Bishop and Reisberg 1996; Gupta and MacWhinney 1993;Smith, Wilson, and Reisberg 1996.) Nonetheless, these data do seem tosuggest that memory rehearsal relies on the top pathway in figure 10.3,

The Detachment Gain 149

and so is disrupted by the noise produced by concurrent articulation. Theimagined verbal transformations, in contrast, rely on the bottom path-way, and so are truly dependent on subvocalization, and disrupted if sub-vocalization is blocked—even if silently blocked, as in the clamping task.

More broadly, these results confirm the claim that there are indeed twodifferent ways to create an auditory image. Both are disrupted by thepresence of noise; only one is disrupted by (silently) occupying the artic-ulators. In other words, both pathways involve the auditory system, butonly one involves articulation. (For related data, see Bishop and Reisberg1996; Gupta and MacWhinney 1993; Smith, Wilson, and Reisberg 1996.)

This now forces us to confront a series of questions that have obvi-ously been waiting in the wings from the start: How do these two paths,each forming a mental image, differ? Why do some tasks require onepath rather than the other? What is it about the D-2-R task or the imag-ined verbal transformations that demands enacted imagery (and thus isdisrupted by silent clamping) even though another route toward creat-ing auditory representations is available?

Here is a hypothesis: When someone creates a mental image, theimage comes into being with certain understandings attached to it—un-derstandings similar to the ones inherent in the Gestalt principles ofperception. If the task requires visual imagery, this initial understandingwill include some specification of the figure-ground organization of theimaged scene, the orientation of the imaged form, its arrangement indepth, its segmentation, and the like. There is, of course, room for de-bate about how this organization is implemented—as part of the imageitself, or as a series of constraints on how the image is inspected—but,one way or the other, this organization is in place, and so subjects’ im-ages are not unorganized pixel patterns. Instead, subjects form imagesof organized perceptual objects, and thus their images reflect their per-ceptual understanding of that organization.

To make this concrete, consider the simple form shown in figure 10.4.This form can be perceived as a black square on a white background, or as

150 Daniel Reisberg

Table 10.1Summary of results from Bishop (1996)

Main task Interference type

None Clamp Concurrent (control) articulation

Memory span 65.3 65.7 57.2(percent correct)Verbal Transformations 1.97 1.60 1.17(words transformed)

an aperture in a white surface, through which a (more distant) black sur-face can be (partially) viewed. Similarly, the black form can be perceivedas a square, or as a rotated diamond. The picture shown in figure 10.4 isthus ambiguous. More precisely, the depiction itself is quite neutral as tointerpretation, and is compatible with each of the interpretations just de-scribed. The perception of this figure, however, is not ambiguous; the formis perceived in one of the fashions just described or another. The percep-tion, in other words, isn’t neutral, and somehow does, in one way or an-other, incorporate the perceiver’s understanding of the depicted form.

Many other examples can also serve to illustrate this point. The draw-ing of the Necker cube is neutral as to interpretation, and so open tomultiple interpretations. But the Necker cube is perceived in some de-terminate fashion, and, in fact, is perceived first in one way and then inanother. This shift, from one perceptual organization to another, under-scores the fact that the percept is indeed organized, and thus already in-terpreted in a way that pictures are not.

My claim is that images are, in this regard, just like percepts, and notlike pictures—perceptually organized depictions, and not mere pixelpatterns.4 So far I have illustrated this point with examples from vision,but the same ideas apply both to perception and to imagery in othermodalities. If a task requires auditory imagery, the subject will create animage with a certain perceptual understanding of the imaged form—anunderstanding that stipulates how the form is to be parsed, how theunits group together, and the like. In this case, we should not think aboutthe image as though it were some sort of echo or spectrogram. Instead,the image is an image of a perceptually organized auditory event.

This theoretical sketch leaves a great deal unsaid, but even in thisskeletal form this view has clear implications for how imagery func-tions, and for how subjects can use their imagery to make new discov-eries or to reach new insights. In essence, the proposal is this: An image,

The Detachment Gain 151

Figure 10.4.A simple ambiguous figure. Is this a black square on a white background, or a square-shaped aperture in a white surface, through which one can see a portion of a black field?

I have suggested, is more than a mere depiction; instead, the image is aninterpreted, organized depiction. Processes leading to image-based dis-covery, therefore, must take as their starting point this organized depic-tion, and this puts important constraints on image-based discoveries.

To see how this matters, consider two contrasting cases. In one, sub-jects are seeking to make an image-based judgment or an image-baseddiscovery that is entirely compatible with their initial organization ofthe target form. In this case, image-based discoveries flow quite readily.If the image is visual, subjects gain little advantage from “reproducing”the image in an actual (out-in-the-world) picture. If the image is audi-tory, subjects gain little advantage either from overtly voicing the im-aged sound, or from subvocalization.

In other cases, though, subjects seek to create an image, but then mustreanalyze it. Perhaps the discovery depends on a change in one’s as-sumptions about how the form is segmented, or perhaps the discoveryrequires that subjects blur together two or more of the imaged elements.These are cases in which the subjects’ task requires them to change theirunderstanding of the form, such that their goal is incompatible with theimaged object as initially organized. Image-based discovery of this sortturns out to be surprisingly difficult and heavily dependent on hints(typically: hints suggesting exactly how the imaged form might be reor-ganized). This sort of discovery is also much easier with an actual stim-ulus (auditory or visual) than with a mental image in the correspondingmodality. And this sort of discovery also seems, in the auditory case, torequire subvocalized support. The D-2-R task provides an obvious ex-ample in this category. An image is initially formed with three temporalsegments (“dee,” “two,” “are”) and these must be collapsed into twosegments, plainly demanding a shift in how the form was initially understood. This task, therefore, requires subvocalized support. Theverbal-transformation effect provides another example, in this case de-manding a shift in how the auditory stream is parsed. Other examplesshow the same pattern with nonlinguistic stimuli (e.g., Smith et al.,Experiment 3). In all cases, if a change in parsing is required, or if amemory chunk must be decomposed into its elements, then subvocal-ization seems to be needed.

But why does subvocalization help with these tasks? One plausiblesuggestion is that, with these tasks, you need to set aside your initialperception, and, in essence, reperceive. This creates the possibilitythat you will arrive, on the “second look,” at a different perceptualproduct than you did the first time around. To make this possible, youneed to jettison your initial view of the imagined event and go back tothe “raw material,” the initial data specifying that event. For this, youneed access to this raw material—you need, in short, a stimulus. And

152 Daniel Reisberg

what subvocalization provides is a stimulus—the raw material, serv-ing as grist for perception, existing independent of your understand-ing of it.

There is no need for this restructuring, no need for a “second look,” ifa task depends on a judgment that is compatible with your initial under-standing. For such tasks, you do not need to jettison one organization inorder to find another. Hence there is no need to return to the raw mater-ial, and, indeed, no need for the raw material. Therefore, there is noneed for subvocalized stimulus support, and no cost if subvocalizationis unavailable.

To put this in slightly different terms, the idea is that subvocalizationallows you to gain some distance, and, crucially, some detachment, fromyour own mental products, in order to inspect these products as astranger might. This allows a “new start” with the information, andmakes possible a fresh perspective that (potentially) can lead to freshdiscoveries.

An impressive range of data fits with these suggestions. As just oneexample, consider these tasks: Would these strings sound alike, if pro-nounced aloud: HEJ and HEDGE? RANE and REIGN? These judg-ments can be made about an intact unaltered representation, and, as itturns out, these judgments can be done with subvocalization blocked,in a noisy environment, or with both subvocalization blocked and anoisy environment (Smith et al. 1996). In contrast, would these stringsrhyme if pronounced aloud: TAPE and FAIP? GAUZE and PAWS? Thesejudgments do seem to rely on subvocalization, and are compromised ifsubjects are blocked from covert speech.

The rhyme and homophone tasks are obviously quite similar—bothrequire a judgment about how two strings are related; both hinge on acomparison between mental representations of sound (or, at least, rep-resentations of phonology). With appropriate stimulus selection, thetwo tasks can also be well matched in terms of overall difficulty. Despitethese commonalities, however, a crucial difference remains: The homo-phone judgment rests on an assessment of intact representations—cre-ated and then judged as whole units. The rhyme judgment, in contrast,rests on an assessment of representations that must be constructed andthen dissected, in order to ignore all-but-the-final sound of the targetstring. It is this dissection that requires stimulus support—requires astimulus that subjects can “step away from,” in order to change theirthinking about the form’s temporal boundaries. This stimulus supportcould be provided by an actual overt stimulus—if, for example, the sub-ject chose to read these visual strings aloud. Or it can be providedcovertly, via subvocalization. It is this latter path that subjects generallyuse, and it is this latter path that is blocked by concurrent articulation.

The Detachment Gain 153

(For other data relevant to these themes, see Reisberg 1996; Reisbergand Logie 1993; Reisberg et al. 1989; Smith et al. 1996.)

Let me close this chapter, though, by putting these ideas into a largercontext, in order to emphasize the breadth of issues that may be at stakehere. First, I have on occasion described this work to psychotherapists,and they mention the fact that it’s sometimes useful in therapy to tape-record the client, and then allow the client to listen to himself or herself.Equivalently, the client is sometimes asked to write down an account ofthis or that issue and then the client and the therapist together read thisaccount. The idea is that, in these cases, there is a lot to be gained by rip-ping ideas out of their initial mental context, undoing, in essence, encoding specificity, and then examining these ideas with some detach-ment. This seems to me an interesting idea, and it says something, I be-lieve, about the processes of discovery, and also about the way ourthoughts take some definition from a broader mental context. But, inany event, we can put this example alongside of our auditory imagerywork, with the suggestion that we may be seeing parallel phenomenahere—similar effects in highly different settings.

A second example comes not from the clinic but from the classroom.All instructors know that teaching material, be it psych 1 or an ad-vanced course, is a terrific way to learn that material. Why is this? Partof the answer, of course, lies in the thought one has to give to the mater-ial in order to figure out the best sequence of presentation, the best levelof specificity, the right explanatory metaphors, and so on. But there isalso, I believe, a gain from the teaching activity itself, from getting theideas outside of you, into the public arena, so that, again, you can hearthose ideas as a stranger might. This, too, seems a parallel phenomenonto the one at stake in our research. It is also one of the reasons, by theway, why I believe it is particularly important to train graduate stu-dents to teach—as a way of making them better teachers, of course, butalso as a way of making them better thinkers. And, of course, this is anapproach to graduate education strongly associated with the Universityof Pennsylvania’s psychology department, and, in particular, associ-ated with both Henry and Lila Gleitman.

Third, and finally, we still need to ask: So why does Henry think outloud? I’m willing to entertain the idea that this, too, is the same phe-nomenon. Thinking out loud is, in some ways, just like writing a paper.In both cases, you’re forced to fill in the gaps in your thinking, so thatthe argument is fluent. In both cases, you’re forced to articulate ideasthat had, until that point, been vague intuitions. In both cases, the taskof remembering all of one’s points is eased, and this frees up capacityfor other chores, including, of course, the evaluation of these points.

154 Daniel Reisberg

But, beyond these fairly obvious suggestions, I also believe there’ssomething I’ll call a “detachment gain,” a gain from being able to rip theideas out of their initial context. It’s interesting, for example, that yougain something by writing your ideas down, and you gain even moreby then setting the written document aside and returning to it a fewdays later. Why is this? If the gains come merely from filling-in of gaps,or from a reduction in memory load, then the full benefit should be re-alized by writing the ideas down in the first place. Why, therefore, dowe observe a further gain from “increasing the distance” between theauthor and the product? My suggestion is that this sort of “detachmentgain” is similar in its origins to the effects I have described in our im-agery experiments. In both cases, this “externalization” allows one toremove the idea from the context of understanding in which it was cre-ated. And that removal leads to the possibility of new discoveries thatmight not have been obtained in any other fashion. This is a substantialgain. It is a gain easily demonstrable in the laboratory. It is a gain thatmotivates many of us to write. It even motivates some of us to writetextbooks. It’s part of what motivates many of us to teach. And it is part,I think, of why Watson was correct in suggesting that some of us, withlaryngitis, are impressively reduced in our mental capacities.

Notes

1. Of course, the aspects of covert speech relevant for this task are likely to be differentfrom the aspects needed for the D-2-R task. To judge pitch, subjects need informationabout their own vocal chords (or at least the planning mechanisms for the vocalchords). For the D-2-R task, they need access to information about the articulators.Thus, although both of these tasks depend on subvocalization, they may rely on differ-ent parts of subvocalization. For discussion of this point, see Smith, Wilson, andReisberg 1996, especially their discussion of Experiment 1.

2. To compute the reliability of this correlation, one first needs to determine the samplingdistribution for this statistic; this can be done via a Monte Carlo technique, asking whatthe correlation would be if there was no linkage between these two data sets, that is, ifthere was no tendency for positions to be “preserved” as one moved from one of theseplots to the other. This can be calculated by considering the correlations obtained whenthese two sets of distances—in the imagery data and in the perception data—are ran-domly paired. This technique indicates that the expected value for these correlations,under the null hypothesis, is r = –0.005, with a standard deviation of 0.093. For furtherdiscussion, see Besag and Diggle (1977).

3. To be more precise, these activities block usage of the planning mechanisms needed to co-ordinate and control subvocalization. Thus the locus of interference is not the articula-tors themselves, but the more central systems that control the articulators. Fordiscussion of this point, see, for example, Baddeley and Wilson 1985, or Smith et al.1994. For relevant data, also see the chapter by Jonides, this volume.

4. For related arguments, see Pylyshyn (1981). I might add, however, that my views of im-agery have been misconstrued by a number of colleagues, and so a note of clarificationseems worthwhile. The claim is not that images are “mere descriptions.” Instead, itseems undeniable that images do depict the represented form—that is, images show

The Detachment Gain 155

(don’t describe) what the form looks like. The key, however, is that images depict in aspecial way: Unlike pictures, images are perceptually organized. Thus the historical de-bate, asking whether mental images are more like “pictures” or more like “proposi-tions,” simply gave us the wrong alternatives. Images (like pictures) depict, but also(like propositions) are structured in a fashion that renders interpretation unambiguous.

References

Baddeley, A. D. and Andrade, J. (1995) The impact of concurrent articulation on auditoryimagery. Unpublished manuscript; Applied Psychology Unit; Medical ResearchCouncil; Cambridge, England.

Baddeley, A. and Logie, R. (1992) Auditory imagery and working memory. In D. Reisberg(ed.), Auditory Imagery. Hillsdale, NJ: Erlbaum Associates.

Baddeley, A. and Wilson, B. (1985) Phonological coding and short-term memory in pa-tients without speech. Journal of Memory and Language 24:490–502.

Besag, J. and Diggle, P. J. (1977) Simple Monte Carlo tests for spatial pattern. AppliedStatistics 26:327–333.

Bishop, D. and Reisberg, D. (1996) The role of articulation in generating, maintaining, andmanipulating speech-based representations. Unpublished manuscript; AppliedPsychology Unit; Medical Research Council; Cambridge, England.

Crowder, R. and Pitt, M. (1992) Research in memory/imagery for musical timbre. In D.Reisberg (ed.), Auditory Imagery. Hillsdale, NJ: Erlbaum Associates.

Crutcher, R. (1994) Telling what we know: The use of verbal report methodologies in psy-chological research. Psychological Science 5:241–244.

Ericsson, K. and Simon, H. (1980) Verbal reports as data. Psychological Review 87:215–251.Gupta, P. and MacWhinney, B. (1993) Is the phonological loop articulatory or auditory?

Journal of Memory and Language 33:1–26.Iverson, P. and Krumhansl, C. (1993) Isolating the dynamic aspects of musical timbre.

Journal of the Acoustical Society of America 94:2595–2603.Payne, J. (1994) Thinking aloud: Insights into information processing. Psychological Science

5:241–248.Pylyshyn, Z. (1981) The imagery debate: Analogue media versus tacit knowledge. In N.

Block (ed.), Imagery (pp. 151–206). Cambridge, MA: MIT Press.Reisberg, D. (1996) The non-ambiguity of mental images. In Cornoldi, C., Logie, R.,

Brandimonte, M., Kaufmann, G., and Reisberg, D. (eds.), Stretching the Imagination:Representation and Transformation in Mental Imagery (pp. 119–172). New York:Oxford University Press.

Reisberg, D. and Logie, R. (1993) The ins and outs of working memory. In M. Intons-Peterson, B. Roskos-Ewoldsen, R. Blake, and K. Clayton (eds.), Imagery, Creativity,and Discovery (pp. 39–86). Hillsdale, NJ: Erlbaum Associates.

Reisberg, D., Smith, J. D., Baxter, D. A., and Sonenshine, M. (1989) “Enacted” auditory im-ages are ambiguous; “Pure” auditory images are not. Quarterly Journal of Ex-perimental Psychology 41A:619–641.

Schooler, J., Ohlsson, S., and Brooks, K. (1993) Thoughts beyond words: When languageovershadows insight. Journal of Experimental Psychology 22:166–183.

Smith, J. D., Reisberg, D., and Wilson, M. (1992) The role of inner speech in auditory im-agery. In D. Reisberg (ed.), Auditory Imagery (pp. 95–119). Hillsdale, NJ: ErlbaumAssociates.

Smith, J. D., Wilson, M., and Reisberg, D. (1996) The role of subvocalization in auditoryimagery. Neuropsychologia 33:1433–1454.

Wilson, T. (1994) The proper protocol: Validity and completeness of verbal reports.Psychological Science 5:249–252.

156 Daniel Reisberg

Chapter 11

An Update on Gestalt Psychology

Philip J. Kellman

Long ago, in my first year of graduate school at Pennsylvania, I heardrumors about a mythical gathering of zealous researchers, whoseweekly deliberations extended far into the night and disbanded onlywhen exhaustion finally overcame insight. That year, the Gleitman re-search seminar was only myth, as Henry was ill. His recovery led to theseminar’s revival, and both were sources of joy in the department.

When I joined the seminar, I was just beginning to learn about per-ception. Henry was and is a great interpreter of all things psychological,but in his heart, I believe, perception holds a special place. No doubtsome of his interests in perception were traceable to his time on the fac-ulty at Swarthmore College. He was there during the height of the Ges-talt influence, interacting with Köhler, Wallach, and others. As ElizabethSpelke and I began research on the developmental origins of principlesfrom Gestalt psychology, Henry richly conveyed much of that tradition.He was always encouraging about our chances of answering some veryold questions about perceptual organization; his support meant muchto our fledgling project. Meanwhile, he kept testing my emerging eco-logical theoretical leanings with his own empiricist ones. In the seminar,Henry’s insights and those of others improved many a research project.Lila, in particular, had me baffled. I wondered if there were another L.Gleitman who was famous in psycholinguistics, as her comments aboutperception research showed such depth and wisdom that that theycould only have come from a specialist in perception.

Although we used them to guide our initial studies of the develop-ment of perceptual organization (e.g., Kellman and Spelke 1983), theGestalt principles of object segregation, which had been applied to oc-clusion situations by Michotte, Thines, and Crabbe (1964), were vagueand a bit numerous. In time, my own research has come back to theseprinciples, trying to develop from them more precise ideas that couldform part of computational and neural models of perception. In thisgratifying and challenging enterprise, I have worked closely with TimShipley, whose dissertation Henry and I co-supervised in 1988. Much of

what we have accomplished and much of what remains to be done canbe characterized as an update of Gestalt psychology. Transforming theGestalt insights into a detailed understanding of perceptual computa-tions is important for diverse reasons: It advances our understanding ofadult human visual perception, sheds light on perceptual development,and informs attempts to make artificial vision systems that could pro-duce descriptions of physical scenes from information in reflected light.In this chapter, I will try to make clear what has become of variousGestalt principles in some current research.

The Gestalt principles have been applied in many domains, but theiroriginal and most familiar home is the domain of object perception, inparticular the problems of visual segmentation and grouping. This isthe domain I will consider in discussing the computational legacy ofGestalt ideas. The basic problems in segmentation and grouping areeasy to describe and illustrate. In a sheaf of light rays arriving at the eye,no ray of light is physically connected to any others. Some image de-scriptions likewise preserve information separately for each physicallocation. A digitally encoded image might list for each location (pixel)intensity and chromatic values. There is no linkage between pixels 384and 385, for example.

What we get perceptually from the light rays coming from a real scene

158 Philip J. Kellman

Figure 11.1.Examples of boundary and surface interpolation. a) Partially occluded object. b) Illusoryobject. c) Apparent transparency.

or a digitized image is quite different. It is a description of objects andsurfaces in a three-dimensional (3-D) space. The objects are seen as de-tached or separable from adjacent objects and surfaces, and each objectunites many visual directions or pixels. The problems of segmentationand grouping involve a mapping from the optic array onto these repre-sentations of objects.

How do we group together and separate regions to achieve these,usually accurate, representations of the objects in a scene? Perhaps themost vexing part of the problem is that parts of an object often reflectlight to the eyes from several spatially-separated areas. In figure 11.1a,the black object appears as a single entity whose contours are partly oc-cluded in several places. How can a human viewer or a computer visionsystem connect separate visible regions and represent their hidden con-tours, surfaces, and overall shapes? This set of problems will be ourfocus.

The Identity Hypothesis in Object Completion

There is one idea it will be helpful to introduce and place in the back-ground: what we have called the identity hypothesis in object completion.There are a number of different-looking phenomena in which the visualsystem accomplishes segmentation and grouping by supplying hiddencontours and connecting regions. Some of the phenomena are shown infigure 11.1. Along with the partial occlusion display in figure 11.1a, fig-ure 11.1b shows what is usually called an illusory object or illusory fig-ure, and figure 11.1c shows an apparently transparent (translucent,really) object. The identity hypothesis states that these different-lookingperceptual completion phenomena are caused by the same underlyingprocess. In these displays, the same parts of the central figure are de-fined by luminance edges, and the gaps across which edges are interpo-lated are the same. Those are formal similarities, but what I amsuggesting is that the same gap-surmounting process is at work in all ofthese cases as well. The differences in our phenomenology for the vari-ous cases have to do not with differences in interpolation processes, butwith how the interpolated edges and surfaces are situated relative toother surfaces in the array (especially whether they are in front or be-hind).

The arguments and data suggesting a common interpolation processcan be found elsewhere (Kellman and Shipley 1991; Kellman, Yin, andShipley 1998; Ringach and Shapley 1996). Here I give one example toconvey the general idea. In figure 11.2, we see yet another perceptualsegmentation and completion phenomenon, called a self-splitting fig-ure or SSO. The particular SSO shown is one constructed by Petter

An Update on Gestalt Psychology 159

(1956) and later discussed by Kanizsa (1979). The display has several in-teresting properties. As noted above, it resolves into two distinct ob-jects—boundaries get constructed through homogeneously coloredregions of the display. A second interesting property, and our immedi-ate concern, is the depth relationship between the two objects. At thetop of the display, the righthand ring appears to pass in front of the left,whereas at the bottom, the lefthand ring passes in front of the right.These effects of perceptual organization appear to be strong and consis-tent across observers.

At the top, where the righthand ring crosses in front, its contours areclassic illusory contours and its surface is said to be modally completed(Michotte et al. 1964), meaning it has a sensory presence. In the same vi-sual direction, the lefthand ring has a partly occluded surface and con-tours, sometimes called amodal completion. Amodal means that thehidden surfaces are perceived or represented, but they do not have localsensory presence. (You could not answer a question about the presenceor absence of a smudge on the occluded surface because, after all, it isoccluded.) The phenomenological difference between illusory and oc-cluded contours and surfaces has led many to think that these are phe-nomena of very different character, the former explainable by sensorymechanisms and the latter involving cognitive processes. On the iden-tity hypothesis, these involve, at least in part, the very same interpola-tion mechanisms, and the phenomenological difference concernswhether, in the final percept, the interpolated surface forms in front ofor behind some other surface in the scene.

Here is where Petter’s effect comes into the story. In displays such asthe rings in figure 11.2, Petter observed that a simple rule governswhich object will be seen as in front, having illusory contours, andwhich will be seen as going behind. The object that must be completedacross the smaller gap always ends up in front, and the object that tra-verses the larger gap ends up behind. From this observation, which ap-pears to be correct, we can make the following logical argument. If thefinal “illusory” or “occluded” status of a contour depends on somecomparison with another interpolated contour, then some mechanismthat interpolates contours must operate before the final status as illu-sory or occluded is determined. For an explicit or implicit comparisonto take place, the visual system must recognize both contour comple-tions crossing at that site. In other words, the mechanism that interpo-lates contours is not “modal” or “amodal.” (For other phenomena anddata that converge on the same point, see Kellman, Yin, and Shipley1998.)

The idea of a common underlying mechanism producing phenomenawhose subjective appearance differs so greatly is somewhat surprising,

160 Philip J. Kellman

and there is residual controversy about what exactly is shared and whatmust differ in different-looking cases of visual completion. In what fol-lows, the identity hypothesis will not be our focus, but it will allow us tomove between experiments and data involving illusory and occludedobjects and boundaries without distinguishing these cases.

Gestalt Principles and Unit Formation

Segmentation and grouping, illusory contours and transparency phe-nomena all involve issues of unit formation, determining what goeswith what. The Gestalt psychologists first inquired into these problems,and Gestalt principles have been applied to all of these phenomena(Kanizsa 1982; Michotte et al. 1964). It is a nice consequence of the iden-tity hypothesis that our “updates” of certain Gestalt principles willapply to all of them as well. We now look at particular principles andexamine their legacies in more recent work.

Good Continuation

In his classic (1921) paper “Untersuchungen zur Lehre von der Gestalt”(“Laws of organization in perceptual forms”), Wertheimer gave a num-ber of examples illustrating what he called the “Factor of Direction” orthe “Factor of Good Curve.” Despite offering these two formal names,another of Wertheimer’s phrases used in passing—”good continua-tion”—has stuck as the name of this principle. Figure 11.3 shows someexamples redrawn from Wertheimer (1921).

Despite the compelling and varied nature of the demonstrations,Wertheimer’s definition of this principle is rather vague. In fact, the dis-plays are meant to convey the following idea, without any formal defi-nition:

An Update on Gestalt Psychology 161

Figure 11.2.Self-splitting Object (SSO) after Petter (1956). Although the display is homogeneous incolor, it is perceived as two bounded objects. The ring on the right tends to appear infront of the ring on the left at the top of the display but appears to pass behind the ring onthe left at the bottom. (See text.)

On the whole, the reader should find no difficulty in seeing whatis meant here. In designing a pattern, for example, one has a feel-ing how successive parts should follow one another; one knowswhat a “good” continuation is, how “inner coherence” is to beachieved, etc.; one recognizes a “good Gestalt” simply by its own“inner necessity.”

Despite its intuitive importance, it is hard to find in the seventy or soyears since Wertheimer any explicit definition of the “good” in goodcontinuation. One obvious candidate is to relate “good” to mathemat-ical descriptions of contours. A function is often called “smooth” inmathematics if it has no discontinuities in the first derivative. A discon-tinuity would correspond to a sharp corner, that is, a point at whichthere is no unique slope of the function. But there are other notions of

162 Philip J. Kellman

Figure 11.3.Examples of the “Factor of Direction” (Good Continuation) from Wertheimer (1923). a)The segments labeled A and D appear to form a unitary object, as do those in C and B. b)Despite the possible appearance of three closed regions, the display is usually seen ascontaining a unitary curved edge and a square-wave.

smoothness, involving higher derivatives. In the design of automobilebodies, for example, smooth might mean differentiable at least two orthree times (Prenter 1989). Which notion captures the phenomena ofhuman visual segmentation and grouping? Surprisingly, this issue hasbeen the subject of little empirical investigation. Some of the issues, andsome clues to the answers, are illustrated in figure 11.4. In (a), there isfirst-order continuity between parts A and B, but a second derivativediscontinuity between them. Between A and C there is a first-order ortangent discontinuity (TD). Perceptually, A and B appear unitarywhereas C appears separable. In (b), there is a smaller direction changebetween A and B than between A and C. Direction change might there-fore predict that A and B will be linked more than A and C. On the otherhand, both B and C have a TD with A. Perceptually, neither B nor C ap-pears to have continuity with A, suggesting the importance of TDs insegmentation. In (c), parts A and B are not distinguishable as separateparts; as parts of a constant curvature arc, they agree in all derivatives.The case is different in (d). Here, a straight segment (B) meets a constantcurvature segment (A). The two parts agree in the first derivative, butthere is a second-order discontinuity. Nevertheless, the two parts ap-pear smoothly joined. All of these examples suggest that TDs lead tosegmentation and their absence—agreement in the first derivative—fa-cilitates joining.

Apart from these considerations about continuous contours, we mayask what relationships between separated contours lead to their percep-tual connections, as in partially occluded and illusory objects? Hereagain, the notion of good continuation has been invoked (e.g., byMichotte et al. 1964), but without any specific definition.

Parts (e) and (f) of the figure illustrate the same contour relations as(c) and (d) with gaps now caused by occlusion. Both displays producethe appearance of a unitary contour passing behind the occluder.

Formalizing Good Continuation: Ecological Constraints and ComputationalTheory

Seventy years after Wertheimer, the intuition behind the principle ofgood continuation is still important. Making this idea useful in modelsof human and computer vision requires first of all a precise mathemati-cal specification. It also requires placing continuity in the context of thegeneral problem of scene segmentation and object perception. Insteadof starting with particular contours and patterns, we need to posebriefly the question of how objects reflecting light make available infor-mation that might be used to segment and group the world into discreteobjects and surfaces. These are questions of ecological optics Gibson

An Update on Gestalt Psychology 163

164 Philip J. Kellman

Figure 11.4.Examples illustrating the importance of first-order continuity and discontinuity. a)Segments A and B, which are first-order continuous, appear connected moreso than Aand C or B and C. b) A first-order or tangent discontinuity divides A, B and C. c)Apparently unbroken contour made from A and B segments, where all derivatives agreeat the point of connection. d) Apparently unbroken contour made from A and B seg-ments where the first derivatives of A and B agree at the point of connection, but there isa discontinuity in the second derivative. e) and f) Contours in c) and d) under partial oc-clusion. (See text.)

(1966, 1979) and computational analysis (Marr 1982). I consider thiscontext first and then return to good continuation.

Multiple Tasks in Object Perception

Object perception involves multiple computational tasks. The first isedge detection. Different physical objects, having different material com-position, will tend to produce reflected light of differing luminance andspectral composition. Accordingly, abrupt changes in luminance andspectral characteristics are likely to indicate locations of object edges.Not all such changes are boundaries of objects, however. Some areshadows; others are textural markings on continuous surfaces, and soon. Some classification process must distinguish surface edges fromthese other cases.

Much of edge classification may be achieved by coordinating informa-tion from a luminance map of a scene with a depth map, gotten fromstereoscopic information. For a moving observer, there will also beavailable a motion map, assigning to each location a velocity vector(see, e.g., Lee 1974). Discontinuities in the depth and motion maps willcorrespond to true surface edges with less ambiguity.

The most common type of edge emerging from these initial analysesis the occluding edge. It is a contour that bounds an object or surface onone side. Each occluding edge indicates where something ends in thescene, but each also marks a mystery. If a person is seen standing infront of a car, the image contour separating the visible surfaces of the carand the person is a boundary of the person but not the car. At this con-tour, the car’s surface disappears behind. The mystery is where it goes.

Determining which side bounds the object is called boundary assign-ment (Koffka 1935). Nakayama, Shimojo, and Silverman (1989) sug-gested that an image contour be labeled intrinsic to a surface region if itbounds that region and extrinsic if it does not. Boundary assignmentmay not be implemented as a separate process. When depth or motioninformation is available, it is computationally simple to recover edgesand the relative depth order of two surfaces at those locations. Becausedepth order determines boundary assignment (the nearer surface al-ways owns the boundary), boundary assignment and edge classifica-tion may occur together.

To this point, we have a representation of occluding edges, partiallybounding surface regions. Now we are in a position to consider howGestalt notions of continuity can be implemented in perceptual process-ing. The story has two parts. The first involves particular locations inimages in which edge continuity is disrupted—what I called above tan-gent discontinuities (TDs). A TD is nothing more than a sharp corner

An Update on Gestalt Psychology 165

where contours meet.1 At such a point there is no unique slope of thecontour. TDs thus characterize all the standard types of contour junc-tions—“T,” “X,” “Y,” arrow, and so on. To be a junction means to be aTD. In our updated notion of good continuation, a TD is the key con-cept. TDs in one of the displays from figure 11.1 are marked in figure11.5. Referring to figure 11.1, it can be seen that all of the interpolatededges, in all displays, begin and end at TDs.

The ecological importance of TDs is straightforward. It can be proven(see Kellman and Shipley 1991, appendix B) that every instance inwhich an object boundary is partly occluded produces a TD at the placewhere the boundary goes out of sight. Thus TDs are potential loci of oc-clusion. They also mark the transition points from extrinsic to intrinsiccontours of a surface region. For hidden parts of objects, TDs are wherewe pick up the trail of where their hidden edges might go.

Relatability

Some TDs are merely the visible corners of objects. Not all are loci of oc-clusion. Moreover, even when a TD is a locus of occlusion, there re-mains the question of where the occluded part of the boundary goes.More is needed to determine object boundaries. Here we come to thesecond part of the implementation of the Gestalt idea of good continua-tion, what we have called relatability (Kellman and Shipley 1991, 1992).

Relatability formalizes good continuation. It constrains unit forma-tion based on an assumption that object boundaries tend to be smooth.Specifically, relatability expresses the conditions required to connect twoedges by a smooth (at least once differentiable) and monotonic (singly

166 Philip J. Kellman

Figure 11.5.Occlusion display from figure 11.1 with all tangent discontinuities indicated by arrows.

inflected) curve that agrees with the tangents of the two edges beingconnected at the point where each leads into a TD. We will define re-latability with edges separated in the optical projection and show thatthe case of continuous edges (zero gap) is a limiting case. Relatabilitycan be defined using the construction shown in figure 11.6a.

E1 and E2 are edges of surfaces. Let R and r be perpendiculars tothese edges at the point where they lead into a TD. Let R be the longer ofthe two perpendiculars, and let the angle � be the angle of intersectionof R and r. Intuitively, when relatability holds, there will always be asmooth, monotonic curve that can be constructed, starting from theendpoint of E1 (and matching the slope of E1 at that point) and pro-ceeding through not more than a 90-degree bend to the endpoint of E2

An Update on Gestalt Psychology 167

Figure 11.6.Construction used to define relatability. a) E1 and E2 are surface edges; R and r are per-pendiculars to the tips (points of tangent discontinuity) of E1 and E2, assigned so that R >r. � is the angle between R and r, E1 and E2 are relatable if 0 ≤ R cos � ≤ r. b) Illustrationof relatable edges. c) Illustration of nonrelatable edges. Either a doubly inflected curve orintroduction of tangent discontinuities are required to connect two nonrelatable edges.

(and matching the slope of E2 at that point). When R cos � > r, any con-nection between E1 and E2 would have to be doubly inflected (if itmatched the slopes at E1 and E2) or would have to introduce sharp cor-ners where the interpolated edge meets E1 and E2. (See figure 11.6c.)According to this model, visual boundary interpolation does not occurin such cases.

Formally, E1 and E2 are relatable iff:0 ≤ R cos � ≤ rThis statement can be unpacked in two steps. The righthand side of

the inequality simply states that the projection of R onto r (R cos �) fallswithin the extent of r. Whenever the length of r is less than projection ofR onto r, the edges are not relatable. Second, the curve constructed toconnect the two edges cannot bend more than 90 degrees. This limita-tion is expressed by the lefthand side of the inequality, because cos �will be negative for � > 90.

Below we will see that relatability should involve all three spatial di-mensions, although we have defined it here in terms of two. A gooddeal of work, however, can be done with 2-D edge relations alone, be-cause the smoothness of objects in the 3-D world has consequences fortheir 2-D projections. It can be shown using elementary projectivegeometry that collinear edges, smooth curves, and sharp corners in 3-space always project onto collinear edges, smooth cuves, and sharpcorners in a 2-D projection (excluding degenerate cases, such as projec-tion of a line to a single point). Thus, much of the information about ob-ject smoothness and edge relations is preserved in the opticalprojections reaching the eyes, even in a static, 2-D image.

Experimental Evidence about Relatability

A variety of experimental evidence supports relatability as a formal de-scription of connections formed by the visual system under occlusionand in illusory contours (Kellman and Shipley 1991; Shipley andKellman 1992a). Some of the best comes from an elegant paradigm in-troduced by Field, Hayes, and Hess (1993). Field et al. used arrays oforiented Gabor patches, small oriented elements consisting of a sinu-soidal luminance pattern multiplied by a Gaussian window. A Gaborpatch closely approximates the best stimulus for the oriented filtersfound in simple cells of V1, the first visual cortical area. Displays usedby Field et al. contained randomly placed, spatially separated elementsvarying in orientation. Some displays contained a “path.” A path wasconstructed by having the a sequence of several nearby elements hav-ing the same angular relationship, for example, successive elementswere collinear, or successive elements differed by 15 degrees, etc. In the

168 Philip J. Kellman

experiments, subjects on each trial judged which of two successivelyand briefly presented arrays contained a path. When the positional andangular relations satisfied the relatability criterion, subjects performedvery well at this task. When the path consisted of a sequence of ele-ments rotated 90 degrees, so that relatability was violated, performancewas much poorer. It appears that certain edge relationships lead to edgeconnections which become salient, perhaps in parallel across large re-gions of the visual field. The study also supported the idea that edgeconnections decline as the angle varies from collinearity, with a cutoffaround 90 deg.

Strength of interpolation also depends on the relative extents of thephysically specified edges and gaps in a scene. Interpolation strengthappears to be a linear function of the “support ratio”: the ratio of physi-cally specified edge lengths to total edge length (physically given edgesplus gap length) over a wide range of display sizes (Shipley andKellman 1992b; Lesher and Mingolla 1993). This relationship makesprecise a version of the Gestalt law of proximity, that nearer elements aremore likely to be grouped together.

Relatability in Cases of Minimal Gaps

We have defined and illustrated relatability in the context of occlusionand illusory contours—cases in which the visual system constructs con-nections across spatial gaps. In the classic Gestalt examples, good con-tinuation was illustrated as determining the breakup of unoccludeddisplays, without appreciable gaps, into separate objects (as in figures11.3 and 11.4). Unoccluded displays may be considered as a limitingcase of relatability—the case where the gap is zero. (Actually, nearlyzero. The contours of the perceived figures do overlap, producingminute occlusions and illusory contours.) In such cases, the “connec-tion” of edges is the continuation of the edge that proceeds smoothlythrough a junction. We saw relevant examples in figure 11.4. These ex-amples fit the definition of relatability in that smoothness resides in thefirst derivative. Connecting a straight segment (zero curvature) with asegment of positive curvature yields a well-defined first derivative atthe point of connection but a discontinuity in the second derivative, yetfigure 11.4d appeared to have perceptual continuity. In contrast, thesharp corner in figure 11.4b disrupts continuity of segment A with bothB and C.

This analysis of relatability at the limit sheds light on typologies ofcontour junctions in human and artificial vision (Clowes 1971; Waltz1972). In a “T” junction, the contour that does not change direction indi-cates the boundary of a surface, whereas the other contour passes be-

An Update on Gestalt Psychology 169

neath. A “Y” junction is different in that no contour continues smoothly;all come to an end at that point in space. It has been suggested that the“Y” provides information for an object corner. Relatability subsumesthese observations about contour junctions under a more general prin-ciple for connecting and segmenting visual arrays.

3-D Relatability: Depth Information in Object Completion

For convenience, we defined the notion of relatability in a plane.Perception of object unity and boundaries in the 3-D world requires tak-ing into account 3-D relationships of contours, however. Over the years,several demonstrations of 3-D contour completion have been devised.One is shown below in figure 11.7. If this display is viewed stereoscopi-cally (free-fuse by crossing or diverging the eyes), it gives rise to a 3-D il-lusory contour on one side and a 3-D occluded region on the other.Binocular disparity places the inducing edges at particular 3-D orien-tations, and contour interpolation processes build the connections,smoothly curving through three dimensions, across the gaps. Thedemonstration suggests that interpolation processes take 3-D positionsand relations as their inputs and build connections across all three spa-tial dimensions.

Until recently, these phenomena have not been addressed experimen-tally. Recently, we carried out a series of experiments to test 3-D rela-tions in object completion. A full report will appear elsewhere (Kellman,Yin, Shipley, Machado, and Li, in preparation); here I note some of themain results.

We used 3-D illusory object stimuli such as those shown in figure 11.8.Such displays appear to produce vivid 3-D illusory contours and sur-

170 Philip J. Kellman

Figure 11.7.Example of 3-D illusory and occluded contours. (Free-fuse by crossing or diverging theeyes.)

faces. We hypothesized that these occur when the physically given con-tours satisfy a 3-D criterion of relatability. The extension from the 2-Dcase is this: Bounding contours are relatable in 3-D when they can bejoined by a smooth, monotonic curve. This turns out to be equivalent tothe requirement that, within some small tolerance, the edges lie in a com-mon plane (not necessarily a frontoparallel plane), and within thatplane, the 2-D relatability criterion applies. Another way of saying thesame thing is that the linear extensions of the two edges meet in their ex-tended regions in 3-D space (and form an angle greater than 90 degrees).

Three-dimensional relatability can be disrupted by shifting one piecein depth, as shown in figure 11.8b. Another relatable display and a cor-responding shifted, nonrelatable display are shown in figures 11.8c and11.8d.

The experimental paradigm used these displays as follows. Subjectswere shown a stereoscopic display on each trial. Stereoscopic dispari-ties were produced by outfitting the subject with liquid-crystal-diode

An Update on Gestalt Psychology 171

Figure 11.8.Stimuli in depth relatability experiments. Each display is a stereo pair. (Free-fuse bycrossing the eyes.) Below each stereo pair is a side view of the display with the relation tothe observer’s eye shown. a) 3-D relatable display. The top and bottom white areas lie inintersecting planes and appear connected by a 3-D illusory surface. b) Non-relatable dis-play made by depth-shifting one inducing surface in (a) relative to the other. c) 3-D relat-able display with top and bottom areas in a common plane. The top and bottom areasappear connected by a planar illusory surface, slanted in depth. d) Non-relatable displaymade by depth-shifting one inducing surface in (c) relative to the other. (From Kellman,Yin Shipley, Machado, and Li, in preparation.)

(LCD) shutter glasses, synchronized with alternating computer images.Subjects made a speeded judgment on each trial about the positions ofthe upper and lower parts of the display. Displays like those in figure11.8a and 11.8b were said to be in intersecting or converging planes.Those in figure 11.8c and 11.8d were said to be in parallel planes (includ-ing coplanar). Note that the classification required from the subject oneach trial was orthogonal to the display’s status as relatable or nonrelat-able. The key predictions were that (1) perception of a unified objectwould facilitate classification performance, and (2) perceived unitywould depend on relatability. The former was expected based on resultsin 2-D displays showing that object completion produces an advantagein detecting boundary orientation (Shapley and Ringach 1996; Kellman,Yin, and Shipley 1998).

Results of the initial experiment (Kellman, Yin, Shipley, Machado,and Li, in preparation) are shown in figure 11.9, which shows discrimi-nation sensitivity (d’) in a signal detection analysis by condition. Twovalues of depth displacement (used to disrupt relatability) were used.These corresponded to a 5 cm and a 10 cm shift in depth of one of thepieces from the observer’s viewing distance (100 cm). Results indicate aclear superiority for the relatable displays. (Note that performance onparallel and converging displays are combined in the sensitivity analy-sis.) Response times reflected the same advantage: Both parallel andconverging relatable displays produced faster responding.

On the surface, these results suggest that object completion producesa performance advantage in this task and that 3-D relatability, to a firstapproximation, predicts unit formation in these displays. Even thesmaller value of depth shift disrupted performance markedly. As this isa new paradigm and new data, however, there are several alternativeexplanations to be considered. Some of these are still occupying us inthe lab, but we can relate a couple of important results here.

First, it is possible that performance in our task might not really re-quire object completion. Perhaps relatable displays were better pro-cessed because their pieces were more nearly at the same distance fromthe observer. Comparing two parts’ orientations might be easier whenthe parts are equidistant. Our design allowed us to check this hypothe-sis using a subset of the data. As figure 11.8d illustrates, a subset of par-allel displays used a shift away from the canonical (relatable) stimulusthat actually made the two parts more nearly equidistant. We comparedthese displays (which had either 0 or 5 cm depth differences) with relat-able parallel displays having parts that differed substantially in depth(10 cm for the largest slant condition). Results showed that relatability,not similarity in depth, produced superior accuracy and speed. Morerecently we have tested even more subtle alternatives to the idea that

172 Philip J. Kellman

our effects are due to object completion. Results support the object com-pletion hypothesis.

But are these truly three-dimensional effects? Introducing binoculardepth differences involves monocularly misaligning contours in eacheye. Perhaps these monocular effects, not true depth effects, cause theperformance decrement. It is known that misalignment of parallel ornearly parallel contours disrupts 2-D object completion (Shipley andKellman 1992a; Kellman, Yin, and Shipley 1998).

In designing the original study, we aimed to produce significantdepth shifts using misalignments that remained within the tolerancesfor 2-D completion. It has been estimated that contour completionbreaks down at about 15 minutes of misalignment of parallel edges(Shipley and Kellman 1992a). Our misalignments were on the order ofabout 10 minutes in the maximum depth shift condition. To check theeffect of monocular misalignment, we carried out a separate experi-ment. In our binocular, depth-shifted displays, each eye had the samemisalignment with opposite sign. In this experiment, we used the samedisplays, but gave misalignment of the same sign in both eyes. Thus theamount of monocular misalignment was exactly identical in every dis-play as in the original experiment. Because both members of each stereopair had misalignments of the same sign, shifted displays appeared to

An Update on Gestalt Psychology 173

Figure 11.9.Sensitivity as a function of slant in the depth completion experiment. Relatable displayswere more accurately and rapidly classified, suggesting that the upper and lower induc-ing areas were processed as a connected unit. (From Kellman, Yin, Shipley, Machado,and Li, in preparation.)

be at the same depths as relatable displays, but with some lateral mis-alignment. Results showed no reliable accuracy or speed differences be-tween shifted and relatable displays in this experiment. This outcome isconsistent with the idea that perceived depth relationships affected ob-ject completion in the first study. The effects are not explainable bymonocular misalignment.

This line of research is just beginning, but it suggests that our up-dated notion of good continuation—contour relatability—applies inthree spatial dimensions.

Good Form

The principle of good form (or more generally, Prägnanz) describes thetendency of perceptual processing to maximize simplicity and or regu-larity. Whether perceptual systems act in accordance with such a princi-ple remains controversial. The principle has been difficult to defineprecisely, in part because it seems to refer to perceptual ourcomes ratherthan stimulus relationships. Some attempts have been made to formal-ize the notion of overall figural simplicity (e.g., Buffart, Leeuwenberg,and Restle 1981).

It is difficult to separate good form from other factors. Common illus-trations almost invariably involve edge continuity besides good form.Figure 11.10 shows two illustrations of good form redrawn from a text-book on perception. Both can be explained in terms of edge relatability.In the display in (a), the edges leading into the TDs are relatable so thatthe physically specified plus interpolated edges produce two closedforms—the triangle and the rectangle. The second example involves acase of relatability across minimal gaps. At each contour intersection,edges entering and leaving with no TD in between are classified visu-ally as connected. In contrast, a TD between entering and leaving con-tours indicates a possible breakpoint. In the figure, the continuity ofedges gives the two closed forms shown. Kanizsa (1979) argued thatthat global symmetry is a questionable or weak determinant of objectcompletion, using demonstrations that pitted global factors againstlocal edge continuity. Two of these are redrawn in figure 11.11.

The debate about local vs. global determinants of segmentation andcompletion has persisted, however. Sekuler, Palmer, and Flynn (1994),for example, reported evidence from a priming paradigm suggestingthat global completion occurs in displays like the one shown in figure11.12a. (Global completion entails seeing a fourth articulated part be-hind the occluder, making the display radially symmetric.) Others havereported evidence for both global and local completions using priming(Sekuler 1994; van Lier, van der Helm, and Leeuwenberg 1995). Van

174 Philip J. Kellman

An Update on Gestalt Psychology 175

Figure 11.10.Putative examples of good form or Pragnanz. a) A triangle and a rectangle are seen. b) anellipsoid and a square are seen. Both outcomes are explainable by relatability with no ad-ditional principle of good form or Pragnanz. (Redrawn from Goldstein 1995).

Figure 11.11.Kanizsa’s Demonstrations pitting local continuity against global symmetry. a) (Redrawnfrom Kanizsa 1979.)

Lier et al. interepreted their results in terms of dual or multiple repre-sentations activated by partly occluded displays.

This suggestion is close to our own hypothesis: Various experimentaleffects reflect two distinct categories of processing. One is a bottom-up,relatively local process that produces representations of boundaries ac-cording to the relatability criterion. This process is perceptual in that itinvolves a modular process that takes stimulus relationships as inputsand produces boundaries and forms as outputs. The other process ismore top-down, global, and cognitive, coming into play when familiaror symmetric forms can be recognized. For lack of a more concise label,we call it recognition from partial information (RPI).

One factor pointing toward such a distinction involves the identitybetween partly occluded and illusory objects, which we have alreadydescribed. The identity hypothesis has received considerable support(Kellman, Yin, and Shipley 1998; Ringach and Shapley 1996; Shipleyand Kellman 1992a), and certain types of displays, such as the Petter ef-

176 Philip J. Kellman

Figure 11.12.Displays pitting local continuity and global symmetry. a) Occluded object for which localand global completion hypotheses make differing predictions. b) Illusory object versionof a. Although subjects are willing to report a global (symmetric) completion in the oc-cluded version, the symmetric completion is not seen in the illusory object display.

fect which we considered earlier, suggest that an identity at some pointin processing is logically required (Kellman, Yin, and Shipley 1998).

If true, the identity hypothesis sheds light on the global-local contro-versy, for this reason. Global completion phenomena are not observedin illusory object displays. Figure 11.12b shows the illusory object dis-play with physically defined edges equivalent to those in figure 11.12a.The reader may observe that there is no appearance of a fourth articu-lated part in the illusory figure display.

If the identity hypothesis is true, why should global completion occurin occluded but not illusory object displays? The answer may be that thedisplays are the same in terms of the perceptual processes of contourand surface interpolation but different in terms of RPI. An occluded sur-face is an interpolated surface that is not the nearest to the observer insome visual direction (i.e., there is something in front of it). An illusorysurface is nearest to the observer among all surfaces in a certain visualdirection. The crucial consequence of this difference is this: An observerviewing an occluded display is aware that part of the object is hiddenfrom view. This allows certain kinds of reasoning and responses that arenot sensible when no part of an object is occluded. In particular, despiteany local completion process, the observer can notice what parts are visible(unoccluded) and whether they are consistent with some familiar or symmetricobject.

Consider a concrete example. If the tail rotor of a helicopter is seenprotruding from behind a building, an observer may easily recognizeand report that such a helicopter is present, even though the particularcontours and surfaces of the hidden parts are not given perceptually. Astored representation of the helicopter may be activated and a beliefabout the presence of the helicopter may be formed. But RPI differsfrom perceptual processes that actually specify the positions of bound-aries and surfaces behind an occluder.

This separation of processes might explain conflicting reports aboutglobal and local processing. First, the only objective data supportingglobal outcomes come from priming studies. It is well known that prim-ing occurs at many levels, from the most basic representation of thestimulus to higher conceptual classifications involving the stimulus(e.g., Kawaguchi 1988). Unfortunately, there have been no attempts todistinguish these influences in the priming literature on occlusion.Studies reporting global completion have typically used large numbersof trials with a small set of familiar and/or symmetric figures, such ascircles and squares. Even if the subjects start out with little familiarity ordo not notice the possibility of symmetry under occlusion, repeated ex-posure may produce familiarity or symmetry responses.

An Update on Gestalt Psychology 177

The Dot Localization Paradigm

Priming may not be suitable for separating perceptual processes ofboundary and surface completion from more cognitive influences. Totest the possibility of different processes, we developed a new experi-mental paradigm. We focused on the idea that perceptual boundarycompletion processes lead to specific perceived boundary locationswhereas RPI will not in general do so, as in our occluded helicopter ex-ample. We measured the precision of boundary location by showing anoccluded display and briefly flashing a probe dot in front of the oc-cluder. Subjects were instructed to respond on each trial whether theprobe dot fell inside or outside the occluded object’s boundaries (i.e.,whether the projection of the occluded object to the eye would or wouldnot encompass the dot).

We used an adaptive staircase procedure. In this procedure, the stim-ulus value for each trial changes depending on the subject’s responses.Systematic changes allow a single point on the subject’s psychometricfunction to be estimated. For each display, we used both a “two-up, onedown” and a “one up, two down” staircase to estimate two points: the0.707 probability of seeing the dot as outside the boundary and 0.707probability of seeing the dot inside the boundary (= 0.293 probability ofoutside). We took the difference between these estimates as a measure ofthe precision of boundary perception, and the mean of these estimates asan estimate of the perceived location of the boundary. Staircases for sev-eral stimulus patterns were interleaved, that is, patterns appeared in arandom order, and screen position was varied randomly.

We realized that competing perceptual and recognition processesmight lead to different strategies across subjects. Therefore, we gavesubjects explicit strategy instructions. In the global instruction condition,we told subjects that they should see the display as symmetric; for thedisplay in figure 11.12a, for example, they were told that there was afourth protrusion behind the occluder identical to the three visible pro-trusions around the circle. In the local instruction condition, we told themthat we wanted them to see the display as containing a simple curveconnecting the two visible edges. In this manner, we sought to find sub-jects’ best abilities to localize boundaries under a global or local set.

A number of interesting findings have emerged (Kellman, Shipley,and Kim 1996). Localization of boundaries in displays where comple-tion is predicted by relatability is extremely precise. This is true forstraight (collinear) and curved completions. A very different outcomeoccurs in cases where completion is predicted to follow global symme-try. Here, the precision (difference between “out” and “in” thresholds)is an order of magnitude worse. It is about 15 mm in a display of about70 cm diameter (in visual angle, about 20 arcmin in a display 87 arcmin

178 Philip J. Kellman

in diameter). Moreover, the midpoint of the range is close to 1 cm awayfrom the theoretically predicted location of the boundary. This resulthas shown up consistently in a range of displays testing symmetry andrelated global notions of object completion. There are a number of is-sues still under investigation in this new paradigm. What is alreadyclear is that global influences do not lead to specification of preciseboundary position in the way local perceptual completion does. Theseoutcomes are consistent with the idea of separate perceptual comple-tion and more cognitive RPI processes.

Similarity

An interesting feature of edge relatability is that it does not seem to besensitive to similarity of surface quality (e.g., lightness, color, or tex-ture). Figure 11.13 gives two examples. In (a) the visible parts are seen asa unified object despite differences in their surface lightness and con-trast polarity from the occluding object. In (b) an illusory figure isformed from connections between pieces of very different luminances.Shipley and Kellman (1992a) found that magnitude estimations of ob-ject completion under occlusion in a large sample of randomly gener-ated figures showed no reliable differences whether the relatable pieceswere the same or different in luminance and chromatic color. TheGestalt principle of similarity thus seems to have little effect on relata-bility or the boundary interpolation process in general.

Does this mean that there is no role for similarity in object comple-tion? Kellman and Shipley (1991) proposed a surface-spreading process

An Update on Gestalt Psychology 179

Figure 11.13.Surface color insensitivity of boundary interpolation. a) A unitary partly occluded objectis seen despite differences in lightness of its visible regions. b) Illusory contours form be-tween surfaces of different lightnesses.

that complements boundary interpolation (cf. Yarbus 1967; Grossbergand Mingolla 1985). Surface quality spreads within physically specifiedand interpolated boundaries. In figure 11.14a the circle appears as a spoton a background. In figure 11.14b, the righthand circle still looks thesame way but the lefthand circle may appear as a hole in the occludingsurface. This effect appears to be dependent on similarity between thesurface lightness and texture of the circle and the partly occluded el-lipse. Because the circle has no TDs, it does not participate in the bound-ary interpolation process. What connects the circle with the surfacebehind the occluder appears to be a separate connecting process relatedto surface similarity. This surface process appears to be confined withinthe boundaries of the completed partly occluded figure in figure 11.14b.Figure 11.14c suggests, however, that surface spreading also occurswithin the extended tangents of the boundaries of a partly occludedarea (the half of the ellipse above the occluder), even when they are notrelatable to others.

In her dissertation, Carol Yin tested these two hypotheses—that sur-face quality spreads within relatable edges and also within extended

180 Philip J. Kellman

Figure 11.14.Examples illustrating the surface completion process. a) The circle appears as a spot infront of a background. b) The lefthand circle now appears as a hole, due to surface com-pletion, based on similarity of lightness and texture. c) Surface completion can occur evenwithout edge relatability. (See text.)

tangents of nonrelatable edges continuing behind occluding surfaces(Yin, Kellman, and Shipley 1997). In a series of experiments, subjectsmade a forced choice of whether a circular area appeared to be a hole ina surface or a spot on top of the surface in a number of displays varyingedge and surface similarity relations. In a variant of the method, sub-jects made forced-choice responses of which of two displays lookedmore like it contained a hole for all possible pairs of displays in a partic-ular experiment. These studies confirmed the hypotheses of surfacespreading within relatable edges and tangent extensions. Yin also stud-ied the surface completion process from an objective performance para-digm, pitting the effects of surface completion in making a circle looklike a hole or a spot against small amounts of stereoscopic disparity. Shefound that surface completion interactions reduced sensitivity tostereoscopic depth (Yin, Kellman, and Shipley in press).

Surface similarity and edge relatability seem to play complementaryroles in object perception. Interpolated edges establish connectionsunder occlusion, and surface qualities (lightness, color, and texture)spread within physically given and interpolated boundaries.

Common Fate

Wertheimer (1921) defined the “Factor of Common Fate” in this way.Suppose one sees a row of dots in which some are closer to others, lead-ing to grouping by proximity. Now suppose some dots are shifted up-ward while others remain at rest. The shift will seem more disruptive ifonly dots that were initially grouped together are moved. If the shift in-volves some dots from different groups, it appears to change the group-ing.

The principle of common fate received little emphasis in later Gestaltdiscussions of perceptual organization. In Koffka’s (1935) treatise, forexample, the principle is not even mentioned. In some ways, however,the nugget of insight in the principle of common fate connects to themost important modern developments in understanding perception.Owing in part to the development of ecological analyses of perception(Gibson 1966; Johansson 1968), we know that motion relationships pro-vide a wealth of information about object structure and spatial layout.

For perceiving unity under occlusion, there are two distinct types ofinformation (Kellman and Shipley 1991). One, a direct descendant ofWertheimer’s common fate, we have called the edge-insensitive process.Certain motion relationships lead two visible parts to be seen as con-nected. This connecting principle does not require any particular rela-tionships among the visible edges of the parts for unity to be seen.Computational and psychophysical research has revealed processes

An Update on Gestalt Psychology 181

that can determine whether particular 2-D motion patterns are consis-tent with a rigid 3-D structure, and if so, what structure it is. Wert-heimer’s notion of common fate includes at least the stimulusrelationships that allow recovery of rigid structure (Ullman 1979; Todd1981). They may also include many nonrigid motions, such as thejointed motions characteristic of a moving human body, and elastic mo-tions, characteristic of organisms or inanimate objects that stretch andcontract during movement (Johansson 1975).

Spatiotemporal Relatability of Edges

A complementary process—the edge-sensitive process—does involveedge relationships in information given over time by motion. If a sta-tionary observer looks through dense foliage, she may see meaninglessfragments of color from the scene behind. If the observer moves whilelooking, however, the objects and spatial layout behind the foliage maybe revealed. Sequential projection of parts seems to allow visual percep-tion of complete objects, although this ability has not been much stud-ied. There is evidence that sequential perception of inducing elementscan produce illusory contours and figures (Kellman and Cohen 1984;Bruno and Bertamini 1988). Perception under these circumstances re-quires not only integration of information over time, but interpolation,because some parts of the object never project to the eyes. The situationis one encountered often in ordinary perception.

What stimulus relationships in both space and time lead to percep-tion of complete objects? With the extra degree of freedom given by mo-tion, attempting to answer this question might seem daunting. It mightbe possible, however, to extend the criterion of spatial relatability to ac-count for completion in dynamic scenes. A simple hypothesis abouthow this might be done is illustrated in figure 11.15. In (a), a movingopaque panel containing two apertures moves in front of an object.Suppose one part of the figure becomes visible through an aperture attime t1 and another part becomes visible at time t2. If the position andedge orientation of the part seen at t1 is encoded in a buffer and persistsuntil the part at t2 appears, the standard relatability computation can beperformed on the currently visible part and the earlier encoded part.The situation in (b) adds a step. Here the object moves, revealing onepart through the bottom aperture at t1 and another through the topaperture at t2. Here the hypothesis is that when the part appears at t1,the visual system encodes not only its position and edge orientation buta velocity signal. This velocity signal could be used to update the spatialposition of the earlier visible part over time, either in a special-purposebuffer or by triggering a pursuit eye movement. When the second part

182 Philip J. Kellman

becomes visible, it is combined with the updated position of the firstpart in the standard spatial relatability computation.

The Dynamic Occlusion Paradigm

Evan Palmer, Tim Shipley, and I recently developed an experimentalparadigm to test these ideas (Palmer, Kellman, and Shipley 1997). Theparadigm works as follows. On each trial, an object passes behind anoccluder with several narrow slits, vertically separated so that someparts of the object never project to the eyes. This feature makes the taska completion or interpolation task as opposed to only an integrationtask (where visible parts are integrated over time). On each trial an ob-ject passes once back and forth behind the occluder. Subjects then make

An Update on Gestalt Psychology 183

Figure 11.15.Spatiotemporal relatability. a) A moving occluding panel with two windows passes infront of an object, projecting parts of the object to the eyes at different times. If a trace ofthe first visible part can be preserved until the second appears, spatial relatability can op-erate. b) A moving object’s parts are projected at two different times in two differentplaces. If velocity information is available, the position of the initially viewed part can beupdated (by an eye movement or in a visual buffer) so that it’s position relative to the sec-ond visible part can be extrapolated. Spatiotemporal relatability applies the spatial relatability computation to the currently visible and previously visible, positionally ex-trapolated parts. (From Palmer, Kellman, and Shipley, in preparation.)

a forced choice between two test displays, choosing which matched themoving target display. The design is illustrated in figure 11.16.

Two display conditions were used. Relatable displays (apart from theshift manipulation; see below) met the criterion of spatiotemporal re-latability. The upper test display in figure 11.16 is an example. The othertest display differs from the first by having one of the three fragmentsshifted by some amount. Five different amounts of shift (ranging from1.67 arcmin to 8.33 arcmin of visual angle) were used. The targetmatched the unshifted test display on half of the trials and the shifteddisplay on the other half.

We predicted that relatability would facilitate encoding of the visibleparts in the target display. If three parts moving behind slits weregrouped into a single, coherent object, this might lead to more economi-cal encoding and memory than for control displays (see below) in whichthree detached pieces were encoded. For simplicity, I will consider hereonly the cases in which either a test display or both the target and a testdisplay were relatable. In these cases, it was predicted that the greaterease of encoding a relatable display would lead to better performance.

Displays in a second condition were compared to the first. These non-relatable displays consisted of the identical three pieces as in the relat-able condition, but the top and bottom pieces were permuted. (Seefigure 11.16b.) With these nonrelatable displays, it was hypothesized

184 Philip J. Kellman

Figure 11.16.Design for studying dynamic object completion. A target array consisting of three visibleparts moves behind the occluder, visible only through narrow apertures. After each pre-sentation, the subject makes a forced choice between two displays. a) Relatable display.b) Nonrelatable display. (See text.) (From Palmer, Kellman, and Shipley, in preparation.)

that visual completion would not occur; each nonrelatable target mighthave to be encoded as three distinct pieces, which would lead to greaterencoding demands and lower sensitivity to the relative spatial positionsof the three parts.

These experiments are just beginning, but we can present some earlyresults. Figure 11.17 shows accuracy data (discrimination d’) from 16subjects for relatable and nonrelatable displays as a function of shift.Relatable displays were far more accurately discriminated than dis-plays made of the identical physical parts but placed in nonrelatable po-sitions. The results provide tentative support for generalizing thenotion of relatability from the spatial to the spatiotemporal domain.There are a whole range of issues raised but not yet addressed by the re-sults. For example, we did not control fixation, and it is unclear whethereye movements based on velocity signals from the moving fragmentsfacilitate spatiotemporal object completion. Likewise, we have not yetinvestigated effects of a number of other parameters. One of special im-portance is velocity. We suspect from other research (Shipley andKellman 1994) that spatiotemporal completion will occur within a re-stricted temporal window of integration, around 165 msec. So the

An Update on Gestalt Psychology 185

Figure 11.17.Results of dynamic object completion experiment. Sensitivity is shown as a function ofthe misalignment difference between the canonical display and the other test choice.Separate plots are given for relatable and nonrelatable displays. (From Palmer, Kellman,and Shipley, in preparation.)

results of our initial studies of dynamic occlusion raise more questionsthan they answer. They do provide some basis for connecting dynamicobject perception to previous work with static displays, by means of theextended notion of relatability.

Neural Models

The theoretical ideas about boundary interpolation and surface fillingthat I have sketched are largely formal or computational in nature. Thatis, they characterize stimulus relationships that underlie object comple-tion. They provide only hints about a precise process model or neuralrealization. I think it is worth concluding by mentioning some clues inthese areas that are central to some of our current thinking and work inprogress, as well as some work by others.

We defined relatability in edge interpolation as a simple mathemati-cal relationship between edge pairs. A number of considerations areleading us to consider interpolation effects as resultants of excitationfields that arise from individual edges. For example, there is some evi-dence that edges and the surface of a single region continue behind anoccluder even when they do not connect to any other region (Kanizsa1979; Nakayama and Shimojo 1992). We call this edge continuation todistinguish it from edge completion or interpolation. In this case, edgesseem to continue along linear extensions of edge tangents at the point ofocclusion. Surface spreading along such tangent extensions was foundin Yin’s research, described above.

One way to account for edge continuation and interpolation is to as-sume that each physically specified edge at its endpoint gives rise to afield of excitations at nearby locations. A vector field would identifywith each spatial location and at each orientation (perhaps in a 3-D net-work) a certain excitation. Excitation would decrease with distance andwould also depend on the orientation and positional relations as speci-fied in the geometry of relatability. An interpolated boundary in thisscheme arises when excitation fields from two separate physically spec-ified edges meet, with a winner-take-all inhibition scheme preventingmultiple completions. The temporal component of spatiotemporal re-latability could be realized by adding the dimension of time to the vec-tor field.

Our research group and others are working on the specifics of thiskind of model. For now it may be sufficient to note that this approach isconsistent with some other psychophysical work, including that ofField and colleagues, Polat and Sagi (1994), Das and Gilbert (1995), andothers. Both neurophysiological and psychophysical experiments sug-gest that cortical cells sensitive to orientation trigger the kinds of spatialinteractions that could implement relatability. There is, of course, more

186 Philip J. Kellman

work to do in pursuing these general ideas. A meaningful theory willbuild on previously proposed frameworks (Grossberg and Mingolla1985; Grossberg 1994; Heitger and von der Heydt 1993) but specificquantitative relationships faithful to psychophysical data must beadded. New dimensions must also be added. Our research suggeststhat successful models must incorporate relationships across all threespatial dimensions and relationships in information given over time. Asdaunting as the theoretical task appears, it may be made tractable byprecisely characterizing the grammar of object completion. In particu-lar, we are encouraged by the idea that a simple piece of geometry—thenotion of relatability—may provide a common thread knitting togetherpictorial, 3-D, and spatiotemporal object completion. This unifying ideamay provide a platform for precise process modeling and investiga-tions into the underlying neural mechanics.

Conclusion

Understanding perceptual organization—and segmentation and group-ing in particular—still poses deep mysteries to researchers in biologicaland artificial vision. Yet often, when progress is made, we can trace itsroots to insights made more than a generation ago by the Gestalt psy-chologists. It is amazing to realize that not only did the Gestaltists pro-vide some of the clues about how to solve these problems, but they werethe first to articulate clearly that these problems existed at all. At thesame time, it must be admitted that their principles lacked precisionand coherence. That these principles can still be recognized in more re-cent computational models, however, attests to the robustness of theoriginal insights. In this chapter, I have attempted to make explicit someof these connections between the old and the new.

A simple piece of geometry—the relatability criterion—appears tocapture much of the grammar of edge interactions that lead to objectcompletion. With rather simple extensions, relatability can be applied tocontour interactions in depth and to dynamic object completion. Under-lying this principle—and the Gestalt idea of good continuation—is theidea that object boundaries tend to be smooth. An alternative ecologicalinterpretation might be that objects are not all that smooth, but for mak-ing inferences about where objects go under occlusion, smoothness isthe best general assumption for a visual processor to use. Relatabilitymight be implemented by simple interactions of units responding tooriented edges. Evidence is beginning to suggest that such interactionsoccur surprisingly early in cortical visual processing.

Complementary to the boundary completion process is the spreadingof surface quality within boundaries. Here, the Gestalt principle of sim-ilarity lives on. Some other principles, such as an idea of Prägnanz or

An Update on Gestalt Psychology 187

global symmetry, may turn out not to be determinants of perceptualrepresentations per se, but may exert their effects more in memory andrecognition.

Of the original Gestalt principles, it is the notion of good continuationthat emerges as having the most important legacy in models of objectperception. This is the principle that also stands out when I reflect onthe impact of the Gleitmans and the Gleitman Research Seminar. Thesemany years later, Henry’s and Lila’s insight, dedication, and high stan-dards continue to help all of us in our academic endeavors. That weseek to emulate them in our own research and teaching is perhaps thebest principle of good continuation.

Acknowledgments

Portions of this research were supported by National ScienceFoundation grant SBR-9496112. I thank Thomas Shipley, Carol Yin,Sharon Guttman, and Evan Palmer for useful discussions, and JohnJonides and Dan Reisberg for helpful comments on an earlier draft ofthis chapter. Address reprint requests to Philip J. Kellman, Departmentof Psychology, UCLA, 405 Hilgard Avenue, Los Angeles, CA90095–1563 or by email to <Kellman@cognet.ucla.edu>.

Note

1. Even the language we use to describe the idea contains the idea implicitly. We say a TDis a point where “contours meet,” but the presence of the TD is what makes it sensibleto say “contours” (plural). Without the TD there is only a single contour.

References

Bruno, N. and Bertamini, M. (1990) Identifying contours from occlusion events. Perceptionand Psychophysics 48 (4):331–342.

Buffart, H., Leeuwenberg, E., and Restle, F. (1981) Coding theory of visual pattern com-pletion. Journal of Experimental Psychology: Human Perception and Performance7(2):241–274.

Clowes, M. B. (1971) On seeing things. Artificial Intelligence 2:79–112.Das, A. and Gilbert, C. D. (1995) Long-range horizontal connections and their role in cor-

tical reorganization revealed by optical recording of cat primary visual cortex.Nature 375(6534):780–784.

Field, D., Hayes, A., and Hess, R. F. (1993) Contour integration by the human visual sys-tem: Evidence for a local “association field.” Vision Research 33 (2):173–193.

Gibson, J. J. (1966) The Senses Considered as Perceptual Systems. Boston: Houghton-Mifflin.Gibson, J. J. (1979) The Ecological Approach to Visual Perception. Boston: Houghton-Mifflin.Grossberg, S. (1994) 3-D vision and figure-ground separation by visual cortex. Perception

and Psychophysics 55 (1):48–120.Grossberg, S. and Mingolla, E. (1985) Neural dynamics of form perception: Boundary

completion, illusory figures, and neon color spreading. Psychological Review92:173–211.

188 Philip J. Kellman

Heitger, F. and von der Heydt, R. (1993) A computational model of neural contour pro-cessing: Figure-ground segregation and illusory contours. Proceedings of the FourthInternational Conference on Computer Vision. Los Alamitos, CA: IEEE ComputerSociety Press, 32–40.

Johansson, G. (1970) On theories for visual space perception: A letter to Gibson.Scandinavian Journal of Psychology 11(2):67–74.

Johansson, G. (1975) Visual motion perception. Scientific American 232(6):76–88.Kanizsa, G. (1979). Organization in Vision. New York: Praeger.Kawaguchi, J. (1988) Priming effect as expectation. Japanese Psychological Review, Special

Issue: Problems of repetition in memory 31(3):290–304.Kellman, P. J. and Cohen, M. H. (1984) Kinetic subjective contours. Perception and

Psychophysics 35(3):237–244.Kellman, P. J., Machado, L., Shipley, T. F. and Li, C. C. (1996) 3-D determinants of object

completion. Investigative Ophthalmology and Visual Science 37(3):685.Kellman, P. J. and Shipley, T. (1991) A theory of visual interpolation in object perception.

Cognitive Psychology 23:141–221.Kellman, P. J. and Shipley, T. F. (1992) Visual interpolation in object perception. Current

Directions in Psychological Science 1(6):193–199.Kellman, P. J., Shipley, T. F., and Kim, J. (1996) Global and local effects in object comple-

tion: Evidence from a boundary localization paradigm. Paper presented at the32nd Annual Meeting of the Psychonomic Society, St. Louis, Mo., November 1996.

Kellman, P. J. and Spelke, E. S. (1983) Perception of partly occluded objects in infancy.Cognitive Psychology 15:483–524.

Kellman, P. J., Yin, C., and Shipley, T. F. (1998) A common mechanism for illusory and oc-cluded object completion. Journal of Experimental Psychology: Human Perception andPerformance 24(3):859–869.

Kellman, P.J., Yin, C., Shipley, T.F., Machado, L. and Li, C.C. The 3-D geometry of objectcompletion. Manuscript in preparation.

Koffka, K. (1935). Principles of Gestalt Psychology. New York: Harcourt.Lee, D. N. (1974) Visual information during locomotion. In R. B. MacLeod and H. L. Pick

(eds.), Perception: Essays in Honor of James J. Gibson. Ithaca, NY: Cornell UniversityPress.

Lesher, G. W. and Mingolla, E. (1993) The role of edges and line-ends in illusory contourformation. Vision Research 33(16):2253–2270.

Marr, D. (1982) Vision. San Francisco: Freeman.Michotte, A., Thines, G., and Crabbe, G. (1964) Les complements amodaux des structures

perceptives. Studia Psycologica. Louvain: Publications Universitaires de Louvain.Nakayama, K., and Shimojo, S. (1992) Experiencing and perceiving visual surfaces.

Science 257(5075):1357–1363.Nakayama, K., Shimojo, S., and Silverman, G. (1989) Stereoscopic depth: Its relation to

image segmentation, grouping and the recognition of occluded objects. Perception18(1):55–68.

Palmer, E., Kellman, P. J., and Shipley, T. F. (1997) Spatiotemporal relatability in dynamicobject completion. Investigative Ophthalmology and Visual Science 38(4):256.

Palmer, E., Kellman, P. J. and Shipley, T. F. Spatiotemporal relatability in dynamic objectcompletion. Manuscript in preparation.

Petter, G. (1956) Nuove ricerche sperimentali sulla totalizzazione percettiva. Rivista di psi-cologia 50:213–227.

Polat, U. and Sagi, D. (1994) The architecture of perceptual spatial interactions. VisionResearch 34(1):73–78.

Prenter, P. M. (1989) Splines and Variational Methods. New York: Wiley.

An Update on Gestalt Psychology 189

Ringach, D. L. and Shapley, R. (1996) Spatial and temporal properties of illusory contoursand amodal boundary completion. Vision Research 36:3037–3050.

Sekuler, A. B. (1994) Local and global minima in visual completion: Effects of symmetryand orientation. Perception 23(5):529–545.

Sekuler, A. B., Palmer, S. E., and Flynn, C. (1994) Local and global processes in visual com-pletion. Psychological Science 5(5):260–267.

Shipley, T. F. and Kellman, P. J. (1992a) Perception of partly occluded objects and illusoryfigures: Evidence for an identity hypothesis. Journal of Experimental Psychology:Human Perception and Performance 18(1):106–120.

Shipley, T. F. and Kellman, P. J. (1992b) Strength of visual interpolation depends on theratio of physically-specified to total edge length. Perception and Psychophysics52(1):97–106.

Shipley, T. F. and Kellman, P. J. (1994) Spatiotemporal boundary formation. Journal ofExperimental Psychology: General 123(1):3–20.

Todd, J. T. (1982) Visual information about rigid and nonrigid motion: A geometric analy-sis. Journal of Experimental Psychology: Human Perception and Performance 8(2):238–252.

Ullman, S. (1979) The Interpretation of Visual Motion. Cambridge, MA: MIT Press.van Lier, R. J., van der Helm, P. A., and Leeuwenberg, E. L. J. (1995) Competing global and

local completions in visual occlusion. Journal of Experimental Psychology: HumanPerception and Performance 21(3):571–583.

Waltz, D. L. (1972) Generating semantic descriptions from drawings of scenes with shad-ows (Tech. Rep. AI-TR–271). Cambridge, MA: MIT.

Wertheimer, M. (1921) Laws of organization in perceptual forms. In W. D. Ellis (ed.),Readings from Gestalt Psychology. New York: Harcourt Brace, 1938.

Yarbus, A. L. (1967) Eye Movements and Vision. New York: Plenum Press.Yin, C., Kellman, P. J., and Shipley, T. F. (1997) Surface completion complements bound-

ary interpolation. Perception, special issue on surface appearance, 26:1459–1479.Yin, C., Kellman, P.J., and Shipley, T.F. (in press). Surface integration influences depth

discrimination. Vision Research.

190 Philip J. Kellman

Chapter 12

Beyond Shipley, Smith, and Gleitman: YoungChildren’s Comprehension of Bound Morphemes

Katherine Hirsh-Pasek

In the fall of each year, as leaves turn bright against the New Englandlandscape, psycholinguists make their annual pilgrimage to the BostonLanguage Conference. One of the highlights of the trip to Boston is theGleitman dinner, a gathering of all those fortunate enough to be Lilaand Henry’s intellectual children, grandchildren, and great grandchil-dren. As you look around the dining room, you can’t help but be im-pressed by the large number of scientists who have been touched by theGleitman tradition, a tradition characterized by outstanding scholar-ship, first-rate teaching, and personal friendship.

There is no match for the scholarship that we witnessed during ourgraduate years. Lila always understood the big picture of language de-velopment, constantly reframing our narrow questions into ones thataddressed major issues in the field. I remember marveling at the way inwhich she made our first-year research projects seem so much more im-portant than we had imagined. (She magically molded my research onyoung children’s understanding of jokes into a key project on the rela-tionship between metalinguistic processing and reading.) Lila also had(and still has) the insight and common sense to know just where to lookto test her account of a developmental story. She has that rare ability tointegrate data from linguistic and psychological journals with examplesfrom the TV guide, Star Trek, and a neighborhood two-year-old. WhileLila helped us ask the questions, however, it was Henry who wouldsculpt those questions into psychologically interesting research. The re-sult was a constant stream of papers in child language, each of which fitinto a larger program of research, many of which became classics in thefield.

Their scholarship is unquestioned, yet their style of teaching and ad-vising stand out as the shining light of my graduate years. When mythirteen-year-old son recently asked Henry what he would describe ashis greatest accomplishment in psychology, he answered without hesi-tation, “My students.” No one who worked with Henry or Lila wouldbe surprised by that answer. The Thursday night cheese seminars at the

Gleitman home showed us how much they cared. Every week we metuntil all hours of the night, learning how to respect each person’s ideas,even when we disagreed. We learned that there were no simple an-swers, and that every result had alternative explanations. Beyond ourweekly meetings, Henry and Lila were always available, never too busyto read our drafts or to look at our preliminary analyses. They workedwith us side-by-side to ensure that our papers were of high quality.Then, they graciously offered us first authorships on our collaborativeefforts.

Perhaps the reason that Lila and Henry were such good mentors,however, is that they were not just academic advisors. They were alsogood friends. When you became a Gleitman student you entered intothe rich world of the Gleitman’s life—on the tennis court, at the theater,and at the finest local restaurants. I am counting on a long and contin-ued collaboration and friendship with both Lila and Henry. I have nodoubt that they will define the field of psycholinguistics as we moveinto the next millennium.

In this paper, I take the opportunity to demonstrate one way in whichtheir insights continue to shape my research. Using the now classicShipley, Smith, and Gleitman (1969) as a springboard, my collaboratorsRoberta Golinkoff, Melissa Schweisguth and I ask anew, “When arechildren first sensitive to grammatical morphemes in the input lan-guage?” and “Do they show this sensitivity sooner in comprehensionthan in production?” Almost thirty years after these questions were ad-dressed in the Shipley, Smith, and Gleitman paper, they remain centralto the study of grammatical development.

Grammatical Morphemes and Their Role in Language Acquisition

One of the key issues in language development concerns the youngchild’s ability to discover the building blocks of grammar: the nouns,verbs, and other parts of speech in the ambient language. Only withthese building blocks in hand (or in head) can children come to recog-nize syntactic patterns in the input and to construct the grammar oftheir native language. Indeed, every theory of grammar acknowledgesthat the discovery of grammar is one of the fundamental problems oflanguage acquisition. Throughout the years, a number of proposalshave been advanced for how children might go about finding cate-gories like nouns and verbs. Among them are syntactic distributionalproposals, phonological proposals, and semantic proposals. In the syn-tactic distributional view, for example, children can use fairly regulardistributional properties of the grammar to begin the process of assign-ing words to form classes (see Maratsos and Chalkley 1980). For exam-ple, nouns generally occur at the ends of sentences in child-directed

192 Kathy Hirsh-Pasek

speech (Aslin, Woodward, LaMendola, and Bever 1996). Nouns alsogenerally occur after grammatical morphemes like “the.” Through attention to these structural cues children might come to create a cate-gories of nounlike and verblike words based on distributional regulari-ties in the input.

Nouns and verbs also have different prosodic properties. Nouns, forexample, are more heavily stressed within sentences than are verbs.They also tend to have more syllables per word, longer durations, morevowels, and more phonemes overall than do verbs. Perchance these sta-tistical regularities assist the child in finding the relevant form classes(Kelly 1992, 1996).

Finally, accompanying these structural and prosodic distinctions aresemantic differences that can assist children in locating nouns andverbs. Nouns often, though not invariably, refer to persons, places, andthings, whereas verbs are more likely to refer to actions. These grosscorrelations have become the fodder for semantic bootstrapping theo-ries (Pinker 1984; Grimshaw 1981; Bowerman 1973; but see Gleitmanand Gillette 1995 for a contrasting view). Such proposals can get thelearner started in form class assignment, but semantic “bootstrapping”can only take her so far. A word’s meaning does not define its formclass. To use Pinker’s (1994) example, the word interest can be a noun in“Her interest in bowling,” a verb in “Bowling started to interest her,” andan adjective in “She seemed interested in bowling from the start.” All ofthese instantiations of interest share a similar meaning. Yet, they are notall classified into the same part of speech.

In sum, then, learners have a number of partially redundant cues tolinguistic form class through syntactic distribution, prosody, and se-mantics. Undoubtedly, they can capitalize on this redundancy by at-tending to the coalition of cues to solve the problem of finding thebuilding blocks of grammar (Hirsh-Pasek and Golinkoff 1996; Morgan,Shi, and Allopenna 1996). Among all of these available cues, however,one set of cues to nouns and verbs stands out as more reliable than therest; one set of cues is sufficient (though not necessary) for distinguish-ing between the major form classes: grammatical morphemes.

Grammatical morphemes are the closed-class words (such as “the”)and bound morphemes (such as /ing/) associated with particular formclasses. Although in English these elements are usually weakly stressedin the input, they are fairly reliable cues for form class assignment. Forexample, nouns (or noun phrases) follow the grammatical morpheme“the,” and the morphological ending /ing/ tends to signal a verb. Thus,even though many cues operate in tandem to allow children a way toassign words into basic grammatical units of nouns and verbs, gram-matical morphemes might well provide the easiest and most reliablecue of all those available.

Beyond Shipley, Smith, and Gleitman 193

The potential role of grammatical morphemes in syntactic develop-ment has not gone unnoticed (Maratsos and Chalkley 1980; Morgan1986; Morgan and Newport 1981; Morgan, Meier, and Newport 1987).As noted above, Maratsos and Chalkley thought that grammatical mor-phemes would be central to a distributional view of how children learngrammatical categories. Further, Morgan and his colleagues arguedthat grammatical morphemes were key to the “prosodic bracketing”that allowed adults to parse artificial grammars into linguistic constit-uents. These cues are certainly available in the input. Yet, while adultsmay notice and be able to capitalize on these cues, significant contro-versy exists as to whether young children could even attend to, let alonemine, these weakly stressed cues in the service of grammatical acquisi-tion. Children do not reliably produce grammatical morphemes untilthey are about twenty-four months of age (Brown 1973; deVilliers 1973;Valian 1986; P. Bloom 1990). Thus, many think that these cues could notbe used by young language learners to assist them in discovering gram-mar. Pinker (1984) wrote,

In general, it appears to be very common for unstressed closed-class morphemes not to be present in the earliest stages in the ac-quisition of many languages. Thus, as much as it would suit mypurposes to claim that Stage I children have latent control over themorphemes whose presence defines the categorization of certainconstituents, it does not seem to be tenable given available evi-dence. (p. 103)

Though grammatical morphemes would assist children in form classassignment, Pinker is suggesting that children might not be able to usethese cues until the grammar is at least partially acquired.

This argument is powerful one. Yet, there is a rub. Pinker’s assertionsare based on production data. Shipley, Smith, and Gleitman (1969),however, claim that children could potentially be sensitive to thesemarkers in the language even though they do not produce them. Thatis, children might well comprehend grammatical morphemes (andtherefore use them in form class assignments) before they can say them.On this account, the lack of grammatical morphemes in children’sspeech represents a production constraint rather than a portrait of tod-dlers’ linguistic competence. It was this insight that Shipley, Smith, andGleitman (1969) captured in their paper, “A study in the acquisition oflanguage: Free responses to commands.” The authors noted,

It seems clear, however, that the study of spontaneous speech doesnot provide a sufficient basis for understanding what the child“knows” about language at various stages of development. . . .

194 Kathy Hirsh-Pasek

[A] study of spontaneous speech, however objective and compre-hensive, forms a poor basis even for the study of adult language.(p. 103)

It was Shipley, Smith, and Gleitman (1969), then, who set the stage forthe study of language comprehension as a metric for emerging lan-guage development.

In Shipley, Smith, and Gleitman (1969) two questions were posed.First, did infants and toddlers understand more than they could say?Second, were holophrastic and telegraphic listeners—who did not useany grammatical morphemes—sensitive to grammatical morphemes inthe input that they heard? Subjects 18 to 33 months of age participatedin an “act out” task in which they responded to three simple types ofcommands. Appropriate commands had obligatory grammatical mor-phemes, as in “Throw the ball.” Omissions were telegraphic commandsthat omitted the obligatory morphemes, as in “Throw ball.” Finally,nonsense commands placed nonsense words in places in which theobligatory morphemes belonged, as in “Gor ronta ball.”

In answer to the first question, results differed depending on the lan-guage level of the children. Those in the holophrastic group carried outmore commands when the commands omitted obligatory morphemesthan when they included them. Children in the telegraphic group, incontrast, carried out fewer commands when they omitted grammaticalmorphemes than when they included them. As Shipley et al. wrote,“What is surprising is that just those utterance types they themselvesdid not use were more effective as commands” (p. 331). These findingssuggest something that most researchers did not consider in 1969—thatchildren may be sensitive to grammatical morphemes even when theyare not yet producing them in their own speech.

In response to the second question on grammatical morphemes,Shipley et al. (1969) made an even more remarkable discovery. Whenthey presented telegraphic speakers with requests in which nonsensewords replaced the grammatical morphemes, the response pattern wasdisrupted. This further confirms the finding that the telegraphicspeaker is not a telegraphic listener. These children were sensitive togrammatical morphemes in the input that they heard.

The Shipley, Smith, and Gleitman (1969) findings opened the door formore investigations that probed young children’s sensitivity to gram-matical morphemes and their use of these markers in the constructionof grammar. A number of studies followed that confirmed and ex-tended the findings of Shipley, Smith, and Gleitman (1969). By way ofexample, Katz, Baker, and MacNamara (1974) and Gelman and Taylor(1984) found that infants as young as 17 months of age were sensitive to

Beyond Shipley, Smith, and Gleitman 195

distinctions between “the” and “a.” In enactment tasks, these toddlerswere more likely to retrieve a particular block when requested to get“the” block than when requested to get “a” block.

An even more dramatic example comes from Shafer, Gerken,Shucard, and Shucard (1992) who used an evoked potential procedureto demonstrate that 10- and 11-month old children could attend to thephonological properties of grammatical morphemes. When normalfunction morphemes (such as “a” and “of”) were replaced with non-sense functors (such as “gu”), infants noticed the change and paid moreattention to the sentences containing nonsense functors. It appears as if“infants are sensitive to enough of the canonical phonological proper-ties of their language to begin to identify function morphemes as aphonological class” (Gerken 1996, p. 417).

Finally, Gerken and McIntosh (1993) offer a compelling demonstra-tion of toddler sensitivity to grammatical morphemes in comprehen-sion. Using a picture-pointing task with four choices, toddlers 21 to 28months of age were requested to (a) Find the dog for me (correct mor-phology); (b) Find * dog for me (morphology absent); (c) Find was dogfor me (ungrammatical morpheme); or (d) Find gub dog for me (non-sense morpheme). Consistently, children performed better in the gram-matical than the ungrammatical task. For toddlers with MLUs of under1.5 hearing the stimuli in female Motherese, the proportions correctwere 86% in the correct condition and 75% in the missing condition,with a dramatic drop to 56% and 39% in the ungrammatical and non-sense conditions, respectively. Thus, children who were not producingmorphemes in their own speech were nonetheless sensitive to this in-formation in comprehension. Most importantly, children with lowMLUs have obviously learned something about particular phonologi-cal forms within the input. If they had not yet noted the particulargrammatical morphemes, then all except the absent morpheme shouldhave been treated similarly. If they had just classified the input asprosodically or lexically familiar versus unfamiliar, the ungrammatical“was” condition should have been as good as the grammatical “the”condition. Thus, children are not only sensitive to grammatical mor-phemes in the input, but seem to know something about their appropri-ate locations in the sentence—a crucial fact reopening the possibilitythat they could use different morphological cues to classify differentconstituents into the correct form classes.

In short, the studies that followed Shipley, Smith, and Gleitman(1969) reaffirmed their interpretation that children are sensitive to mor-phological cues in the input. The studies also confirmed the role thatcomprehension can play in providing an important window on lan-guage development.

196 Kathy Hirsh-Pasek

Expanding This Literature: A Study of Bound Morpheme Comprehension

The findings about sensitivity to grammatical morphemes are encour-aging. Most of the studies performed to date, however, have either beenconducted with older toddlers or have used only free morphemes like“the” that signal noun phrases (e.g., Brown 1973; Taylor and Gelman1988). To make the comprehensive case that toddlers note morphologi-cal cues in the input, one must demonstrate that they can attend to thefull range of morphological cues. That is, one must demonstrate thatchildren are equally sensitive to bound morphemes like “ing” that markverb phrases, or “ly” that mark adverbial phrases. It can be argued thatbound morphemes should be even more difficult to notice because theyare not only weakly stressed, but are affixed to the ends of the wordsthat support them.

To address this gap in the literature, Roberta Golinkoff, MelissaSchweisguth, and I tested toddler sensitivity to the bound morpheme“ing” (Golinkoff, Hirsh-Pasek, and Schweisguth, in press). Borrowingdirectly from the Gerken and McIntosh study, we presented three typesof stimuli to the children: grammatical morphemes (“dancing”), ungrammatical morphemes (“dancely”), and nonsense morphemes(“dancelu”). The logic of the design is as follows: If children do not at-tend to bound morphemes, they might see the three stimuli describedabove as virtually identical—interpreting each word as the stressedstem dance. If, on the other hand, toddlers make any distinction amongthe three stimuli, then there is evidence that bound morphemes are de-tectable in the input. Again, paralleling Gerken and McIntosh, ourclaim is that if the toddlers have more correct responses in the gram-matical condition (“ing”) than in the other two conditions (“lu,” “ly”),there would be evidence that the children are distinguishing among thephonological forms and that they could potentially use the informationin categorizing grammatical constituents. The hypothesis driving thisresearch was the latter one. We predicted that the children would in-deed differentiate among the three conditions, performing best in thegrammatical condition (“ing”), less well in the ungrammatical (but fa-miliar) condition (“ly”), and not at all well in the nonsense condition(“lu”). In what follows, I present data on this issue, for it not only un-derscores what Shipley, Smith, and Gleitman (1969) suggested, butagain gives us reason to look for unfolding linguistic competencethrough measures of language comprehension.

The subjects for this experiment were 108 toddlers, distributedequally and randomly into the three conditions of “ing,” “ly,” and “lu,”balanced for gender, and ranging in age from 18 and 21 months. All of the children had been screened by phone to ensure that they

Beyond Shipley, Smith, and Gleitman 197

comprehended at least 6 of the 8 verbs being used as stimuli. At the timeof the visit, the children were also asked if they were producing “ing.”Very few (about 6 of the children) occasionally produced “ing.”

Children were tested individually in the intermodal preferentiallooking paradigm developed to assess language comprehension in tod-dlers (Golinkoff, Hirsh-Pasek, Cauley, and Gordon 1987; Hirsh-Pasekand Golinkoff 1996 a,b). In the intermodal preferential looking para-digm (IPLP), children are seated on their blindfolded parent’s lap mid-way between two television monitors spaced 2.5 feet apart. Figure 12.1provides a schematic drawing of the procedure and the set-up. On onescreen children might see a woman dancing. On the other screen, and atthe same speed of delivery, they would see the same woman waving tothe viewer. Through a hidden speaker located between the monitors,children heard the test stimulus, which in this case was either, “Where’sdancing? Do you see dancing?” or “Where’s dancely? Do you seedancely? Or “Where’s dancelu? Do you see dancelu?” The logic of thisexperiment, confirmed in numerous previous experiments (see Hirsh-Pasek and Golinkoff 1996a,b), is that children will look longer at thescreen that “matches” or “goes with” the linguistic stimuli than at thenonmatching screen. Thus visual fixation serves as the dependent mea-sure. All dependent data were collected by videotaping the children’sresponses so that they could be coded “off-line.” Agreement betweencoders for these experiments has been consistently high, averaging ataround 91% (see Hirsh-Pasek and Golinkoff 1996a).

Before I describe the design further, it is important to note some of theadvantages of this procedure over others for testing the early languagecomprehension of bound morphemes. The first is that, unlike the pic-ture pointing tasks, the IPLP allows the experimenter to deliver dynamicstimuli to children in a controlled fashion. The bound morphemes usedwith young children are often (though not always) attached to actionverbs. It was therefore important to be able to test for linguistic compre-hension of these forms within the context of dynamically presentedstimuli. The second advantage is that while action can be displayed, theprocedure does not require any action on the child’s part. Thus, chil-dren are not lured into impulsive responses that might test their com-pliance (or lack thereof) rather than their linguistic competence. A merelooking response indicates the child’s preference for one screen over theother. For both of these reasons, then, the IPLP seems like an ideal wayto examine children’s budding knowledge of these linguistic forms.

The layout for these experiments is presented in table 12.1. The chil-dren were exposed to 8 different continuous actions (e.g., dancing, wav-ing, pushing, turning) that appeared in four pairs. Each pair of verbs

198 Kathy Hirsh-Pasek

was represented two times for 6 seconds each. The trials were also sep-arated by intertrial intervals during which a light between the twoscreens came on to draw children’s attention to the center. Thus, eachtrial required the child to make a new choice of which screen to attendto. The video tapes were tightly synchronized so that both members of apair appeared in front of the child at the same time. The stimuli werealso balanced for salience so that one member of a pair was not more en-ticing than the other. Finally, presentation of the actions were counter-balanced such that half of the time “dancing” would appear on the leftscreen and half of the time it would appear on the right screen.

The linguistic stimuli determined which of the screens would be thematching versus the nonmatching screen and also differentiated thethree test groups. In a between-subject design, 24 of the children heardall of the pairs in the “ing” condition, 24 heard the stimuli in the “ly”condition and 24 heard the stimuli presented in the “lu”condition. Note

Beyond Shipley, Smith, and Gleitman 199

Figure 12.1.

that the same sentence frames accompanied the words with the excep-tion of the bound morpheme. Table 12.1 contains a sample of the designfor the “ing” condition.

Thus the total design for this preliminary experiment had within-subject variables of verb (four pairs) and match (matching vs. non-matching) and between-subject conditions of linguistic group(grammatical, ungrammatical, nonsense) and gender.

Results

Before reviewing the results, note that there were no stimulus salienceproblems in the simultaneous trials. That is, when the pairs of actionswere presented with a neutral linguistic stimulus, neither verb in apair was intrinsically more interesting than the other member of thatpair.

The first important result comes from the grammatical, “ing” condi-tion. A three-way ANOVA with between-variables of verb and genderand within-variables of match revealed a main effect of match. The chil-dren looked at the matching screen (x = 4.01 sec.) significantly morethan the nonmatching screen (x = 3.31 sec.). In the “ing” condition, boththe boys and the girls responded correctly across all of the verbs. There

200 Kathy Hirsh-Pasek

Table 12.1Sample Block of Trials for “ING” Condition

Left Screen Linguistic Stimuli Right Screen

Simultaneous Trials

Black Hey boys and girls! BlackWhat do you see on TV?

woman drinking from cup What’s going on on those TVs? woman blowingWhat are they doing? air at piece of paper

Black Hey! Look up here! Blackwoman drinking from cup What’s going on on those TVs? woman blowing

What are they doing? air at piece of paper

Test Trials

Black Which one is drinking? BlackCan you find drinking?

woman drinking from cup Where’s drinking? woman blowingDo you see drinking? air at piece of paper

Black Whoa! Find drinking. Blackwoman drinking from cup Look up here again? woman blowing

Which one is drinking? air at piece of paper

were no interactions with verb or with gender. This result is criticallyimportant because it suggests that children are responding to the stim-uli. We do not know from this result alone, however, whether they arejust listening to the verb stem, (e.g, dance), or whether they actually no-tice the bound morpheme “ing.”

The “ly” (ungrammatical) condition produced more interesting re-sults. Here, the ANOVA revealed a main effect of match and an interac-tion between verb and match. The children—both boys andgirls—failed to watch the matching screen (x = 3.07 sec.) more than thenonmatching screen (x = 4.21 sec.) in the first verb and then looked atthe matching screen (x = 3.83 sec.) significantly more than the non-matching screen (x = 3.02) for the last three verbs. One possibility is thatthe children recognized “ly” as familiar but were puzzled at first by itsplacement on a verb. This would suggest that children are sensitive tothe ungrammatical use of a familiar morpheme and that this usage iscapable of disrupting sentence comprehension as in the Gerken andMcIntosh (1993) “was” condition. Though confused at first, however,they later decide that perhaps the familiar ending could be an endingfor the verb. Note here that if the children were only responding to theverb stem (e.g., dance) then no verb by match interaction should be ex-pected, since the verb stems were the same in the “ing” and the “ly”conditions. Thus the pattern of results for the “ly” condition suggeststhat by 18 months of age children possess more sophistication aboutgrammatical morphemes than we imagined. They appear to be awarenot only of which morphemes are found in English but of the type ofwords on which the morphemes are typically to be found. These datasuggest that children may be segmenting a verb into a stem and a mor-pheme. In the end, however, they let input rule the day and decide thatthe ending can be placed on the verb.

Finally, the “lu” (nonsense) condition offers yet a third piece of evi-dence that children are attending to bound morphology. Here, compre-hension is completely disrupted and neither the match nor thenonmatch is watched to a greater degree throughout the four blocks oftrials. Mean visual fixation time across the four blocks of trials is identi-cal across the match and nonmatch conditions at 3.56 seconds. Childrenwere not sure which screen to watch in response to words like“dancelu” and “wavelu.” Again, the only difference in the three linguis-tic conditions is the difference in the bound morpheme. Thus the boundmorpheme “lu” abolished all preferences for the verb stems.

DiscussionThese findings are suggestive and parallel to those of Gerken and McIn-tosh (1993). Even in late infancy, children are sensitive to the grammatical

Beyond Shipley, Smith, and Gleitman 201

morphemes in the input. What we saw in this experiment is that merelychanging the weakly stressed, bound morpheme at the end of a sen-tence frame significantly influenced children’s sentence processing.With “ing” at the end of the main verbs, all children responded appro-priately. With the familiar but ungrammatical “ly” at the end of thesame verbs, responses were initially confused and then resolved on thematching screen. Finally, with the nonsensical ending “lu,” children’sresponses were totally disrupted such that the only consistent trendwas from the girls who preferred the nonmatching screen.

What is clear from this pattern of responses is that the children didnote grammatical morphemes even in the difficult case of bound mor-phemes in which the functor is not only weakly stressed, but is also at-tached to a verb that carries the primary semantic force. These childrenwere not simply relying on the verb stem to determine their choices. Ifthey had been, their responses should be equivalent across the three testconditions. What is less clear is exactly what the differential pattern ofresponses does indicate. Below, I consider three possible interpretationsof these results. I then conclude by echoing Shipley, Smith, and Gleit-man’s (1969) concern that if we are truly to understand the differencesbetween children’s spontaneous speech and their knowledge, we mustdevelop new techniques for the systematic observation of this knowl-edge. I will argue that in this study as in others, the intermodal prefer-ential looking paradigm is a tool for the systematic observation ofcomprehension and that the study of early comprehension might pro-vide a crucial way to explore linguistic competence.

Three interpretationsThree possible interpretations could be used to explain these results: thewhole word explanation, the particular morpheme explanation, and the fa-miliar morpheme explanation. Let me take each in turn.

The first possibility is that children do not analyze a word into a stemand an accompanying morpheme. For years, psycholinguists have ar-gued about whether words are stored as whole units or as base wordsplus morphemes (Taft and Forster 1975; Rubin, Becker, and Freeman1979). If words are stored as whole units, the lexicon would require sep-arate storage of each morphological variant of a word. Thus teach, teacher,teaching, and teaches would each be stored as a separate and indepen-dent word. If, on the other hand, a complex word is stored as a base orstem plus the morpheme, then the word teach would be stored in lexicalmemory as would a set of morphemes that could be affixed to the baseword. For example, teach would be stored as would -er, -ing, and -es.Rules would then be required for adjoining base words and bound mor-phemes. It has been suggested that the whole word option requires

202 Kathy Hirsh-Pasek

more memory storage, but that the stem plus morpheme solution to lex-ical storage allows for more productivity and increases the processingload. Most of the current evidence from adult lexical decision experi-ments supports the stem plus morpheme interpretation (Taft and For-ster 1975).

It is possible that children who are first learning words might favor asystem for storing whole words as units. Without enough words in theirrepertoire, they might not be able to recognize the patterns of endingsthat comprise bound morphemes. Indeed, one could hypothesize thatchildren might need to have a critical mass of words before such analy-sis into stem plus morpheme could take place. A similar argument hasbeen offered by Jones and Smith (1993), who suggest that the shape biasin word learning does not occur until children have enough words intheir lexicons to do an internal analysis. Under this interpretation, thechildren in our experiment might have learned the whole unanalyzedwords for “dancing,” “pushing,” and “waving” and thus would performbetter when these words were used as stimuli than when unfamiliarwords like “dancely” or “dancelu” served as stimuli. This interpreta-tion, however, is not entirely supported by the data. The children did dobetter in the “ing” condition than they did in the “lu” condition; yet,they gave mixed results in the “ly” condition. Had they been using awhole word strategy, the “ly” condition should have elicited the sameresponses as did the “lu” condition. Both are equally unfamiliar. Yetthat was not the case. Hence we can tentatively reject the whole word al-ternative in favor of one of the two base plus morpheme alternative ex-planations.

The particular morpheme hypothesis holds that the child has alreadylearned something about particular morphemes in the input and thusknows, to a certain extent, that “ing” signals verbs and that “ly” signalsadverbs. Of the three explanations, this alternative gives the child themost sophisticated knowledge, suggesting that children use boundmorphemes to label constituent phrases. Again, the fact that the chil-dren watched the correct verb in the “ing” condition and watched bothverbs in the “lu” condition supports this alternative. Again, however,the findings from the “ly” condition make this hypothesis less likely.One could argue that the children in the “ly” condition were faced witha forced-choice alternative and that no adverbial alternative was avail-able. Choosing the lesser of two evils, they favored the “ing” alterna-tive—thus explaining the results in total. Yet there is another reason toquestion this explanation. Children who are just beginning to learngrammar must be open to the full range of bound morphemes that theywill encounter. If they fully restricted the class of morphemes to thosethat they currently knew, they would not be able to master new bound

Beyond Shipley, Smith, and Gleitman 203

morphemes. Thus, instead of supporting this alternative, we turn to aslightly more flexible explanation offered by a third position.

The final hypothesis, the familiar morpheme hypothesis, explains theresults and also leaves room for further learning. On this scenario, chil-dren know that certain phonological forms heard in the input serve asbound morphemes. That is, the children store acoustic information thathas no meaning for them as yet, but that has been repeated with somestatistical frequency. Several recent experiments attest to the fact that in-fants as young as 8 months of age can perform this kind of statisticalacoustic analysis (see, e.g., Saffran, Aslin, and Newport 1996). Oncestored as acoustic templates, some of these sound sequences could thenbecome associated with specific form classes and come to have moreparticular meanings. Familiar phonological patterns like “ing” may beso frequently encountered that they become associated with particularstems that the children have heard before. The morpheme “ly, “ how-ever, may sit longer in this undifferentiated phonological class untilenough information becomes available to classify it reliably (see Gerken1996 for a similar proposal). On this scenario, the highly familiar “ing”and the familiar “ly” would pattern in somewhat the same way, whilethe unknown “lu” ending would pattern in quite a different way. Thatis, children might have mastered that “ing” can occur with verbs. Theymight also know that “ly” is a familiar ending, but not know its func-tion. Thus, after some hesitation, they may trust the input and assignthe interpretation to the verb stem. The “lu” condition, in contrast, pre-sents an unfamiliar morpheme to the child. Since it is not in the famil-iar phonological or undifferentiated class, the children may assumethat it is not an ending attached to the base form and in fact choose thenonmatching alternative. To borrow from other work that Roberta andI have done, the children might see the “lu” form as so different thatthey (or at least the more sophisticated children) apply the lexical strat-egy of novel-name-nameless-category and choose the nonmatchingpicture for the linguistic stimulus (Golinkoff, Mervis, and Hirsh-Pasek1994). As this alternative permits the learning of new bound formsfrom the input, we favor that interpretation here and are preparing fur-ther studies with less familiar bound morphemes such as -ness to as-sess this hypothesis.

In sum, the intriguing pattern of results presented above allows us tosay with some conviction that children who are just beginning to usetwo-word sentences can detect (and perhaps use) bound morphology toassist them in constructing the grammar of their language. To learngrammar children must (1) be sensitive to these cues for constituentstructure; (2) be able to use these cues among others to label the con-stituents of grammar; and (3) be able to figure out how these constituent

204 Kathy Hirsh-Pasek

structures pattern in their own native tongue. Over the last severalyears, we have begun to make advances on the first two of these levels.The results presented here are yet another step in this progress.

What these results also highlight is the critical role that comprehen-sion data can play in our understanding of language acquisition. AsGolinkoff and I noted (Hirsh-Pasek and Golinkoff 1996a):

There can be little doubt that studies on young children’s lan-guage production in the past 25 years have provided a rich sourcefor language acquisition theories. Language production, the ob-servable half of the child’s language performance, however, isonly part of the story. Just as astronomers were not satisfied tostudy only the light side of the moon, so researchers in languageacquisition have long recognized that access from the “dark” sideof their topic—namely, language comprehension—illuminates thelanguage acquisition process far more than the study of produc-tion alone. (p. 54)

As can be seen in the analysis of the bound morpheme data, a numberof advantages can be obtained by looking at comprehension data. First,these data can be used to falsify theoretical assertions about the youngchild’s linguistic competence. In this case, data from language produc-tion have suggested that grammatical morphemes could not be used toassist the Stage I child in the learning of grammar (Pinker 1984). Datafrom comprehension present a different picture, suggesting that chil-dren are sensitive to both free and bound morphemes in the input andthat they might in fact be able to use this information to segment andperhaps identify grammatical constituents.

Second, comprehension data allow a clearer picture of the processesof language acquisition. By the time children are producing a structure,they have already acquired that structure. The steps leading up to mas-tery of the structure may be masked. Comprehension data, however,allow us to examine this process. If our hypothesis is correct and chil-dren do store familiar phonological information in an undifferentiatedstate before associating it with particular form classes, such storagemight only be visible in comprehension tasks.

Finally, comprehension studies allow for methodological control thatis often not possible in tests of production. With the exception of elicitedproduction tasks (Crain and Thornton 1991), those who examine pro-duction data are often in a position of “wait and see” in which theymust wait for the child to produce something in the hopes that they willsee the full repertoire of what the child can produce. Taking the boundmorpheme data as an example, comprehension allows us to look,specifically, at bound morphemes before they are produced.

Beyond Shipley, Smith, and Gleitman 205

The research presented in this chapter, then, both replicates and ex-pands some of the classic findings of Shipley, Smith, and Gleitman(1969). Children are sensitive to grammatical morphemes in the inputthat they hear. They are even sensitive to what is arguably the most dif-ficult class of grammatical morphemes—bound morphemes. Further,as noted in Shipley, Smith, and Gleitman’s original study, comprehen-sion does indeed precede production, and systematic examination oflanguage comprehension can provide a more accurate measure of thechild’s developing language. To fully understand what children bringto the language-learning task, how they can mine the input for cues togrammatical structure, and how they utilize a coalition of these cues tofind the building blocks of grammar, we will need to conduct extensiveand focused studies of their language comprehension.

Conclusions

The now classic Shipley, Smith, and Gleitman (1969) paper representsone area in which Lila set the stage for language research to come. Sheand Henry continue to be architects for our field. They not only frameresearch questions that must be addressed if we are to understand howyoung children acquire their native tongue, but they also point us in thedirection of new methodologies that can address these questions. Lilaand Henry will continue to influence psycholinguistic research foryears to come. The field is indebted to them and I feel honored to beamong those at the Gleitman dinner, among those to have been touchedby their brilliance.

Acknowledgments

The data reported here are the product of collaborative research withRoberta Golinkoff of the University of Delaware and Melissa Schweis-guth now of the University of California at San Diego. We gratefully ac-knowledge the support of the University of Delaware’s HonorsPsychology Program through which Melissa Schweisguth helped to de-sign the project and to collect the data. This research was also supportedby an NSF grant (#SDBR9601306) awarded to Hirsh-Pasek and Golinkoffand by an NICHD grant (#HD25455-07). Finally, we thank RebeccaBrand and He Len Chung for their able assistance in the data collectionand Elissa Newport for her thoughtful comments on this chapter.

References

Aslin, R., Woodward, J., LaMendola, N., and Bever, T. (1996) Models of word segmenta-tion in fluent maternal speech to infants. In Signal to Syntax, ed. J. Morgan and K.Demuth. Cambridge, MA: MIT Press, pp. 117–135.

206 Kathy Hirsh-Pasek

Bloom, P. (1990) Syntactic distinctions in child language. Journal of Child Language,17:343–356.

Bowerman, M. (1973) Structural relationships in children’s early utterances: Syntactic orsemantic? In Cognitive Development and the Acquisition of Language, ed. T. E. Moore.New York: Academic Press.

Brown, R. (1973) A First Language. Cambridge, MA: Harvard University Press.Crain, S. and Thornton, R. (1991) Recharting the course of language acquisition. In

Biological and Behavioral Determinants of Language Development, ed. N. A. Krasnagor,D. M. Rumbaugh, R. L. Schiefelbusch, and M. Studdert-Kennedy. Hillsdale, NJ:Erlbaum.

de Villiers, J. and de Villiers, P. (1973) A cross-sectional study of the acquisition of gram-matical morphemesin child speech. Journal of Psycholinguistic Research 2:267–278.

Gerken, L. (1996) Phonological and distributional information in syntax acquisition. InSignal to Syntax, ed. J. Morgan and K. Demuth. Cambridge, MA: MIT Press, pp.411–427.

Gerken, L. and McIntosh, B. J. (1993) The interplay of function morphemes in young chil-dren’s speech perception and production. Developmental Psychology 27:448–457.

Gleitman, L. and Gillette, J (1995). The role of syntax in verb learning. In The Handbook ofChild Language, ed. P. Fletcher and B. MacWhinney. Oxford: Blackwell, pp.413–429.

Golinkoff, R., Hirsh-Pasek, K., Cauley, K. M., and Gordon, L. (1987) The eyes have it:Lexical and syntactic comprehension in a new paradigm. Journal of Child Language14:23–46.

Golinkoff, R. M., Mervis, C., and Hirsh-Pasek, K. (1994) Early object labels: The case for adevelopmental lexical principles framework. Journal of Child Language 21:125–155.

Golinkoff, R., Hirsh-Pasek, K., and Schweisguth, M. A. (in press) A reappraisal of youngchildren’s knowledge of grammatical morphemes. In J. Weissenborng and B.Hoehle (eds.), Approaches to Bootstrapping: Phonological, Syntactic, and Neuro-physiological Aspects of Early Language Acquisition. Amsterdam, Philadelphia: JohnBenjamins.

Grimshaw, J. (1981) Form, function, and the language acquisition device. In The LogicalProblem of Language Acquisition, ed. C. L. Baker and J. McCarthy. Cambridge, MA:MIT Press, pp. 163–182.

Hirsh-Pasek, K. and Golinkoff, R. (1996a) The Origins of Grammar. Cambridge, MA: MITPress.

Hirsh-Pasek, K. and Golinkoff, R. M. (1996b) The intermodal preferential looking para-digm reveals emergent language comprehension. In Methods for Assessing Chil-dren’s Syntax, ed. D. McDaniel, C. McKee, and H. Cairns. Cambridge, MA: MITPress.

Jones, S. and Smith, L. (1993) The place of perception in children’s concepts. CognitiveDevelopment 62:499–516.

Katz, N., Baker, E., and MacNamara, J. (1974) What’s in a name? A study of how childrenlearn common and proper names. Child Development 45:469–473.

Kelly, M. (1992) Using sound to solve syntactic problems: The role of phonology in cate-gory assignments. Psychological Review 99:349–364.

Kelly, M. (1996) The role of phonology in grammatical category assignments. In Signal toSyntax, ed. J. Morgan and K. Demuth. Cambridge, MA: MIT Press, pp. 249–263.

Morgan, J., Meyer, R. P., and Newport, E. L. (1987) Structural packaging in the input tolanguage learning: Contributions of prosodic and morphological marking ofphrases to the acquisition of language. Cognitive Psychology 19:498–550.

Morgan, J., Shi, R., and Allopena, P. (1996) Perceptual bases of rudimentary grammaticalcategories: Toward a broader conceptualization of bootstrapping. In Signal toSyntax, ed. J. Morgan and K. Demuth. Cambridge, MA: MIT Press, pp. 263–287.

Beyond Shipley, Smith, and Gleitman 207

Pinker, S. (1984) Language Learnability and Language Development. Cambridge, MA:Harvard University Press.

Pinker, S. (1994) The Language Instinct. New York: William Morrow.Saffran, J. R., Aslin, R. N., and Newport, E. L. (1996) Statistical learning by 8-month-old in-

fants. Science 274:1926–1928.Shafer, V. L., Gerken, L. A., Shucard, J., and Shucard, D. (1995) An electrophysiological

study of infants’ sensitivity of English function morphemes. Unpublished manu-script, State University of New York, Buffalo.

Shipley, E., Smith, C., and Gleitman, L. (1969) A study in the acquisition of language: Freeresponses to commands. Language 45:322–342.

Taylor, M. and Gelman, S. (1988) Adjectives and nouns: Children’s strategies for learningnew words. Child Development 59:411–419.

Valian, V. (1986) Syntactic categories in the speech of young children. DevelopmentalPsychology 22:562–579.

208 Kathy Hirsh-Pasek

Chapter 13

Language and Space

Barbara Landau

While preparing for the event to celebrate Henry and Lila, I lookedthrough the many papers I had saved from my graduate studies atPenn. Among these were a draft of a manuscript that would be my firstpublished paper, some notes from one of my first “seminar” presenta-tions, and the penultimate version of the manuscript written by Landauand Gleitman on the subject of language learning by children who wereborn blind. These three artifacts remain, for me, palpable evidence ofthe impact that Lila and Henry have had on my professional life.

Looking at the draft manuscript of what would eventually becomethe “grandmother” paper, I can recall bringing the raw data to Lila andHenry during my first year of graduate school. I had actually collectedthe data as part of a Masters thesis at Rutgers, directed by Adele Abra-hamsen (who had introduced me to Lila the year before). When I firstdescribed the study to Lila, she listened patiently, then explained to mewhy the data were important and what they actually meant. She thenrecommended that I show the data to Henry, who spent the next severalmonths with me explaining how to conceptualize, analyze, and presentthe data to make a convincing argument. Following this, I wrote a firstdraft, which was then edited line-by-line by both Henry and Lila. Theresult was a wonderful paper, and I longed for them to be the coauthorsthey deserved to be. But when I suggested this, they declined, tellingme “this is really your work.” Nothing could have been further from thetruth, but this event reflects the first lesson I learned about Lila andHenry: They are great teachers, not only for their gift in educating theirstudents, but for their intellectual and personal generosity.

Looking at my seminar notes, I recall some of my earliest experiencesthere. These were lengthy seminars held in the Gleitman’s living room,sometimes running formally until close to midnight, and then continu-ing in the Gleitman kitchen until people dropped from exhaustion. (Lilaand Henry always were the last to succumb.) But they were the most ex-hilarating intellectual experiences I had ever had, and I rarely left with-out feeling privileged to have been a part of them. During one of my

first presentations, I told the seminar that I was interested in how blindchildren learn language. Henry’s immediate question was “Why?”—aquestion that stunned me, as it seemed self-evident that the blindwould provide an interesting symmetry to the recently published workby Feldman, Goldin-Meadow, and Gleitman on language learning bylinguistically deprived deaf children. But he was right to ask that ques-tion; and it became immediately apparent that the real answer wouldrequire thinking in depth about underlying assumptions, competingtheories, the connections between data and theory, and what the ulti-mate meaning would be of different empirical outcomes. This set thestage for my education under the Gleitmans’ watch.

What also stunned me was Henry’s lengthy sequel to his own ques-tion—one of many times in which he would use the student’s fledglingidea to teach. He set about brilliantly laying out the (il)logic of a ques-tion about language in the blind, followed by the logic of asking aboutspatial knowledge in the blind. Lila disagreed, rearticulating the ques-tion about language learning, and brilliantly reformulating it as shewent. Several other seminar members joined in, and the debate contin-ued all evening, and for many evenings thereafter. At some point dur-ing this lengthy process, my research questions became clearlyformulated and I became capable of defending them to the most pene-trating critic. This, I came to learn, was the format for the seminar: A stu-dent would present an ill-formed research question, Lila and Henrywould rearticulate and refine it (making it sound like the student was agenius along the way), and ultimately, that reformulation would be-come the student’s own. This was the second lesson I learned about Lilaand Henry: They are great scholars, not only for their brilliance, but fortheir dedication to fostering great work in others.

The final item I found was the penultimate version of the manuscriptwritten by Landau and Gleitman on the subject of language learning bychildren who were born blind. On these proofs were copious commentsin Henry’s hand, which reminded me vividly of the intense debates wehad had for the five years that we had worked on studies of the blindchild. The debates revolved around the question of whether the studyof the blind child was really about language, or really about space. Henryargued that the work was really about space, for if one could only un-derstand how spatial knowledge was constructed in the absence of vi-sual experience, it would follow trivially that language could belearned. Lila argued that the work was really about language, for al-though it was fascinating to learn how spatial knowledge could be builtupon nonvisual experience, it was impossible to understand how cer-tain aspects of language could be acquired unless one considered theprinciples of language itself as they interacted with experience.

210 Barbara Landau

Was the study really about language, or was it about space? The ques-tion found itself perfectly poised within the larger group at Penn, whichincluded two critical members of the psychology department (LizSpelke and Randy Gallistel) as well as other members of the SloanGroup (a group of linguists, psychologists, and computer scientists atPenn dedicated to the emergence of cognitive science). Within this con-text, I think we all finally concluded that it was truly about both—thatone could not understand how the blind child learned language unlessone understood how any child could come to represent the spatialworld, come to represent the formal system of human language, and,most critically, come to map these two together. But we only came to thisconclusion, I think, after years of debate, during which I learned to pre-sent ideas, to defend ideas, to criticize ideas, and to admire ideas, all inthe context of early morning coffees, late-night meetings, and perennialsupport, both personal and professional. Thus I learned my third lessonabout Henry and Lila: They are great mentors, for they give to their stu-dents intellectual direction for life. The set of profound and difficult is-sues that were laid out under Henry and Lila’s guidance during theseyears formed the subject matter of Language and Experience (Landau andGleitman 1985), and have continued to guide me since that time.

1.0 Initial Findings and Promissory Notes

In trying to understand how spatial experience is used during languagelearning, we began with the simple hypothesis of John Locke (1690):

If we will observe how children learn languages, we shall findthat, to make them understand what the names of simple ideas orsubstances stand for, people ordinarily show them the thingwhereof they would have them have the idea; and then repeat tothem the name that stands for it, as ‘white’, ‘sweet’, ‘milk’, ‘cat’,‘dog’. (Book 3.IX.9)

In our empirical studies of the blind child, however, we made somerather surprising discoveries that could not be explained by Locke’s hy-pothesis: The blind child developed a normal vocabulary, completewith rich representations of visual terms—spatial terms, color terms,and visual verbs such as look and see, which clearly could have had nobasis in “showing” things and “repeating the name.” At the end ofLanguage and Experience, we concluded with a much more complex hy-pothesis about word learning:

To explain how lexical learning based on different introducing circumstances in some domains yields up categories whose

Language and Space 211

substance and boundaries are much alike (e.g. see to blind andsighted children), we have argued that humans are endowed withrichly specified perceptual and conceptual principles that high-light certain construals of experience and suppress others; en-dowed with linguistic principles about which discriminationsamong millions of salient ones are lexicalizable; endowed withprinciples for manipulating the speech presented to the ear in cer-tain ways, but not in many other potentially available ways; andendowed with principles for pairing the perceptual-conceptualdiscriminanda with the lexical items. (p. 202)

Simply put, we proposed that there are universal principles thatguide the acquisition of new words despite very different kinds of ex-perience. At the same time, we proposed a very specific role for experi-ence: Regardless of how richly structured a child’s innate knowledge,some information from the environment must also be used to deter-mine the meaning of any word. This is because any given word mightbe compatible with an infinite number of possible meanings, but thechild cannot know in advance just which meaning is the one that thespeaker intends. For this reason, information from the environment—together with the learner’s natural predispositions in interpretingwords—can help serve as a “mental pointer” to the correct intendedmeaning. In the case of visual terms such as look and see, we proposedthat the blind child could have used the syntactic contexts in which theverbs occurred together with the nonlinguistic spatial contexts in whichthe word was used—the contexts of haptic exploration, in which shecould truthfully exclaim, “Let me see camera!”

As we knew at the time, the work on the blind child left many prom-issory notes. Fleshing out our hypothesis and testing its truth would de-pend on detailed studies of those richly specified perceptual andconceptual principles that highlight certain construals of experienceand suppress others; the kinds of linguistic principles that specifywhich discriminations are relevant to the lexicon; the kinds of princi-ples used in manipulating speech; and the kinds of principles that existfor pairing the two. A substantial amount of progress has been made ineach of these areas under Lila and Henry’s guiding hands and their de-scendants (see Gleitman and Gleitman 1997; and chapters by Fisher,Hirsh-Pasek, Goldin-Meadow, Naigles, and Newport, this volume).

Over the past twelve years, I have directed my attention to specific as-pects of these problems, focusing on the acquisition of words in two dif-ferent ontological domains—objects and places. In both cases, I havespent a fair amount of time puzzling about the kinds of perceptual, con-ceptual, and linguistic principles that could in fact be brought to bear on

212 Barbara Landau

the indeterminacy problem. What kinds of initial biases are there in theways in which learners represent objects, places, paths, and events?How do languages encode these notions? What skeletal conceptual andperceptual structures might map these onto various formal linguisticdevices, and thereby serve as an engine for further learning? How dolearners use their spatial and linguistic knowledge to learn the mean-ings of new words?

It turns out that the investigation of two domains (object, place) ismore than twice as complex as the investigation of one domain, and thishas necessitated a kind of breadth that is the foundation for cognitivescience today, but that served as a cornerstone of the Gleitman researchgroup long before it became fashionable. For example, the fundamentalorganizing principles of what is “salient” to the learner are qualitativelydifferent in the two cases of object and place. In the case of objects, wemust consider how objects are represented by the learner, how differentindividual objects are grouped into categories based on different kindsof similarity, what kinds of categories deserve to be lexicalized, howthese different kinds of categories are formally encoded in languages ofthe world, and how learners then actually learn the names for specificcategories. None of these is simple.

In the case of places, we must consider how a learner representsplaces geometrically, what kinds of geometric and force-dynamic rela-tions deserve to be lexicalized, how these relations are formally en-coded, and how learners acquire these place terms. The geometricrepresentation of “place” appears to be quite tightly constrained in hu-mans and other species (Gallistel 1990; Hermer and Spelke 1997;Landau, Spelke, and Gleitman 1984). Moreover, it is substantially differ-ent from the representation of objects qua objects, even though objectsoccupy locations and languages very often encode an object’s locationwith reference to other objects (Landau and Jackendoff 1993). Further,these linguistic terms appear to encode both more and less than the geo-metric properties engaged for navigation, and may constitute a distinctkind of semantic category specialized for talking about location. Add tothis substantial cross-linguistic variability: In natural languages, spatialrelationships are universally encoded as predicates—formal expres-sions of relationships—but their specific linguistic form class may bethe verb, preposition, postposition, or even various nominal markers(such as terms for “head-of” or “foot-of”; see Levinson 1992). Acrossthese forms, there is a fair amount of cross-linguistic variability in thekinds of spatial relationships that are encoded (e.g., English on covers abroader group of cases than German aan or auf; see Bowerman 1996), al-though these differences may reflect featural choices based on universalsemantic properties (Landau and Jackendoff 1993; Landau 1996).

Language and Space 213

In the remainder of this chapter, I will confine myself to work on howobjects are encoded for the purpose of naming. In this work, I have triedto fill at least some of the promissory notes left by Language andExperience. Even within this domain, things are quite complex.

2.0 Objects Named

One of the most important findings of the blind study is one that is notcited very often: Blind children develop a vocabulary of object namesthat is virtually indistiguishable from that of sighted children of thesame ages. Thus, with or without visual experience, children acquireroughly the same names for roughly the same kinds of objects and thesenames are generalized appropriately with little explicit tutoring. Whatis the basis for this learning?

It is commonplace to assume that generalization of object names isbased on the child’s understanding that object names are cover termsfor “object kinds”—objects that are considered by the linguistic com-munity to be relevantly similar to each other (Markman 1989). Muchdebate has revolved around the nature of these similarities—whethertheir foundation is innate knowledge of basic ontological kinds (object,substance, etc.) or whether the similarities are learned through lan-guage (for different views, see Keil 1979; Quine 1960; Soja, Carey, andSpelke 1991); whether the similarities holding among “natural kind”objects are qualitatively different from those holding among manmadeobjects (Kripke 1977; Putnam 1977; Malt and Johnson 1992); whetherthe similarities are specific to lexicalized entities, or are general similar-ities that are prepotent in all kinds of similarity tasks (Landau, Smith,and Jones 1988; Markman and Hutchinson 1984; Smith, Jones, andLandau 1996; Waxman and Markow 1995). To some, the very notion ofsimilarity as a theoretical construct is misguided, too slippery to everplay a significant role in theory construction (Goodman 1972).

But things that fall under the same object name are similar to eachother in some sets of ways and not in others; and if we are to understandhow it is that blind and sighted children can easily learn to assign aname to only certain objects (and not others), we must ask what kinds ofsimilarity do matter in object naming, and what kinds do not. The ques-tion thus is not whether similarity matters, but what kinds of similaritymatter, and how these differ for different domains, for different tasks,and for different developmental moments.

Quine (1969) proposed two quite different kinds of similarity: One is“intuitive” similarity, present in many species and rooted in the sensoryand perceptual systems—for example, similarities among colors—which are a strict function of the neural structures and psychological/computational mechanisms that determine color perception. A second

214 Barbara Landau

kind is “theoretical” similarity; this allows us to construct and observesimilarities that go beyond the perceivable properties of objects.Theoretical similarities are especially useful in explaining why thingsfall into the same named category and might include similarities basedon feeding or reproductive behavior, social behavior, evolutionary con-siderations, or highly specific goals guiding categorization (Medin andColey 1998; E. Shipley, this volume). Quine says:

A crude example (of theoretical similarity) is the modification ofthe notion of fish by excluding whales and porpoises. Another tax-onomic example is the grouping of kangaroos, oppossums, andmarsupial mice in a single kind, marsupials, while excluding ordi-nary mice. By primitive standards the marsupial mouse is themore similar to the ordinary mouse than to the kangaroo; by theo-retical standards the reverse is true. (p. 167)

Clearly, our mature knowledge of object categories must engage sim-ilarities that are not necessarily perceptual. People regularly make deci-sions about kinship on the basis of true blood relationships rather thanappearance: A grandmother is the mother of a parent; she may or maynot have gray hair and wrinkles, even though four-year-old childrenmay indiscriminately call all gray-haired, wrinkled women “grand-mas” (Landau 1982). Analogously, when scientists decide how to clas-sify objects in nature, they rely on properties deemed to be important incurrent scientific understanding of the nature of different kinds, for ex-ample, an animal’s digestive or reproductive system, its lineage, its eco-logical niche. Recent work in cognitive development has shown that,from the age of around four, children also make judgments based onsimilarities other than the perceptual (Gelman and Wellman 1992; Keil1989).

Yet acknowledging that such bases for classification are possible doesnot mean that perceptual—intuitive—similarity is unimportant, noreven that it is less important. Consider the recent headline in the NewYork Times Science Section (Sept. 19, 1995); and see fig. 13.1.

Strange Bird Must Think It’s a Cow

The article describes the work of Alejandro Grajal, an ornithologist whostudied the hoatzin, a tropical bird with a digestive system called“foregut fermentation” similar to that of cows as well as Colombinemonkeys, kangaroos, and tree sloths. This discovery clearly is impor-tant in understanding the nature of the species; but it is unlikely tocause a change in what the thing is called—bird. Presumably, this ani-mal was originally named at a time when the nature of the animal’s digestive system was unknown; this “original dubbing ceremony”

Language and Space 215

(Kripke 1977) may have been conducted on the basis of perceptual similarities. What was known about the animal and what therefore mayhave determined its name was its appearance and probably, how it be-haved. Similarly, in Quine’s example, the marsupial mouse is similar toordinary mice in its appearance, hence it we call it “mouse,” despite itscloser theoretical similarity to the kangaroo. These uncontroversial factsraise an important question: Why is intuitive similarity sometimes abetter predictor of an object’s name than theoretical similarity?

The answer may have to do with learning by young children. First,even if theoretical similarities do play an important role in early devel-opment, such similarities can be hard for young learners to discover.Even when they are relatively easy to discover, there must be somemechanism for them to be linked with a relatively quick and reliable“identification” function—the function that gives us our first hypothe-ses about which of the objects in the world do belong to a given cate-gory (see, e.g., Armstrong, Gleitman, and Gleitman 1983; Landau 1982;Smith and Medin 1981; Keil 1994). For it to be a useful function forlearning by infants and young children, the mechanism should select aproperty or properties that are easily picked up by learners and highlypredictive of category membership. If so, then the learner will not goastray even in the earliest stages of word learning.

2.1.0 Object shape: A privileged kind of similarity for early object namingWhat kinds of similarities could serve this function? Abundant research

216 Barbara Landau

Figure 13.1.Strange Bird Must Think It’s a Cow (Reproduced with permission of Dr. Grajal)

in visual perception tells us that three-dimensional object shape is criti-cal to object recognition in adults (Marr 1982; Biederman 1987). In somecurrent theories, basic-level objects—airplanes, cups, cars—are recog-nized through decomposition into parts, either by analysis of contourminima (Hoffman and Richards 1984) or by specific arrangements ofvolumetric primitives (Biederman 1987). Perhaps not coincidentally, thebasic level seems to provide an easy entry point in object naming(Brown 1957; Rosch, Mervis, Gray, Johnson, and Boyes-Braem 1976;Waxman and Markow 1995). Two-dimensional outline drawings canengage representations of objects as well, producing rapid and error-free object recognition and identification at the basic level in adults.Surface color appears to be much less important in the process of iden-tification (Biederman and Ju 1988).

Could representations of object shape underly object recognition byinfants as well—and hence be a plausible candidate representation to beengaged during early learning of object names? Recent results showthat four-month-old infants can recognize the similarity among chairs,compared with couches or tables, even when the objects in questionhave quite complex configurations (Behl-Chadha 1996), thus suggest-ing that complex perceptual similarities are computed well before thechild learns names for things. This idea is consistent with the classicfindings of Hochberg and Brooks (1962): These investigators preventedtheir own infant from observing any two-dimensional representationsof objects over the first year and a half of life. At the end of the period,the child had acquired a reasonable vocabulary of object names, pre-sumably on the basis of observing real three-dimensional objects andhearing them named. His parents now showed him line drawings of fa-miliar objects, and asked him to name them—which he did.

Results such as these suggest that object shape may provide a privi-leged kind of similarity in the early acquisition of object names.Although there is in principle an infinite number of possible interpreta-tions of a novel word (one manifestation of the indeterminacy problemfor language learning), learners who entertained each of these interpre-tations would be lost in the wilderness while trying to learn the word“dog.” Fortunately, this does not appear to happen.

2.1.1 The shape biasIn a number of studies, my collaborators and I have shown that youngchildren do show a preference for generalizing the names of novel ob-jects on the basis of shape (Landau, Smith, and Jones 1988). The taskhere is simple: Two- and three-year-olds are shown a novel artifact-typeobject and hear it named, for example, “See this? This is a dax.” Thenthey are shown a series of test objects, one at a time, and asked eachtime, “Is this a dax?” When a test object is the same shape as the original

Language and Space 217

object but differs from it in size, color, or surface texture, subjects asyoung as two years of age and as old as adults will accept the object asan instance of “a dax” (Landau et al. 1988; Landau et al. 1992; Smith,Jones, and Landau 1992). However, when a test object has a differentshape from the original—even if it is just the same in size, color, or tex-ture—children and adults alike tend to reject it, saying it’s “not a dax.”

It is important to note that shape is not equally salient across all con-texts, but rather appears to be especially salient in the context of objectnaming. For example, it is not the preferred pattern of generalizationwhen the task is converted to a similarity task that does not involve aword: If just asked whether a test item “matches” or “goes with” or is“the same as” a standard object, children are much more likely to showpreference patterns that are based on overall stimulus salience (such asbrightness or surface texture; Smith et al. 1992), thematic sorting prefer-ences (Markman and Hutchinson 1984), or perhaps the overall frame-work of stimulus choices and context, which serves as a mental pointerto the “relevant” dimension of similarity for adults (Medin, Goldstone,and Gentner 1990).

The preference for shape in object naming shows up not only in ex-perimental contexts, but in very many naturalistic contexts that reflectthe use of our mental representations of objects for goals other than lan-guage. One need not be a scholar of art to recognize the important rolethat shape plays in explicit representations of objects. Two sculptures ofClaes von Oldenburg are excellent examples: One is his monumentalmetal “The Clothespin” in downtown Philadelphia. Sixty feet tall andsculpted of steel, The Clothespin looks just like one. A second is his“Bicyclette Ensevelie” (Buried Bicycle)—sections of handlebar, wheeland seat arranged on the ground over the span of a large outdoor parkin Paris, yet immediately recognizable as a representation of a bicycleprotruding from the ground. Our eager adoption and quick under-standing of these names as they label explicit external representationsof objects suggests a profound importance for object shape in the task ofobject naming.

2.1.2 The critical role of syntaxThe results on shape and naming begin to tell us about preferred objectrepresentations as the entry point for language learners, but they alsotell us about constraints on the linguistic side: The preference for shapeis specific to syntactic contexts that are appropriate for object naming.Landau and Gleitman (1985) and Gleitman (1990) proposed that syntac-tic contexts are critical for establishing the meanings of verbs. Our re-cent work on object naming extends work started by Brown (1958),

218 Barbara Landau

showing that count nouns, mass nouns, and adjectives also serve asmental pointers to different basic aspects of meaning.

In English, the count noun context is appropriate for object naming;in these contexts, a noun is combined with determiners such as “a” and“an” and quantifiers such as numerals. This syntactic context marks thefact that the named entity is discrete and countable; such entities rangeover concrete and abstract objects (such as “dog” and “belief,” respec-tively), and might best be characterized as “individuated entities”(Bloom 1996b). For entities that are not countable—such as sub-stances—English uses the mass noun context, in which the noun is in-troduced by determiners and quantifiers such as “some” and “more.”Mass nouns can also be quantified by classifiers such as “a piece of,” “apile of,” “a hunk of,” (granite, sand, chocolate). Adjectives name prop-erties, including object properties such as specific shape, texture, andcolor.

Young children are quite sensitive to the syntactic context in which aword occurs, and their generalization differs accordingly. Althoughchildren generalize on the basis of shape in the count noun context, theygeneralize on the basis of surface texture or coloration in the context ofadjectives (e.g., “This is a daxy one,” Landau et al. 1992; Smith et al.1992), on the basis of material substance in the context of mass nouns(e.g., “This is some dax,” Subrahmanyam, Landau, and Gelman 1997),and on the basis of the object’s location in the context of prepositions(e.g., “This is adax the box,” using a novel form that is morphologicallysimilar to known prepositions such as “across” or “adjacent” [Landauand Stecker 1990]). In each of these contexts, children’s attention to par-ticular properties is strongly modulated by syntactic context. Thus thesyntactic context serves as a critical mental pointer to different funda-mental ontological categories—object, property, substance, and place.

Many results now indicate that the influence of syntactic context inconstructing meaning grows over development, allowing children tomove beyond their initial biases for representing objects and events.Consider objects. If a speaker wishes to refer to the object itself, anddoes so describing it with the sentence “This is a dax” or “What a nicedax,” the child’s preferred representation—one engaging the object’sshape—will often be sufficient for the learner to attend to just what thespeaker has in mind, namely, the object itself. But suppose the speakerwishes to talk about the material of which the object is made, ratherthan the object itself. If object shape is a preferred representation, thenother properties—including material—might be ranked lower, andtherefore it might be more difficult to switch attention from one’s pre-ferred representation to the one that the speaker actually has in mind.

Language and Space 219

The argument here is not that material, color, surface texture, or locationcannot be represented, nor that they are somehow less “natural” thanobject shape. Rather, understanding what someone is saying requiresthat the listener direct his or her attention to just that interpretation in-tended by the speaker. If object shape is highly salient under a variety ofconditions, then it should be relatively difficult for the young learner topry his or her attention away from shape toward some other property,until syntax plays a strong enough role.

Recent results from Subrahmanyam et al. (1999) have shown signifi-cant growth in children’s ability to use syntactic context to modulatetheir interpretations of the speaker’s intended meaning. At three yearsof age, children who observe a rigid three-dimensional object are biasedto generalize a novel noun on the basis of shape whether they hear thenoun in the context of count or mass noun. By five years of age, how-ever, children who observe such an object will generalize on the basis ofshape when they hear a count noun, but on the basis of material sub-stance when they hear a mass noun. Adults do the same, strongly andabsolutely.

A very similar developmental course has been found in the domain ofevents. Fisher, Hall, Rakowitz, and Gleitman (1994) found that three-year-olds show a strong “agency” bias in interpreting the meanings ofnovel action verbs that are presented without syntactic context. For ex-ample, if children observe a scene in which one toy animal hands a ballto another toy animal, and they hear “Look! Ziffing!” the children as-sume that “ziffing” means “giving” rather than “taking,” even thoughboth verbs are plausible descriptors of the scene. This bias to encode theverb as one that focuses on a causal agent is present among adults aswell. However, when the novel verb is presented in a syntactic context,this agency bias is overriden: Subjects at all ages interpret the verb as“give” if they hear “The elephant is ziffing the ball to the bunny” or“take” if they hear “The bunny is ziffing the ball from the elephant.”Importantly, Fisher et al. found that the role of syntax grows over devel-opment, starting out by modulating children’s interpretations onlyprobabilistically, but ending by modulating adults’ interpretationsstrongly and absolutely.

Thus, for object names and for action verbs, it appears that the devel-opmental course begins with young children interpreting new words inconcert with their perceptual and conceptual biases—especially whenthese correspond to the interpretation offered by the syntactic context.The developmental course ends with older children (and adults) de-pending quite strongly on syntactic context, overruling preferred per-ceptual interpretations. The great genuis of language, of course, is tocarry us beyond our perceptual biases. At the same time, the great ge-

220 Barbara Landau

nius of perceptual biases may be to allow learners a wedge into the lin-guistic system at all.

3.0 Objections and Responses

In the Gleitman research seminar, empirical findings were alwaystreated with respect but also tempered with a healthy dose of skepti-cism. With a scowl, Henry might ask: “Let us suppose that shape is thepreferred dimension of generalization for object naming. Still, I am wor-ried. . . . What could this mean?” Indeed, since the earliest publicationof this work on shape and naming, it has met with many objections andchallenges. These are the most critical:

1. The objects used in most of the shape studies are novel artifacts in-vented for the purposes of experimentation—a poor representation ofthe artifacts that actually exist in the world. Even as artifacts, their sim-ple geometric design suggests no plausible function. Because they donot belong to any existing natural category, they force a preference forshape in the absence of any information suggesting alternatives (p.c.,audience for virtually every colloquium in which I have presented thesefindings).

2. In any case, the true representations underlying an object’s nameare representations of its “kind.” Young children do not seek to put to-gether objects of the “same shape,” but rather, objects of the “samekind” (Soja, Carey, and Spelke 1992). In the case of artifacts, our true cri-terion for membership in the same kind is neither apparent function norappearance, but rather the creator’s intention (Bloom 1996a).

These objections call for empirical and theoretical response. First, theidea that lack of functional information may lead to a default relianceon shape calls for a direct empirical test: We can provide learners withadditional, richer information and find out whether their patterns ofgeneralization change. Second, the idea that young children are reallyseeking to name objects of the same kind with the same name calls formore explicit theoretical discussion of the possible links among sameshape, same kind, and same name.

3.1 The Role of Function and General World Knowledge

Recently, we have investigated the role of functional information inchallenging the shape bias (Landau, Smith, and Jones 1997; Smith,Jones, and Landau 1996). Simply put, we have asked whether providingclear functional information about an object will lead children andadults to generalize its name on the basis of properties that can supportits function. If so, this would suggest that people’s reliance on shape

Language and Space 221

occurs only in circumstances in which they have relatively impover-ished information about other characteristics of the object. If functiondoes not enter into naming, however, this would suggest a rather stronghypothesis about the nature of naming and the role of shape perception,specifically, that naming might be cut off from more thoughtful, reflec-tive processes that act to store and manipulate our general knowledgeabout objects.

To anticipate, we have gone to lengths to make functional informa-tion salient—even using familiar functions—but have found that youngchildren are quite resistant to naming on the basis of functional proper-ties. This is despite the fact that the same young children are perfectlycapable of using functional information to make other (nonnaming)judgments about objects. In contrast to the pattern found among youngchildren, functional information does readily enter into adults’ namingjudgments, suggesting dramatic and important developmental changesin the kinds of information that enter into the learning of object names.

In one set of experiments, we studied two-, three-, and five-year-olds’and adults’ naming patterns with and without functional information(Landau et al. 1997). Subjects in the Function condition were providedwith information about function while they heard the objects named,whereas subjects in the No Function condition only heard the name andthus knew nothing about the objects’ intended functions. Subjects in theFunction condition were told very explicitly what the objects were for,for example: “This is a dax. Daxes are made by a special company justso they can mop up water”; or “This is a rif. And this is what I do with it.I use it to pull toys from across the table.” After hearing the standard ob-ject named (and, in the Function conditon, observing the function andhearing it described), all subjects were asked to generalize the object’sname to new objects. Some test objects were the same shape as the stan-dard, but could not carry out its designated function; others were a dif-ferent shape from the standard, but could carry out the function. Inaddition to asking subjects whether each of these objects was “a dax,”we also independently checked on how much subjects knew about theobjects’ functions by asking them directly which of the objects couldcarry out specific functions.

Over three experiments, the objects varied in shape, material, andother properties relevant to function. For example, in one experiment,we used objects with simple geometric shapes, composed of materialsthat could support specific functions (cork for one set, with the functionof holding stick pins; sponge for a second set, with the function of mop-ping up water). In another experiment, we used objects similar toknown familiar objects with easily understood functions: novel con-tainers (used to carry water) and canes (used to retrieve toys from

222 Barbara Landau

across a table). In a third experiment, we used bonafide artifacts: combsand clothespins.

The results across the three experiments were remarkably consistent.Two- and three-year-olds, whether instructed about function or not,generalized the standard’s name on the basis of shape. This was trueacross the range of objects, from “nonsense” objects with simple geo-metric shapes to well-known objects such as the comb or clothespin. Inthe latter case, this meant that children were more willing to generalizethe name “comb” to a paper cut-out having the identical shape and sizeof the original comb than to objects that could carry out the function(but were clearly different shapes from the standard comb). Similarly,the children were more likely to generalize a novel name “dax” or “rif”to a container that was the same shape as the standard, even if it hadholes in the bottom and so could not carry out the designated functionof carrying water. Thus the pattern of naming among young childrenwas consistent with the critical role of shape in early object naming.

The pattern among adults was quite different, however. Adults whosaw novel objects and were not told about their functions generalizedfreely on the basis of shape—just as they did in earlier shape studies.Adults who saw novel objects and were told about the objects’ functions,however, generalized the name on the basis of the objects’ functionalproperties—either material and substance or global object propertiessuch as length and rigidity that were critical to the demonstrated func-tions. Finally, adults who observed real, familiar objects (the comb,clothespin) also generalized on the basis of shape, but importantly, theywere very conservative, rejecting many same-shape objects as well assame-function objects. For example, adults were willing to call a papercut-out a “comb,” but they did so only with reluctance, as shown bytheir lower rates of acceptance. Those who rejected these items tendedto spontaneously add comments such as “Well, yes, you could call it acomb, but it’s really a piece of paper in the shape of a comb.” Five-year-olds showed a similar pattern to adults, though somewhat weaker.

Thus the developmental picture is complex. For young children,naming seems to be governed by shape similarity (in these contexts),and functional information is unlikely to enter into naming decisions(but see Kemler-Nelson 1995 for a different set of findings). For olderchildren and especially adults, the importance of shape similarityseems to be strongly modulated by a variety of factors: whether the ob-ject is familiar, how much functional information is known, and generalworld knowledge (see also Malt and Johnson 1992). Adults appear tohave a rather complex metric for deciding whether form or functionmatters the most, but children’s decisions appear to be considerablysimpler.

Language and Space 223

It is important to note that when children were directly queried aboutfunction, their responses were quite different. They did not generalizeon the basis of shape, and they often made the correct judgments aboutfunction—especially when the objects were familiar. This is consistentwith research showing that appreciation for object functions begins ininfancy (Brown 1990; Kolstad and Baillargeon 1991). So children clearlyunderstood something about the objects’ functions. Note, however,they were not perfect; and there was pronounced development in howthe children articulated their knowledge. For example, five-year-olds inour studies were quite knowledgeable, able to tell whether each of theobjects could carry out specific functions whether they had been in-structed with the standard or not. But two- and three-year-olds’ knowl-edge was much spottier: Three-year-olds could determine whether anobject could carry water or retrieve a toy, and whether somethingwould work to comb the hair or hang clothes. They were less good atdetermining which objects could mop up water (sponge) or hold a stickpin (corkboard). Two-year-olds’ knowledge was even shallower: Theycould tell which objects would work to comb the hair or hang clothes,but only by actually trying out the objects in question (on a model headof hair or a toy clothesline). Furthermore, their reasoning as they de-cided what would “work to comb your hair” revealed immature knowl-edge of what it would take: In many cases, merely contacting the hairwith the comb appeared to suffice for the judgment that “it worked” to“comb hair,” confirming what every parent knows.

To summarize, shape and function appear to play different roles inobject naming for children than for adults. Functional information doesnot appear to enter into object naming among young children evenwhen—in other tasks—the same age children can show that they do un-derstand some aspects of the objects’ functions. A variety of studies, in-cluding ours, have found that functional information begins to befirmly integrated into object-naming judgments from about age four orfive on (Gentner 1978; Merriman, Scott, and Marazita 1993; Landau etal. 1997; Smith et al. 1996; but see Kemler-Nelson 1995 for a differenttimetable). This suggests that early object naming may be cut off fromthe influences of many kinds of general world knowledge. Further-more, children’s understanding of functions undergoes considerableenrichment between the ages of two and five years, only beginning toapproximate adult knowledge well past the time when the object vo-cabulary is first learned (see also Keleman 1995; Matan 1996). In con-trast, functional information does appear to play an important role inobject naming among adults, particularly when one must decide whetherand how to extend a novel object name. For familiar objects, shape andfunction appear to cohabit adult mental representations, with each

224 Barbara Landau

dominating the other under different circumstances (Malt and Johnson1992).

3.2 How Are Same Shape, Same Kind, and Same Name Related in LanguageLearning and Mature Naming?

For a shape bias to make sense, it is necessary to link it to the notion of“same kind.” Names for categories of things are names for things of thesame kind: We call “cups” those things that are grouped together invirtue of their membership in some kind. This kind of observation hasled some to argue that the “true” basis for object naming in young learn-ers and in adults is “same kind,” not “same shape” (Soja, Carey, andSpelke 1992). That is, children seek to name objects in accord with theirkind, not in accord with their shape.

Assuming that this is true for learners as well as adults, it is still in-complete. Although it postulates a link between the notion of objectkind and object name, it does not solve the problem of how the child cantell which objects do, in fact, belong to the same kind. Because there arelarge differences between what young children know and what theircaregivers know, it is likely that the “true” criteria for selecting what be-longs in the same kind will change over development. We have seen anexample of this with function: Although young children do not appearto consider object function in their judgments of what a thing is called,adults do. The same is undoubtedly true for natural kind objects:Because children’s knowledge changes dramatically over development(Carey 1985), it is likely that older children and adults will use differentproperties—and possibly even different principles—to categorizemembers of these categories. Recent evidence shows that, even amongadults with different background, there are large differences acrossgroups in their criteria for categorizing different kinds of plants (Medinet al. 1997).

Why, then, does young children’s naming of objects seem to match—more or less—the naming of adults around them? It seems likely thatshape similarity plays an important role here: If young children gener-alize on the basis of similarity in object shape, then their object namingwill often match that of adults. Thus communication can proceed be-cause of a relatively straightforward commonality in the mind of thechild learner and the adult. Whether or not young children are trulysearching for same kind, an initial reliance on shape similarity will setthem on the road to acquiring an object name vocabulary.

I say “initial” because it is clear that same shape is neither necessarynor sufficient in mature judgments of same kind objects (see Bloom 1996afor some controversies and examples that push our own intuitions).

Language and Space 225

However, shape is an excellent beginning, because it correlates quitestrongly with same kind: Objects of the same shape very often are mem-bers of the same kind. Objects of the same shape and same kind areoften called by the same name by the child’s linguistic community. Thismeans that any child who is sensitive to the correlations between sameshape, same kind, and same name will often generalize object namescorrectly, in agreement with those around them. Of course, objects ofthe same kind do not always share the same shape, and this is why ashape bias can only provide an initial, though crucial, bootstrap into theobject-naming process.

What happens to push children toward more complex information inconsidering which objects should have the same name? In some cases,children will have the opportunity to hear the same object name ap-plied to two objects having very different shapes. Elizabeth Shipley andI have investigated what kinds of generalization occurs in such cases,and we have found that young children are likely to “fill in” the inter-mediate space, generalizing the name to all objects along the similarityline that fits between the two standards (Landau and Shipley 1996). Incontrast, if children hear two different object names applied to the sametwo objects, they will generalize as if two separate categories exist. Thislatter finding is consistent with the fact that, across languages, childrenwill have to create somewhat different distinctions (Imai and Gentner1997). Thus, even starting with a bias to generalize on the basis of shape,young children are still free to modulate this bias in accord with the dis-tribution of names for objects in his or her language.

The importance of a shape bias is that it provides an initial guide tosame kind, and thus same category, through a completely natural link:The same representational system that underlies object recognition islinked to the system that underlies object naming. Note that this guidemay be quite strong: Our initial studies of object function suggest thatthe tendency to map same shape to same name may be separated, cutoff from the influences of many other kinds of knowledge. If so, thiswould prove beneficial as well—a simple hypothesis, which is oftencorrect, may be better for young learners than a complex one that re-quires sustained, thoughtful reflection. Without such a constraint, thelearner might choose any similarity as the basis for belonging in thesame kind—and the chances would be that different learners wouldhave different conjectures, different also from those of the languagecommunity around them. By engaging the simple hypothesis that sameshape licenses same name, the learner is provided guidance—in ad-vance of knowing the category—to “highlight certain construals of ex-perience and suppress others.” Such a conjecture, one that naturally

226 Barbara Landau

links same shape with same kind, is necessary; for no matter how muchinnate knowledge the child has, she will require a means to map thisknowledge to things in the world.

Thus the shape bias can serve as a mechanism for getting learnersstarted. Once a common vocabulary is established, it is possible to com-municate other kinds of information about objects—artifact functions,the intentions of those who design artifacts, animal behaviors, mecha-nisms of respiration in plants, etc. Without a common vocabulary, how-ever, none of this is possible.

4.0 Conclusions

In some respects, work on object naming and its link to the human ob-ject recognition system might seem a far cry from research designed todetermine how the blind child learns language. However, the principlesthat are revealed in the two cases are surprisingly similar: In both cases,it has seemed important to specify not just what the child’s innateknowledge might be like, but also, how he or she might use that knowl-edge to learn the words for objects and events. In both cases, the role ofspatial representation is prominent: We cannot answer the question ofhow one learns words for objects and events without understanding therepresentational systems that underlie our nonlinguistic knowledge ofthese. In both cases, the role of linguistic representation is prominent:We cannot know how learning proceeds without understanding howformal linguistic devices are used by learners to “point” toward differ-ent aspects of meaning.

My own need to understand both systems of knowledge and their in-teraction stems directly from the questions raised by Henry and Lilaearly in my career: Is this work really about language, or is it reallyabout space? The ensuing framework has provided me with directionover the years since work on the blind child—direction which continu-ally reminds me how deep, complex, and mysterious it is that any child,blind or sighted, can learn to talk about what she perceives. For this, Ithank them.

Acknowledgments

Preparation of this paper was supported by grants RO1 HD-28675 fromNICHD, RO1 MH-55240 from NIMH, and a General UniversityResearch Grant from the University of Delaware. I wish to thankCynthia Fisher, Elissa Newport, and Elizabeth Shipley for helpful com-ments on the paper.

Language and Space 227

References

Armstrong, S., Gleitman, L. R., and Gleitman, H. (1983) What some concepts might not be.Cognition 13:263–308.

Behl-Chadha, G. (1996) Basic-level and superordinate-like categorical representations inearly infancy. Cognition 60:105–141.

Biederman, I. (1987) Recognition-by-components: A theory of human image understand-ing. Psychological Review 94:115–147.

Biederman, I and Ju, G. (1988) Surface vs. edge-based determinants of visual recognition.Cognitive Psychology 20:38–64.

Bloom, P. (1996a) Intention, history, and artifact concepts. Cognition 60:1–29.Bloom, P. (1996b) Controversies in language acquisition: Word learning and the part of

speech. In Perceptual and Cognitive Development, ed. R. Gelman and T. Kit-Fong Au.San Diego: Academic Press.

Bowerman, M. (1996) Learning to structure space for language: A cross-linguistic per-spective. In Language and Space, ed. P. Bloom, M. A. Peterson, L. Nadel, and M. F.Garrett. Cambridge, MA: MIT Press.

Brown, A. (1990) Domain-specific principles affect learning and transfer in children.Cognitive Science 14:107–133.

Brown, R. (1958) Words and Things: An Introduction to Language. New York: Free Press.Carey, S. (1985) Conceptual Change in Childhood. Cambridge, MA: Bradford Books/MIT

Press.Fisher, C. (2000) Partial sentence structure as an early constraint on language acquisition.

(Chapter 16 in this volume.)Fisher, C., Hall, G., Rakowitz, S., and Gleitman, L.R. (1994) When it is better to receive

than to give: Syntactic and conceptual constraints on vocabulary growth. In TheAcquisition of the Lexicon, ed. L. R. Gleitman and B. Landau. Cambridge, MA: MITPress.

Gallistel, C. R. (1990) The Organization of Learning. Cambridge, MA: MIT Press.Gelman, S. and Wellman, H. (1991) Insides and essences: Early understandings of the

non-obvious. Cognition 38:213–244.Gentner, D. (1978) What looks like a jiggy but acts like a zimbo? A study of early word

meaning using artificial objects. Papers and reports on child language development15:1–6.

Gleitman, L. R. (1990) The structural sources of verb meanings. Language Acquisition1:3–55.

Gleitman, L. R., and Gleitman, H. (1997) What is a language made out of? Lingua 99:1–27.Goldin-Meadow, S. (2000) Learning with and without a helping hand. (Chapter 9 in this

volume.)Goodman, N. (1972) Problems and Projects. Indianapolis: Bobbs-Merrill.Hermer, L. and Spelke, E. (1994) A geometric process for spatial reorientation in young

children. Nature 370:57–59.Hirsh-Pasek, K. (2000) Beyond Shipley, Smith, and Gleitman: Young children’s compre-

hension of bound morphemes. (Chapter 12 in this volume.)Hochberg, J. and Brooks, V. (1962) Pictorial recognition as an unlearned ability: A study of

one child’s performance. American Journal of Psychology 75:624–628.Hoffman, D. and Richards, W. (1984) Parts of recognition. Cognition 18:65–96.Imai, M. and Gentner, D. (1997) A cross-linguistic study of early word meaning: Universal

ontology and linguistic influence. Cognition 62(2):169–200.Keil, F. (1979) Semantic and Conceptual Development: An Ontological Perspective. Cambridge,

MA: Harvard University Press.

228 Barbara Landau

Keil, F. (1989) Concepts, Kinds, and Cognitive Development. Cambridge, MA: CambridgeUniversity Press.

Keil, F. (1994) Explanation, association, and the acquisition of word meaning. InAcquisition of the Lexicon, ed. L. R. Gleitman and B. Landau. Cambridge, MA: MITPress, 169–198.

Keleman, D. (1995) The Teleological Stance. Ph.D. thesis, University of Arizona.Kemler-Nelson, D. (1995) Principle-based inferences in young children’s categorization:

Revisiting the impact of function on the naming of artifacts. Cognitive Development10:347–354.

Kolstad, V. and Baillargeon, R. (1991) Appearance and knowledge-based responses tocontainers in infants. Unpublished manuscript.

Kripke, S. (1977) Identity and necessity. In Naming, Necessity, and Natural Kinds, ed. S.Schwartz. Ithaca: Cornell University Press.

Landau, B. (1982) Will the real grandmother please stand up? The psychological reality ofdual meaning representations. Journal of Psycholinguistic Research 11(1):47–62.

Landau, B. (1996) Multiple geometric representations of objects in languages and lan-guage learners. In Language and Space, ed. P. Bloom, M. A. Peterson, L. Nadel, andM. Garrett. Cambridge, MA: MIT Press.

Landau, B. and Gleitman, L. R. (1985) Language and Experience. Cambridge, MA: HarvardUniversity Press.

Landau, B. and Jackendoff, R. (1993) “What” and “where” in spatial language and spatialcognition. Behavioral and Brain Sciences 16:217–265.

Landau, B. and Shipley, E. (1996) Object naming and category boundaries. In Proceedingsof the Boston University Conference on Language Development, ed. A. Stringfellow.Brookline, MA: Cascadilla Press.

Landau, B., Smith, L., and Jones, S. (1988) The importance of shape in early lexical learn-ing. Cognitive Development 3:299–321.

Landau, B., Smith, L., and Jones, S. (1992) Syntactic context and the shape bias in chil-dren’s and adults’ lexical learning. Journal of Memory and Language 31:807–825.

Landau, Smith, L., and Jones, S. (1997) Object shape, object function, and object name.Journal of Memory and Language 36(1):1–27.

Landau, B., Spelke, E., and Gleitman, H. (1984) Spatial knowledge in a young blind child.Cognition 16:225–260.

Landau, B. and Stecker, D. (1990) Objects and places: Syntactic and geometric representa-tions in early lexical learning. Cognitive Development 5:287–312.

Levinson, S. (1992) Vision, shape, and linguistic description: Tzeltal body-part terminol-ogy and object description. Working paper no. 12, Cognitive Anthropology ResearchGroup, Max Planck Institute for Psycholinguistics.

Locke, J. (1964) An Essay Concerning Human Understanding. Ed. A. D. Woozley. Cleveland:Meridian Books.

Malt, B. and Johnson, E. C. (1992) Do artifact concepts have cores? Journal of Memory andLanguage 31:195–217.

Markman, E. (1989) Categorization in Children: Problems of Induction. Cambridge, MA:Bradford Books/MIT Press.

Markman, E., and Hutchinson, J. (1984) Children’s sensitivity to constraints on wordmeaning: Taxonomic versus thematic relations. Cognitive Psychology 20:121–157.

Marr, D. (1982) Vision. New York: Freeman.Matan, A. (1996) Knowledge of function in young children. Ph.D. thesis, MIT.Medin, D. and Coley, J. (1998) Concepts and categorization. In Handbook of Perception and

Cognition. Perception and Cognition at Century’s End: History, Philosophy, Theory, ed.J. Hochberg and J. E. Cutting. San Diego: Academic Press.

Language and Space 229

Medin, D., Lynch, E., and Coley, J. (1997) Categorization and reasoning among tree ex-perts: Do all roads lead to Rome? Cognitive Psychology 32:49–96.

Medin, D., Goldstone, R., and Gentner, D. (1993) Respects for similarity. PsychologicalReview 100(2): 254–278.

Merriman, W., Scott, P., and Marazita, J. (1993) An appearance-function shift in children’sobject naming. Journal of Child Language 20:101–118.

Naigles, L. (2000) Manipulating the input: Studies in mental verb acquisition. (Chapter 15in this volume.)

Newport, E. (2000) Biological bases of language learning. (Chapter 8 in this volume.)Putnam, H. (1977) Is semantics possible? In Naming, Necessity, and Natural Kinds, ed. S.

Schwartz. Ithaca: Cornell University Press.Quine, W. V. (1960) Word and Object. Cambridge, MA: MIT Press.Quine, W. V. (1969) Natural kinds. Reprinted in S. Schwartz (ed.), Naming, Necessity, and

Natural Kinds, ed. S. Schwartz. Ithaca: Cornell University Press, 1977.Rosch, E., Mervis, C., Gray, W., Johnson, D., and Boyes-Braem, P. (1976) Basic objects in

natural categories. Cognitive Psychology 8:382–439.Shipley, E. (2000) Children’s categorization of objects: The relevance of behavior, surface

appearance, and insides. (Chapter 6 in this volume.)Smith, L.B., Jones, S., and Landau, B. (1992) Count nouns, adjectives, and perceptual

properties in children’s novel word interpretations. Developmental Psychology28:273–286.

Smith, L.B., Jones, S., and Landau, B. (1996) Naming in young children: A dumb atten-tional mechanism? Cognition 60(2): 143–171.

Smith, E. and Medin, D. (1981) Categories and Concepts. Cambridge, MA: HarvardUniversity Press.

Soja, N., Carey, S., and Spelke, E. (1992) Perception, ontology, and word meaning.Cognition 45:101–107.

Subrahmanyam, K., Landau, B., and Gelman, R. (1999) Shape, material, and syntax: Inter-acting forces in children’s learning of novel words for objects and substances.Language and Cognitive Processes 14(3): 249–281.

Waxman, S. R., and Markov, D. (1995) Words as invitations to form categories: Evidencefrom 12- to 13-month old infants. Cognitive Psychology 29(3):257–302.

230 Barbara Landau

Chapter 14

The Psychologist of Avon: Emotion in ElizabethanPsychology and the Plays of Shakespeare

W. Gerrod Parrott

I Introduction

When considering what I should write for this collection of essays, Iknew right away that I wanted somehow to pay tribute to Henry’s in-terest in drama. For, as everyone knows, part of what makes HenryHenry is his interest and skill in acting and directing. And part of whatguides Henry’s approach to psychology are intuitions springing fromhis appreciation of the themes of great drama, from his appreciation ofthe psychological complexity inherent in an actor’s ability to conveycharacter and emotion, and in an audience’s ability to comprehend andvicariously experience a character’s situation all within the frameworkof “as-if.” Part of what makes Henry’s textbook special is its use ofdrama to illustrate psychological principles. And, most important of allfor me, Henry’s interest in drama led to my becoming his student.

After spending my first year at Penn researching a purely cognitivetopic, I found myself being much more interested in human emotion,and I found Henry interested in talking about it. It turned out thatHenry’s interest in drama and mine in emotion overlapped nicely in anarea we called “the quiet emotions,” which included aesthetic emo-tions, humor, and play. After some preliminary experiments on humor,we began the research that formed my dissertation, investigating the in-fant’s game of “peek-a-boo.” In its simple structure of “appearance dis-appearance reappearance,” Henry and I saw a prototype of the sort ofstructure typical of adult’s dramatic narratives: a suspenseful conflictthat is then resolved. Perhaps, we thought, we might understand thedevelopmental roots of drama in this simple game. Henry and I pub-lished our peek-a-boo findings in the journal Cognition and Emotion, andeven that developed into something nice for me: I published other re-search there, became one of the associate editors, and three years ago Itook over as the editor. So, clearly, I thought I should try to pay tributeto Henry’s interest in drama.

But, how to do so? My interests have continued to be focused onhuman emotion, but not on peek-a-boo or drama per se, and, try as I

might, I really cannot relate any of my empirical research to drama. Forexample, I am interested in the emotion of embarrassment, and my ap-proach is rooted in the theory of Erving Goffman, and that emphasizesdramaturgy, but that is too much of a stretch. I also have some researchon mood and memory, showing that people sometimes recall sad mem-ories when happy and happy memories when sad; that is the reverse ofthe usual finding and hence is possibly dramatic, but, it is not aboutdrama. And I have lately been content-analyzing people’s reports of in-tense jealousy, and these accounts are often quite melodramatic, butthat is not right either.

So I cannot pay tribute to Henry by describing research about dramaper se, but I can bring drama into this essay another way. Lately I havebegun to study folk psychologies of emotion, examining the historicaldevelopment of ideas about emotion in Western cultures, and trying tosee how contemporary American conceptions of emotion evolved fromthem. One can track the development of Western conceptions of emo-tion through a multitude of sources, from legal traditions to works offiction, from medical beliefs to academic philosophy, and I have beenlooking to some of these to learn the history of everyday ideas aboutemotion. One of the periods I have found particularly interesting isElizabethan England, and one of the best sources of examples of Eliz-abethan ideas about emotion are the dramatic works of WilliamShakespeare. It is by describing this aspect of my research that I wouldlike to pay tribute to Henry’s interest in drama. My topic for this essaywill be Elizabethan ideas about psychology, particularly about emotion,particularly as they are evidenced in the plays of Shakespeare. In honorof Henry, I shall focus on the Shakespeare part, but I shall also indicatesome of the ways in which the Shakespeare is relevant to the contempo-rary psychology of emotion.

II Psychology in the English Renaissance

In the plays of Shakespeare one finds many expressions of ideas aboutpsychology that were current during the English Renaissance. These ex-pressions reflected shared, everyday conceptions about people’s behav-ior and mental activities, and could be called the Elizabethan folkpsychology. The best documentation of these conceptions is found inthe moral and scholarly writings of the time, which were widely read byShakespeare’s patrons and audiences. A number of Spanish and Frenchworks from the sixteenth century had become available in Englishtranslation by Shakespeare’s time. These books were the Renaissanceequivalent of Henry’s introductory psychology text: influential, pro-found, and selling large numbers for their day.

232 W. Gerrod Parrott

One was Juan Luis Vives’s 1520 Introduction to Wisdom, a spiritual andeducational treatise that was translated into English in 1540 (Fantazzi1979). A more physiological approach could be found in a book by theSpanish physician Juan de Huarte Navarro, whose Examen de Ingenios,written in 1578, was translated into English by Richard Carew in 1594,going into its fourth edition by 1616 (Newbold 1986). Huarte’s popularbook proposed an innate basis for humors and temperament that madecertain passions and careers more suitable for some individuals than forothers. Pierre de la Primaudaye’s The French Academie, written in 1586,was first translated into English by Thomas Bowes in 1594. A moralwork, it discussed the psychology and self-control of emotions in greatdetail and with insight. Just to give the flavor of the book, I shall shareone quotation, retaining original spelling:

Now against the passion of euill Hatred, amongst a great numberof remedies which may very well be applied thereunto, we hauetwo principall ones that are very good and profitable. The firstremedy is, the example of the loue of God. . . . The second remedyis, the contempt of all earthly things. . . . For if we shall set light byall mortall and corruptible things, and lift vp our hearts to higherthings, we shall very easily breake off all hatred and enmity, nei-ther will wee take any thing greatly to heart, but when we see Godoffended. (La Primaudaye 1618, p. 500)

Another moral work from France was Pierre Charron’s Of Wisdome,written in 1608 and first translated into English by Samson Lennardabout five years later. This work discusses the causes and effects of awide range of passions, including envy, jealousy, revenge, fear, sorrow,compassion, choler, and hatred, and the work’s final part considers thevirtue of temperance and methods of controlling the passions (Charron1608).

About this same time, English authors were writing books catering tothe Renaissance interest in psychology and ethics. Through thesebooks, ideas about psychology and emotion from the works of Aristotleand Plato, Hippocrates and Galen, Cicero, Augustine, and Aquinaswere distilled and mixed with English folk notions. Sir Thomas Elyot’sThe Castel of Helthe (1541/1937) was a popular medical book for laymenthat appeared in nearly twenty editions between 1539 and 1610, and itcontained sections on “affectes and passions of the mynde” (Tannen-baum 1937). In 1586 Timothy Bright’s A Treatise of Melancholie an-nounced a deliberate choice to publish in English rather than in Latin sothat “the benefit might be more common”; the success of this rambling,medically oriented book attests to the popular interest in psychology ingeneral and melancholy in particular (Newbold 1986). Thomas

The Psychologist of Avon 233

Wright’s book, The Passions of the Minde in Generall, was first publishedin 1601 and is a much more impressive work. A former Jesuit, Wrightwrote in English so that the wisdom he gleaned from classical sources,both psychological and moral, might help the English people to practicevirtue, achieve self-knowledge and self-control, and use their passionsfor good purposes. The greatest psychological book of the EnglishRenaissance was surely Robert Burton’s Anatomy of Melancholy, but itspublication in 1621 makes it a bit late to have influenced Shakespearedirectly. But, masterful though it was, it was based in large part on theworks already mentioned, and thus illustrates the sorts of ideas thatwere in the air during the time that Shakespeare wrote his plays. Theseideas are certainly what we would call psychology, but they are not onlypsychology: they are intertwined with what we would now call ethics,religion, medicine, philosophy and even astrology. These ideas havepartly shaped our present culture, and one way they did so was by in-filtrating the plays of the Bard of Avon.1

III Emotion in Elizabethan Psychology and in Shakespeare’s Plays

The psychological writings of the Elizabethan period addressed manyaspects of human emotion, from physiology to mental bias and self-regulation. The texts of Shakespeare’s plays suggest that Shakespeareand his audiences knew this psychology well and took it for granted,for the plays contain a wealth of allusions to it. In this section I shall pre-sent a sampling of the Elizabethan psychology of emotion and illustrateits presence in Shakespeare’s plays.

In Elizabethan psychology there still persisted the Aristotelian ideathat there were three types of soul, hierarchically nested: the vegetable(life), the sensible (consisting of life plus feeling, which includes percep-tion, common sense, imagination, instinct, and memory), and the ratio-nal (consisting of life and feeling, plus reason). The treatises circulatingin Elizabethan England maintained that the rational soul of humans op-erated primarily via three organs of the body, each of which was spe-cialized for activities corresponding to the three types of soul: the liverfor the vegetal, the heart for the sensible, and the brain for the rational.The liver and heart were therefore associated with basic biology and theemotions, whereas the brain served rational thought and the will. Thus,in the opening scene of Twelfth Night, we see the Duke of Illyria ask:

[Orsino]:How will she love when the rich golden shaftHath killed the flock of all affections elseThat live in her when liver, brain, and heart,

234 W. Gerrod Parrott

These sovereign thrones, are all supplied, and filledHer sweet perfections with one self king!(Twelfth Night, I, i, 34–38)2

In Renaissance psychology, the liver, spleen, and gall were all thoughtto be related to the emotions. The liver, when supplied with blood, pro-duced courage and love; the gall produced wrath and bitterness; thespleen purged melancholy and thus was linked to mirth. Knowing thisphysiology is a great help in understanding otherwise cryptic passagesin Shakespeare. For example, to cite some passages culled fromAnderson (1927/1966), in Macbeth, the title character speaks to a fearfulservant (earlier described as a “cream-faced loon”) to announce the ap-proach of ten thousand soldiers:

Go prick thy face and over-red thy fear,Thou lily-livered boy. What soldiers, patch?Death of thy soul, those linen cheeks of thineAre counsellors to fear. What soldiers, whey-face?(Macbeth, V, iii, 16–19)

Or in Henry V:

Grey:Those that were your father’s enemiesHave steeped their galls in honey, and do serve youWith hearts created of duty and of zeal.(Henry V, II, ii, 29–31)

Or in Measure for Measure:

Isabella:His glassy essence, like an angry apePlays such fantastic tricks before high heavenAs makes the angels weep, who, with our spleens,Would all themselves laugh mortal.(Measure for Measure, II, ii, 123–126)

The Elizabethans’ psychophysiology has not fared particularly well inlight of modern biology, but their insights about the more mental as-pects of emotion have fared considerably better. Regarding the expres-sion of emotion and the possibility of deception about one’s emotions,Elizabethan psychology asserted that there should be a correspondencebetween the appearance of the body and the state of the soul, an ideathat had its origins in Plato. The ability to conceal emotions was be-lieved to be quite limited, so when a person did not seem to be movedby matters that normally cause shame or guilt or regret, it was assumed

The Psychologist of Avon 235

that the person had learned not to have the emotion, not that the emo-tion was present but not expressed (Anderson 1927/1966). Yet not tohave one of these moral emotions is to become an immoral person, andthere are wonderful passages in Shakespeare expressing this idea. In 3Henry VI, York says to Queen Margaret:

But that thy face is visor-like, unchanging,Made impudent with use of evil deeds,I would essay, proud Queen, to make thee blush.To tell thee whence thou cam’st, of whom derived,Were shame enough to shame thee wert thou not shameless.(3 Henry VI, I, iv, 117–121)

Even better for demonstrating the process of character alteration in per-fecting deception is Macbeth. Early in the play Lady Macbeth beginscoaching her husband:

Your face, my thane, is as a book where menMay read strange matters. To beguile the time,Look like the time; bear welcome in your eye,Your hand, your tongue; look like the innocent flower,But be the serpent under’t.”(Macbeth, I, v, 61–65)

And Macbeth resolves to do it:

False face must hide what the false heart doth know.(Macbeth, I, vii, 82)

By act 5, Macbeth no longer betrays his purposes with his emotions, yetit is not by hiding the emotions that he succeeds, but by no longer hav-ing them:

I have almost forgot the taste of fears.The time has been my senses would have cooledTo hear a night-shriek, and my fell of hairWould at a dismal treatise rouse and stirAs life were in’t. I have supped full with horrors.Direness, familiar to my slaughterous thoughts,Cannot once start me.(Macbeth, V, v, 9–15)

In some respects the Elizabethan view of deception is similar to preva-lent contemporary views, maintaining that emotional deception is pos-sible but imperfect (e.g., Ekman 1985). In one respect it is notablydifferent, however, because the education of the emotions and its moralimplications are not emphasized in contemporary psychology.

236 W. Gerrod Parrott

Another tenet of the Elizabethan psychology of emotion was thatconcrete objects and events can stir passion and action more readilythan can less vivid stimuli. This tenet was a special case of a more gen-eral belief in the dependence of reason and the imagination on informa-tion supplied by the senses (Anderson 1927/1966). Disrupt the input,and the whole system veers off course. Thus we have Oberon streakingTitania’s eyes with juice to alter her perceptions, and in numerous playscharacters are bound and placed in darkness to aid recovery of theirwits. The role of vivid stimuli in producing angry aggression is illus-trated in King John:

King John:Witness against us to damnation!How oft the sight of means to do ill deedsMake deeds ill done! Hadst not thou been by,A fellow by the hand of nature marked,Quoted, and signed to do a deed of shame,This murder had not come into my mind.(King John, IV, iii, 220–224)

And, for the emotion of fear, there is the reaction of Macbeth, who wasterrified by Banquo’s ghost but calmed immediately after his disap-pearance:

Macbeth:Take any shape but that, and my firm nervesShall never tremble. . . .Unreal mock’ry, hence! Exit GhostWhy so, being gone,I am a man again.(Macbeth, III, iv, 101–102, 106–107)

The Elizabethan’s point about the effectiveness of concrete perceptionsin arousing emotions seems quite consistent with some modern ideasabout cognition and action (e.g., production systems), but about emo-tion in particular modern academic psychology is oddly quiet. Perhapsthis is a case where the Elizabethan writings identify a phenomenon un-derappreciated in our time.

The Elizabethan psychologists drew on Aquinas for a sense that emo-tions can perform useful functions by guiding people toward theirgoals and motivating them to overcome frustration or to resign them-selves to irrevocable loss. An emotion such as grief was believed to beexpressible either as angry frustration or sad resignation, and thus therewas an element of choice concerning the direction of emotional im-pulses. Shakespeare, in 3 Henry VI, has Richard resolve not to weepaway his grief but to vent it in revenge:

The Psychologist of Avon 237

I cannot weep, for all my body’s moistureScarce serves to quench my furnace-burning heart;. . .To weep is to make less the depth of grief;Tears, then, for babes blows and revenge for me!(3 Henry VI, II, i, 79–80, 85–86)

King Lear likewise vows not to weep, but he is unable to obtain re-venge. And, on some interpretations at least, Hamlet’s failure to redirecthis grief may be understood as contributing to his inability to seek re-venge (Anderson 1927/1966). This aspect of the Elizabethan psychol-ogy of emotion is particularly consonant with modern psychology,which ever since Darwin has emphasized the adaptive function of emo-tions.

The psychology textbooks of the Renaissance observed that whenpassions become too intense they can bias thinking. In Shakespeare wesee this phenomenon in an exchange between Bushy and the Queen inRichard II.

Bushy:Madam, your majesty is too much sad.. . .Queen:. . . Yet I know no causeWhy I should welcome such a guest as grief.. . .Bushy:Each substance of a grief hath twenty shadowsWhich shows like grief itself but is not so.For sorrow’s eye, glazèd with blinding tears,Divides one thing entire to many objectsLike perspectives, which, rightly gazed upon,Show nothing but confusion; eyed awry,Distinguish form. So your sweet majesty,Looking awry upon your lord’s departure,Find shapes of grief more than himself to wail,Which, looked on as it is, is naught but shadowsOf what it is not. Then, thrice-gracious Queen,More than your lord’s departure weep not: more is not seen,Or if it be, ‘tis with false sorrow’s eye,Which for things true weeps things imaginary.(Richard II, II, ii, 1, 6–8, 14–27)

That passage nicely depicts the bias of sadness; for that of jealousy, thereare Iago’s descriptions of the force of inflamed suspicion:

238 W. Gerrod Parrott

I will in Cassio’s lodging lose this napkin,And let him find it. Trifles light as airAre to the jealous confirmations strongAs proofs of holy Writ.(Othello, III, iii, 325–328)

And later:

As he [Cassio] shall smile, Othello shall go mad;And his unbookish jealousy must consterPoor Cassio’s smiles, gestures, and light behavioursQuite in the wrong.(Othello, IV, i, 99–102)

Emotion’s ability to bias thought has been rediscovered recently. Influ-ential researchers such as Gordon Bower (1981) have reintroduced thephenomenon after long neglect, and it is now an important part of theo-ries of affective disorders, decision making, and memory (Teasdale andBarnard 1993).

To have the passions control reason, to have the body directing themind, is to upset one of the most important Renaissance ideas about theproper order of nature: Reason should govern the body as the king gov-erns the kingdom and God’s laws govern the universe. The need to pre-vent such disorder leads to the final Renaissance topic I would like toconsider: self-control. Self-control appeared as a virtue in Greek writ-ings as early as the sixth century B.C., and by the time of Aeschylus waswell established among the cardinal virtues; it was considered in depthby Plato and continued to evolve with the Stoics of later Greek andRoman culture, when it was incorporated into early Christian doctrine(North 1966). Called sophrosyne by the Greeks and temperantia by theRomans and early Christians, this virtue can be translated variously asself-control, moderation, temperance, or self-knowledge. Shakespeare’sHamlet praised his friend Horatio for possessing just this Stoic virtue:

Since my dear soul was mistress of her choiceAnd could of men distinguish her election,S’hath sealed thee for herself, for thou hast beenAs one, in suff’ring all, that suffers nothing,A man that Fortune’s buffets and rewardsHast ta’en with equal thanks; and blest are thoseWhose blood and judgment are so well commeddledThat they are not a pipe for Fortune’s fingerTo sound what stop she please. Give me that manThat is not passions’ slave, and I will wear himIn my heart’s core, ay, in my heart of heart,

The Psychologist of Avon 239

As I do thee.(Hamlet, III, ii, 61–82)

Later in the same scene Shakespeare has the Player King describe thecharacter of one who has not developed this virtue:

Purpose is but the slave to memory,Of violent birth, but poor validity,Which now like fruit unripe sticks on the tree,But fall unshaken when they mellow be.Most necessary ‘tis that we forgetTo pay ourselves what to ourselves is debt.What to ourselves in passion we propose,The passion ending, doth the purpose lose.The violence of either grief or joyTheir own enactures with themselves destroy:Where joy most revels, grief doth most lament;Grief joys, joy grieves, on slender accident.(Hamlet, III, ii, 179–190)

The modern psychology of emotion strays from this Stoic conception ofself-regulation as virtue. More typical of modern psychology is a lessmoralistic, more hedonistic approach that focuses on the “mainte-nance” of positive emotions and the “repair” of negative emotions (Par-rott 1993a).

IV The Relevance of Folk Psychologies for the Psychology of Emotion

I think there are valid reasons for a twentieth-century psychologist toconsider Renaissance folk psychology and literature. Folk psychologiescan play a valuable role in guiding and evaluating academic psycholo-gies, although probably more with respect to general concepts than todetailed explanations (Fletcher 1995; Greenwood 1991). Regardless ofthe accuracy of the explanations, folk psychologies and literature helpestablish the meaning of everyday concepts. The basic concepts of psy-chology are all folk concepts: memory, attention, perception, emotion,and so on. Academic psychologists may establish new concepts, distin-guished from the everyday concepts named by the same word, but theymust make clear that their findings are not intended to address theeveryday concept. An example might be autonomic arousal, which doesnot quite correspond to any of the everyday meanings of “arousal.”

I would propose that establishing the everyday meaning of “emo-tion,” “emotional,” and the like is an important thing to do. Academicpsychologists have, I believe, developed their own conceptions of emo-tion, conceptions that have begun to stray from everyday conceptions

240 W. Gerrod Parrott

in important ways. As yet, however, the differences have not been madeexplicit or precise. At this point in the development of the psychology ofemotion it would be good to note which aspects of the everyday con-ception are being retained and which abandoned, to clarify the benefitsand costs of the new conceptual framework, and to remind ourselvesthat some aspects of the everyday conception are not addressed by con-temporary research.

Shakespeare’s plays and Renaissance folk psychologies can be usedto evaluate the scope of modern academic theories. By this measure, thesuccess of contemporary emotion research is rather mixed. For certainaspects of Elizabethan psychology contemporary research does an ex-cellent job. Renaissance psychologies clearly included physiological re-actions as part of emotion, and modern neuroscience is a distinctimprovement on Renaissance efforts here (see Gray 1995). Similarly, theRenaissance psychologies delved into the ways in which emotion altersthought, and contemporary research on memory, perception, and judg-ment, and contemporary journals such as Cognition and Emotion, show acorresponding modern interest in just these phenomena (see Fox 1996;Nasby 1996). The Renaissance interest in the purpose of emotionsshows an awareness that emotions can function in adaptive ways, andthis interest is nicely reflected in modern treatments of functionalism(see Oatley 1992; Parrott and Schulkin 1993).

For other aspects of Renaissance psychology, however, contemporaryresearch does not fare so well. It was common in Renaissance psychol-ogy to distinguish between passion and reason, yet there is no corre-sponding distinction in academic approaches to emotion, althoughthere is in contemporary folk theory. When I had my students ask ac-quaintances to recall an example of “being emotional” and to explainwhat made it so, we found that the most common qualities cited were“being irrational” and “being out of control.” Virtually all respondentsconveyed a sense that “being emotional” carried a negative connotation(Parrott 1995). In contrast, modern academic theories of emotion, al-though distinguishing between emotional appraisal and unemotionaldeliberation, tend not take into account the rationality or social appro-priateness of emotional thinking in making these classifications (Parrott1993b).

The Renaissance psychologies clearly linked emotion to ethics andvirtue in ways not considered appropriate for a modern science, yet it isthese connections that made emotion so important to the understand-ing of human nature, and so central to Shakespeare’s plays. The role ofemotion in the development of the character of Macbeth consists not somuch in his concealing or extinguishing emotion as it does in his be-coming evil. The point of emotion for Shakespeare’s King John is not somuch an isolated psychological event as it is as a part of the moral event

The Psychologist of Avon 241

of temptation. Shakespeare was concerned with emotional dysfunctionas well as emotional function, as when Othello’s jealousy overwhelmshis reason. All in all, one finds the emphasis not so much on the natureof emotion per se as on what emotion reveals about character, includingits role in how people come to do wrong: Consider how Iago’s reason-ing is warped by his resentment and jealousy, or how Macbeth begins todo wrong only opportunistically but, through repetition, makes evilpart of his character. In sum, the Renaissance psychology books and theplays of Shakespeare contrast with modern academic psychology inprecisely that quality said to be most characteristic of the Renaissance,its interest in the entire person. By contrast, modern academic psychol-ogy appears excessively modular and mechanical, paying insufficientattention to the social and moral aspects of emotion.

Thus my motive in investigating Shakespeare’s plays and Renais-sance folk psychologies is to remind myself and others of issues thathelp make important a topic such as emotion. Now, where did I learn todo that? I have come to think that some of the most important lessonswere learned at the research seminar Lila and Henry generously con-ducted in their home, which over the years benefited so many of us con-tributing to this volume. One point always emphasized to students byboth Henry and Lila was that good research always maintains its con-nection to the issues that initially established its importance. In the re-search seminar Henry and Lila reminded their students to keep in mindthe larger framework of their research, and discouraged them from pur-suing laboratory phenomena for their own sake or because they were invogue. In this brief tour of emotion in Shakespeare I hope to have bothdemonstrated this lesson and expressed my gratitude for it.

Acknowledgment

I am grateful to John Sabini for his helpful comments on a previousdraft of this essay.

Notes

1. I ought to make clear that there are two ways of researching a topic such as this. One canbecome an authority on Elizabethan culture and on the works of Shakespeare, or onecan avail oneself of the many wonderful expositions of these topics that are available,and my method, in case it is not obvious, is necessarily the latter! So, before going on, Iwould like to acknowledge my indebtedness to the scholars whose writings have madethis material accessible to the likes of me. In particular, I am indebted to books onElizabethan culture and Shakespeare’s plays by such scholars as E. M. W. Tillyard(1944), Theodore Spencer (1949), William Webster Newbold (1986), and especiallyRuth Leila Anderson (1927/1966).

2. All Shakespearian quotations and line references are drawn from the edition ofShakespeare’s complete works by Wells and Taylor (1986).

242 W. Gerrod Parrott

References

Anderson, R. L. (1966) Elizabethan Psychology and Shakespeare’s Plays. New York: Russelland Russell. (First published in 1927).

Bower, G. H. (1981) Mood and memory. American Psychologist 36:129–148.Charron, P. (1608) Of Wisdome (Samson Lennard, trans.). London: E. Blount and W.

Aspley.Ekman, P. (1985) Telling Lies: Clues to Deceit in the Marketplace, Politics, and Marriage. New

York: Norton.Elyot, T. (1937) The Castel of Helthe. New York: Scholars’ Facsimiles and Reprints.

(Original work published 1541.)Fantazzi, C. (1979) In Pseudodialecticos: A Critical Edition. Leiden: Brill.Fletcher, G. (1995) The Scientific Credibility of Folk Psychology. Mahwah, NJ: Erlbaum.Fox, E. (1996) Selective processing of threatening words in anxiety: The role of awareness.

Cognition and Emotion 10:449–480.Gray, J. A. (1995) A model of the limbic system and basal ganglia: Applications to anxiety

and schizophrenia. In M. S. Gazzaniga (ed.), The Cognitive Neurosciences (pp.1165–1176). Cambridge, MA: MIT Press.

Greenwood, J. D. (1991) The Future of Folk Psychology: Intentionality and Cognitive Science.Cambridge: Cambridge University Press.

La Primaudaye, P. de (1618) The French Academie (T. Bowes, trans.). London: ThomasAdams.

Nasby, W. (1996) Moderators of mood-congruent encoding and judgement: Evidence thatelated and depressed moods implicate distinct processes. Cognition and Emotion10:361–377.

Newbold, W. W. (1986) General introduction to W. W. Newbold (ed.), The Passions of theMind in General (pp. 1–50). New York: Garland Publishing.

North, H. (1966) Sophrosyne: Self-Knowledge and Self-Restraint in Greek Literature. Ithaca,NY: Cornell University Press.

Oatley, K. (1992) Best-Laid Schemes: The Psychology of Emotions. New York: CambridgeUniversity Press.

Parrott, W. G. (1993a) Beyond hedonism: Motives for inhibiting good moods and formaintaining bad moods. In Handbook of Mental Control, ed. D. M. Wegner and J. W.Pennebaker, pp. 278–305. Englewood Cliffs, NJ: Prentice-Hall.

Parrott, W. G. (1993b) On the scientific study of angry organisms. In R. S. Wyer, Jr. and T.K. Srull (eds.), Perspectives on Anger and Emotion: Advances in Social Cognition (vol. 6,pp. 167–177). Hillsdale, NJ: Erlbaum.

Parrott, W. G. (1995) The heart and the head: Everyday conceptions of being emotional. InJ. A. Russell, J.-M. Fernández-Dols, A. S. R. Manstead, and J. C. Wellenkamp (eds.),Everyday Conceptions of Emotions: An Introduction to the Psychology, Anthropology,and Linguistics of Emotion (pp. 73–84). Dordrecht: Kluwer.

Parrott, W. G. and Schulkin, J. (1993) Psychophysiology and the cognitive nature of theemotions. Cognition and Emotion 7:43–59.

Spencer, T. (1949) Shakespeare and the Nature of Man. New York: Macmillan.Tannenbaum, S. A. (1937) Introduction. In T. Elyot, The Castel of Helthe (pp. iii–xi). New

York: Scholars’ Facsimiles and Reprints.Teasdale, J. D. and Barnard, P. J. (1993) Affect, Cognition, and Change: Re-Modelling

Depressive Thought. Hove: Lawrence Erlbaum Associates.Tillyard, E. M. W. (1944) The Elizabethan World Picture. New York: Macmillan.Wells, S. and Taylor, G. (eds.). (1986) William Shakespeare: The Complete Works. Oxford:

Clarendon.

The Psychologist of Avon 243

Chapter 15

Manipulating the Input: Studies in Mental VerbAcquisition

Letitia R. Naigles

“It is a truism, or ought to be, that language acquisition dependscrucially on species-specific endowments . . . and at the same timeis the strict outcome of specific learning opportunities” (L. Gleit-man and H. Gleitman, 1997:29).

It was in the service of figuring out just what role the endowmentsplayed and what role the learning opportunities played that Lila andHenry Gleitman developed the deprivation paradigm for the study oflanguage acquisition. That is, if you think some aspect of input (linguis-tic, perceptual) or some aspect of human physiology (hearing, chromo-somes) is important for children’s development of some linguisticstructure or class or knowledge, then take that aspect away and see howthe relevant language has been affected. As this is certainly not a para-digm that can ethically be imposed on any human population, theGleitmans and their collaborators have relied on cases where the rele-vant deprivations occurred naturally, and have generated tremendousfindings showing, for example, the resiliency of the sociable human tocreate his or her own language even in the absence of linguistic input,and the resiliency of the language to emerge even in humans missingsome critical biological components (e.g., Feldman, Goldin-Meadow,and Gleitman 1975; Landau, Gleitman, and Spelke 1981; Landau andGleitman 1985; Fowler, Gelman, and Gleitman 1994; see also Newport,Goldin-Meadow, Landau, this volume).

These studies, performed by the Gleitmans with my predecessors atPenn, were the “milk” of my graduate school days. Among many otherthings, they showed me how much could be learned about languageand language acquisition from such innovative manipulations of inputor endowment. Recently, I have embarked upon a new line of research,concerning children’s acquisition of mental state verbs, which hasseemed tailor-made for an application of the deprivation paradigm (al-beit in a less dramatic fashion than these earlier studies), and I havethus begun to do some input manipulation of my own. It seems only

fitting, then, that the first report of this research should be the topic ofmy chapter honoring the enormous contribution of Henry and LilaGleitman to my scholarly life.

In some ways, early research on mental verb acquisition was the an-tithesis of the early research on syntax acquisition that the Gleitmansand their colleagues were departing from. That is, one school held thatearly syntax acquisition was directed via the input of “Motherese” (e.g.,Snow and Ferguson 1977); the Gleitmans’ work showed how much ofthis was actually directed by the children themselves (e.g., Newport,Gleitman, and Gleitman 1977). In contrast, input has played an aston-ishingly small role in theorizing about children’s developing mentalstates and mental verb understanding (but see Moore et al. 1994). Mosttheories have targeted aspects of children’s cognitive or emotional mat-uration as the prime instigating factors (e.g., Leslie 1991; Olson andAstington 1988; Wellman 1990). But too much emphasis on matura-tional change in children’s development of mental language and statesmay ultimately be just as obscuring as too much emphasis onMotherese input had been in the acquisition of syntax. My argument inthis chapter will be that input has indeed a critical role to play at a criti-cal transition point in children’s mental verb acquisition. But first, somebackground.

A. Why Care about Mental State Verbs? Let Me Count the Ways

Research on mental state verbs (MSVs) and their acquisition has grownexponentially over the past thirty years or so. Rationales for this re-search vary widely, depending at least in part on whether researcherscome from a linguistic tradition or a cognitive psychological one, andwhether the focus is on adults’ knowledge of mental terms or on chil-dren’s acquisition of them. With my bent toward language and cogni-tive development, I have found the following four rationales mostcompelling. (a) Mental verbs epitomize Quine’s (1960) problem of radi-cal translation for the child learner, as thinking and knowing, for exam-ple, are never ostensively available. (b) Mental verbs are notoriouslypolysemous, in that each is associated with multiple senses. Consider,for example, that different senses of know are conveyed by the followingsentences:

(1) I know that song. (recognize)

(2) I don’t know if that’s gonna come out too well. (conjecture)

(3) I don’t know what you are saying. (understand)

246 Letitia R. Naigles

(4) I know you like that book. (believe)

(5) You let me know if you want Mom to help you. (tell)

(6) You know, the keys are over there. (shared information)

(c) Mental verbs experience a long and gradual period of developmentduring child language acquisition, apparently unlike that which occurswith concrete nouns and verbs. (d) Mental verbs provide insight intohuman cognition, as they can reveal our access to our own internalstates and our notions about the internal states of others (i.e., a folk the-ory of mind). In what follows, I discuss these four rationales in terms ofmy ultimate goal, which is determining how mental verbs in general,and think, know, and guess in particular, are acquired.

1. A theoretical issue in the acquisition of mental verbsEven though all lexical words are subject to the Quinean problem ofradical translation in child language acquisition, mental verbs such asthink and know must be especially challenging in this respect. Whereasthe meanings of verbs such as jump and cry must be at least sometimesmanifested in the ostensive context (i.e., jumping and crying are some-times going on when “jump” and “cry” are uttered), it is hard to imag-ine how the meanings of think and know could ever be ostensivelyavailable. These verbs refer to mental states and processes, which are bydefinition abstract and removed from purely sensory experience (seeScholnick 1987; Gleitman 1990 for more discussion). Even in the mostexplicit cases, thinking is just not perceivable. Imagine a child observingher mother standing in the middle of the living room, eyes darting backand forth, head wagging to and fro. Such a child might be moved to ask,“watcha doin’ mom?” only to be told “I’m thinking about where to putthe new couch when it arrives.” Does this child conjecture that thinkmeans rapid eye and head movements? What if the mother was alsopointing to various places around the room—would these points alsobe incorporated into the child’s meaning of think? No, somehow she fig-ures out that think refers to the internal process that prompted thepoints and the rapid eye and head movements.

Gleitman (1990; Landau and Gleitman 1985) has suggested that thepresence of sentence complements with MSVs (in the above sentence,where to put the new couch when it arrives) provides children with an im-portant clue that the verb refers to mental states or processes. Inessence, the sentence complement instantiates the proposition to whichthe MSV pertains (see also Fisher, Gleitman, and Gleitman 1991).Moreover, studies of maternal speech to young children have shown

Manipulating the Input 247

that mental verbs appear with sentence complements more often thanmotion verbs do, and that sentence complements are more likely to fol-low mental verbs than motion verbs (Naigles and Hoff-Ginsberg 1995).At this point, there is little experimental evidence that children actuallyuse this structural clue in determining that a verb refers to a mentalstate rather than a physically available one. However, learning thatthink (and know, guess, wonder, believe, etc.) refers to an mental state orprocess is actually only half the battle. As discussed in more detailbelow, MSVs also have to be distinguished from each other (what if themother above had said “I now know where to put the couch when it ar-rives”?) and sometimes even from themselves, as none of the most com-mon MSVs has a single, unitary meaning (e.g., know, (1) to (6) above).Clearly, ostention is of even less use for this part of the process. One ofmy goals in this chapter is to suggest another source of information forchildren’s acquisition of distinctions between MSVs.

2. The polysemy of mental verbsMental verbs are acknowledged polysemists. Moreover, the senses as-sociated with the verbs think, know, and guess are particularly numerousand varied. For example, most uses of these three verbs involve mentalstates or processes, but some appear to serve primarily conversationalpurposes (e.g., Y’know what? I think it’s time for your nap and the rhetor-ical Guess who I saw today!). Within the mental domain, the three verbsdiffer on a variety of dimensions, both continuous (e.g., certainty) anddiscrete (e.g., factivity, process/product). In what follows, I discussfirst the linguistic-theoretic and then cognitive-psychological experi-mental approaches to understanding the complexity of MSV meaningand representation. These approaches have tended to operate in igno-rance of each other’s work; my hope in presenting both of them is that,following Lila and Henry’s example, it can be seen that combining thelinguistic and psychological traditions is an enormously fruitful en-deavor.

In the linguistics tradition, the polysemy of mental verbs has beenpartially captured by the participation (or lack thereof) of each verb indiffering grammatical or discourse structures. That is, the structural dif-ferences are often used as diagnostics for semantic or pragmatic differ-ences. Here, we briefly discuss three cases in which distinct semantic orpragmatic aspects of individual verbs are illuminated by considerationof structural differences between verbs. For example, take the factivitydimension. Know is considered to be a factive because it presupposesthe truth of its complement, whereas think is nonfactive and allows nosuch presupposition. The syntactic phenomenon of “neg-raising”1 hasbeen proposed as one diagnostic of the absence of factivity, and indeed,

248 Letitia R. Naigles

think allows neg-raising much more freely than know. That is, the sen-tence B doesn’t think it is raining outside is equivalent in many ways to Bthinks it isn’t raining outside: It’s not B’s thinking that is being negated,but the conditions outside. In contrast, G doesn’t know it is raining outsideis not equivalent to G knows it isn’t raining outside: In fact, given know’sfactivity it must be raining outside (for more discussion, see Kiparskyand Kiparsky 1970; Kempson 1975; Horn 1978; Hooper 1975).

Similarly, a process/product-type dimension of MSVs has been asso-ciated with distinct structures, both morphological and syntactic. Forexample, the morphological inflection “-ing,” which can appear freelywith think but only in restricted contexts with know (Beverly is thinkingabout/*knowing the animal puzzle), has been linked to the processing,cogitating sense of think. And appearance with a direct object, which ispossible with know but not think (Gregory knows/*thinks that song!), hasbeen related to the product sense of know, that which captures the ac-complishment of knowledge (Dixon 1990; Wierzbicka 1988).

Finally, MSVs that appear in the discourse structure of parentheticals(e.g., Beverly went to the store, I think and Your house, I know, is very old)gain the additional sense of indicating the speaker’s attitude toward thestatement in the subordinate clause. The three verbs under considera-tion each specify different modulations about that statement (cf. Shatzet al. 1983): Think indicates a rather uncertain attitude, a belief foundedon relatively weak evidence, whereas know signals a certain attitudeand a well-founded belief, and guess refers to a highly uncertain atti-tude and a belief with little if any foundation (Urmson 1963; Hooper1975; Lysvag 1975; Moore and Furrow 1991). In sum, the structural dif-ferences between think, know, and guess each reveal distinct componentsof meaning, thus shedding light on the polysemy of each verb: Know in-cludes the notions of factivity, accomplishment of knowledge, and cer-tainty of attitude. Think and guess include the notions of nonfactivityand processing or accessing of information, and also implicate varyingdegrees of uncertainty of speaker attitude.

The cognitive psychological tradition has approached the polysemyof MSVs somewhat differently. Here, the methodological emphasis hasbeen empirical rather than analytic, and the theoretical focus more oncontexts rather than forms of use. For example, Hall and his colleagues(Hall, Nagy, and Linn 1984; Hall and Nagy 1987; Hall, Scholnick, andHughes 1987; Frank and Hall 1991; Booth and Hall 1995) have postu-lated distinctions both within and between MSVs along a continuum ofinternal processing. For example, they distinguish between knowingand thinking as perceptual experiences (e.g., I know his shirt is red/I thinkit burst), as cognitive products (e.g., I know that tune/I thought of the num-ber; I know why he did that/I thought of how to do it), and as metacognitive

Manipulating the Input 249

or evaluative processes (e.g., I know that Charlie is happier now/I think thisidea is better; I would like to know more than I do/Thinking can be hard work)(examples from Frank and Hall 1991, pp. 531–532). Evidence for thecontinuum is primarily developmental in nature, as children have beenshown to understand the perceptual and cognitive aspects of know be-fore its evaluative and metacognitive aspects (Booth and Hall 1995; seealso Richards 1982). Morever, Frank and Hall (1991) found that adults’spontaneous speech emphasized different aspects of the meanings ofthink and know: Think was most often used in an evaluative sense,whereas know’s usage was best captured by the perceptual and cogni-tive senses.

The studies of Schwanenflugel and her colleagues (Schwanenflugel,Fabricus, Noyes, Bigler, and Alexander 1994; 1996) provide a nice exam-ple of how empirical studies in the cognitive psychological traditioncan illuminate the same distinctions highlighted by linguists’ analyses.Schwanenflugel et al. (1994) gave adult subjects an intensional task, inwhich they were to judge the similarity of meanings of pairs of mentalverbs, and an extensional task, in which they were given scenarios andasked to select any number of mental verbs that could apply to them.The judgments and selections were subjected to multidimensional scal-ing and hierarchical clustering analyses, and several orthogonal dimen-sions emerged. One dimension appeared to reflect the degree ofcertainty of the verb, as know and guess appeared as polar oppositeswith think situated in between them. Another dimension appeared toreflect the creativity of the mental process, in that discover and invent(highly creative processes) and guess and hear (minimally creativeprocesses) were maximally distinguished. Think, know, and guess didnot differ among themselves on this dimension. Furthermore, a com-plex information-processing dimension also emerged, in which percep-tual verbs first were contrasted with conceptual ones, and think, know,and guess all clustered together as conceptual verbs. The more detailedhierarchical analysis yielded a hint of how know might differ from thinkand guess on information processing: Know, learn, and understand weregrouped together as part of a hypothesized memory component, andthink, guess, reason, and estimate emerged as a cluster related to a con-structive processing component.

In sum, this review of the linguistic and cognitive psychological tra-ditions concerning MSVs leads to the prediction that children’s sourcesof information for the meanings of these verbs are to be found in boththe forms and the contexts of MSV use. Moreover, some components ofmeaning—degree of certainty and type of information processing(process vs. product)—seem more central than others, insofar as theyhave emerged in both traditions with their very different methods and

250 Letitia R. Naigles

purposes. It is not surprising, then, that these dimensions, and particu-larily certainty, have been the primary focus of questions concerningMSV acquisition.

3. The developmental trajectory of mental verb understandingGiven the rampant polysemy described above, it is perhaps not surpris-ing that children’s acquisition of mental verbs encompasses such a longperiod of development, extending from age two until well into the ele-mentary school years. In brief, children’s understanding of MSVs ap-pears to begin with conversational senses, then extends to mentalsenses that are relatively undifferentiated, and then progresses to themore sophisticated senses distinguishing between the MSVs. For exam-ple, think, know, and guess typically begin to appear in children’s spon-taneous speech between two and three years of age (Shatz, Wellman,and Silber 1983; Limber 1973; Bretherton and Beeghley 1982); however,these early uses seem more limited than adult uses. Shatz et al. (1983)tracked children’s uses of mental state verbs over time, and found thatthe first uses, early in the third year, typically served conversationalfunctions rather than mental functions (e.g., “know what?” or “I don’tknow”). By three years of age, mental state uses of these verbs (e.g.,“She doesn’t know all this”) become more frequent, as do verb uses thatcontrast real and mental states (e.g., “I was teasing you; I was pretend-ing ’cept you didn’t know that”). Analysis of three-year-olds’ produc-tion of verbs such as think and know, then, indicates that they haveacquired the mental aspects of these verbs, and have distinguishedthem from verbs that refer to physical or affective states (see alsoWellman and Estes 1986, 1987).

However, three-year-olds have yet to learn how the different MSVsare distinguished among themselves, both semantically and pragmati-cally. As numerous researchers have demonstrated, three-year-olds donot distinguish think, know, and guess according to differences on eitherthe factivity dimension (Johnson and Maratsos 1977; Miscione et al.1978; Hopmann and Maratsos 1978; Abbeduto and Rosenberg 1985) orthe certainty dimension (Moore et al. 1989; Moore and Davidge 1989).2

In contrast, it seems that four-year-olds are beginning to make these dis-tinctions. They typically perform above chance, although not error-lessly, on tasks that ask them to distinguish between knowing thatsomething must be true and thinking that it might be true or it might befalse (e.g., Moore et al. 1989; Abbeduto and Rosenberg 1985). Further-more, Frank and Hall (1991) have found that four-year-olds typicallyuse the verbs think and know distinctively in their spontaneous speech.For example, their modal use of think is evaluative (e.g., I think this idea

Manipulating the Input 251

is better) whereas their modal use of know is perceptual (e.g., I know hisshirt is red).

Young grade-schoolers appear to be close to mastery on distinguish-ing the most common MSVs on the certainty and/or factivity dimen-sions (see note 2); for example, they consistently restrict uses of guess toinstances where no evidence was provided about the location of a hid-den object, and rely on the clues of a puppet who said he knew where ob-jects were hidden over the clues of puppets who say they guessed orthought where objects were hidden (Miscione et al. 1978; Moore et al.1989). Very recently, Schwanenflugel et al. (1996) have provided evi-dence that nine-year-olds include both the certainty and information-processing dimensions in their organization of MSVs, and Booth andHall (1995) have demonstrated that grade-schoolers have begun to dis-tinguish between some of the polysemous meanings of know (i.e., be-tween knowing that a tree house wall is broken, knowing what the wallused to look like, and knowing how to fix it).

Such a long developmental trajectory for children’s acquisition ofthink and know may be contrasted with the much shorter trajectory as-sociated with such equally common but more concrete verbs as jumpand cry, which hold apparently adultlike meanings in the lexicons ofthree-year-olds (Clark 1993). In this chapter, I will argue that the longerdevelopmental trajectory of think and know results not only becausethese verbs are more abstract and more polysemous than jump and cry,but also because the input provided for think and know is more confus-ing—at least initially—than that provided for the concrete action verbs.However, before turning attention to children’s input, one more ratio-nale for the study of mental verbs must be discussed.

4. Mental verbs and theories of mindMental verbs have also been of interest because they can provide cluesto people’s mental activity and to their conceptual and logical represen-tations. Historically, developmental psychologists were the first tostudy mental verbs from this perspective. Their concern was to discoverwhen children could differentiate opinion from fact, when children’segocentrism had receded sufficiently to allow distinctions betweentheir mental state and another’s, and when children began to have ac-cess to their internal processes or psychological experiences (Johnsonand Wellman 1980; Miscione et al. 1978; Shatz et al. 1983; Wellman andEstes 1987).

More recently, MSVs have been studied in the context of children’sdeveloping theory of mind (TOM). The classic definition of a TOM isthe notion that other people have minds and intentions and, crucially,

252 Letitia R. Naigles

that the contents of these other minds and intentions can differ fromone’s own and from reality. Tasks employing contrasting mental stateterms have provided the primary diagnostic for the existence of a TOMin four-year-olds: If children can contrast what person A thinks aboutthe world from what person B knows to be true, then they are capable ofholding a false belief, a representation that is different from reality (e.g.,Wimmer and Perner 1983). It has also been noted that the developmen-tal courses of TOM and early MSV acquisition appear to proceed in par-allel. In broad brush, three-year-olds perform at chance on most TOMtasks (unexpected change of location, unexpected contents; Hogrefe,Wimmer, and Perner 1986), while four-year-olds perform above chanceand five-year-olds are essentially perfect.3 This developmental coursehas been remarkably resistant to alteration; for example, attempts to ex-plicitly instruct three-year-olds on how thoughts may be in conflict withreality have consistently met with failure (e.g., Sullivan and Winner1991; Wimmer and Hartl 1991). And as mentioned earlier, children’searly understanding of the certainty and/or factivity distinctions be-tween think, know, and guess appears to proceed along a similar course(Johnson and Maratsos 1977; Moore et al. 1989). Moreover, MSVs andTOM have been found to correlate with each other in development:When Moore and Furrow (1991) gave preschoolers a variety of TOMtasks as well as a task tapping the certainty distinction between think orguess and know, they found a significant correlation between the twotypes of tasks. That is, the children who passed the unexpected-contentsand the unexpected-change-of-location TOM tasks tended to be thesame children who performed above chance on the think/know andguess/know distinctions.

A final parallel that has been noted between MSVs and TOM is repre-sentational. That is, MSVs must include representations at two inde-pendent levels: If I say I think it is raining outside, the truth of theembedded clause is independent of the truth of the sentence as a whole,which is based on think. Thus it may not in fact be raining, yet I can stillthink that it is. So a complete understanding of this sentence requires anunderstanding of the independence of the two clauses. (It’s not the casethat all verbs that take embedded complements have this requirement:If I saw that it was raining outside, then both clauses of the sentencemust be true—it is raining outside and my seeing this—in order for thesentence as a whole to be true.) The same independence-of-levels holdsfor TOM: For children to understand that someone else (erroneously)represents an object in location A while they (correctly) represent thatsame object in location B, two propositions with contradictory truth val-ues must be represented. Some recent data have suggested that children’s

Manipulating the Input 253

passage of TOM tasks is correlated with their mastery of the structure ofMSV embedded complements (see deVilliers 1994, 1995; Tager-Flusberg1993 for more discussion).

5. SummaryIn this section I have discussed how mental verbs provide a challenge tochildren’s acquisition because their processes are invisible (i.e., mental),because each mental verb shares aspects of its meaning with other men-tal verbs yet is also distinct, and because mental verbs are themselvesextremely polysemous. Moreover, the close empirical relation found be-tween MSVs and TOM development suggests that children’s transitionfrom realizing that think and know refer to mental objects to understand-ing how think and know differ is akin to their transition from realizingthat thoughts exist abstractly to appreciating that thoughts may be inconflict with reality. All of these factors undoubtedly contribute to themental verbs’ long period of acquisition; however, none provides an ex-planation for how the acquisition is ultimately accomplished. In thenext section, I consider another factor that might play a more explana-tory role in the acquisition of MSVs.

B. One Hypothesis for MSV Acquisition

So how ARE mental verbs acquired? Clearly, any theory of acquisitionmust require children to pay attention to both the forms and contexts ofmental verb use, as these are what help distinguish the verbs in adultlexicons. And, indeed, Hall and Nagy (1987) suggest that adults’ ex-plicit use of mental verbs in familiar contexts is what helps draw chil-dren’s attention to the mental processes underlying the verbs.However, as Scholnick (1987) points out, there is as yet no coherent the-ory of how children actually acquire the mental verbs. To be sure, I donot claim to have a well-fleshed-out theory of mental verb acquisitioneither. My goal is more modest; namely, to provide an explanation foran early transition children make in mental verb acquisition. This tran-sition typically occurs around the age of four, when children first distin-guish think and guess from know on the certainty and/or factivitydimensions.

Early explanations for this shift focused on children’s cognitive de-velopment as the instigating factor. It has been suggested, for example,that before the age of four children are unable to distinguish uncertainfrom certain situations (Miscione et al. 1978; Johnson and Wellman1980) and only understand about people’s differing desires and nottheir differing thoughts or beliefs (Wellman 1990; Leslie 1991). More re-cently, however, it has been pointed out that children, especially when

254 Letitia R. Naigles

very young, simply may not be hearing the verbs in the usages neededto make the appropriate distinctions.4 For example, Furrow, Moore,Davidge, and Chiasson (1992) coded maternal MSV use to two- andthree-year-olds and found that almost 75% of their utterances contain-ing think were conversational and served to direct the dyadic interac-tion (e.g., “don’t you think the block should go in here?”). Only 5% ofthink utterances instantiated a true mental state reference, and less than1% of think or know utterances were relevant to the notion of uncer-tainty. In fact, many early parental usages of think may (unintentionally,I am sure) implicate exactly the wrong end of the certainty dimension.For example, if a parent says, “I think it’s time for your nap,” this is notusually intended to convey uncertainty about the temporal situationvis-a-vis the child’s nap. On the contrary, it actually means that it is timefor the child’s nap, and she had better get to bed. Furrow et al. (1992)would probably code this usage of think as directing the interaction, butnotice that from the child’s point of view “I think” in this context couldalso be interpreted as meaning I am certain.

A different picture of input emerges when children are older. Frankand Hall (1991) studied adult (both parent and preschool teacher) utter-ances containing think and know in conversation with 4.5-year-olds, andfound that think was primarily used in its evaluative sense, whereasknow was primarily used in its perceptual and cognitive senses. Thusnot only are think and know now distinguished on semantic (and proba-bly pragmatic) grounds, but it is also likely that many of Frank andHall’s evaluative uses of think highlighted its uncertain sense (seeScholnick 1987). In sum, adults typically use think in its conversationalsense when speaking to very young children, but apparently shift thisusage as the children mature, so that think typically manifests its cer-tainty sense in speech to five-year-olds. So here’s the question: Mightchildren’s change in mental verb understanding between the ages ofthree and five years be linked to this change in input they are experienc-ing?

A first step in investigating this question would be to demonstratethat preschool-aged children were indeed sensitive to the ways (i.e.,senses) that mental verbs are used. The only study performed thus farthat has linked parental input and child mental verb understanding isthat of Moore, Furrow, Chiasson, and Paquin (1994), who found a posi-tive correlation between the sheer frequency of maternal belief-term use(i.e., think, know, and guess combined) when children were two years ofage and those same children’s success at distinguishing the three verbsin a comprehension task when they were four. Unfortunately, Moore etal. did not investigate any relationship between the various uses ofthink, know, and guess in maternal speech and children’s subsequent

Manipulating the Input 255

performance on mental verb comprehension tasks. How might I showthat children are sensitive to the ways these verbs are used? Becausethese verbs are attested (as opposed to nonsense) words, I could notcompletely control the type of mental verb usage children heard, andthere was no “natural” case I knew of where parents continued to use arestricted set of MSV senses in speech to their children. However, a vari-ation on the Gleitmans’ deprivation paradigm suggested itself: Ratherthan deprive children of specific types of usage, my colleagues and I(Naigles, Singer, Singer, Jean-Louis, Sells, and Rosen 1995) sought to en-hance them, via the use of television input. Our idea was to provide ad-ditional MSV tokens within the context of a television show, but restrictthe senses in which these verbs were used, and see if this additionalinput affected children’s MSV understanding (at least in the shortterm).

Study 1: Does “Barney and Friends” influence mental verb understanding?While no one would claim that television input, even in these days oframpant television-watching, provides sufficient linguistic input forchildren to learn everything about a language, there is some suggestiveevidence that some forms of television input have the potential to influ-ence young children’s vocabulary development. For example, Rice et al.(1990) found that the amount of “Sesame Street”-watching by childrenfrom age three to age five was a positive predictor of growth in PeabodyPicture Vocabulary Test-Revised (PPVT) scores over the two-year pe-riod. Moreover, Rice and Woodsmall (1988) found significant gains inpreschoolers’ understanding of low-frequency nouns and adjectivesafter they watched short animated film clips from a children’s cablechannel whose voice-over narration included those words. In sum, re-cent research suggests that contemporary television that has been de-signed for children has a significant effect on their overall vocabularydevelopment.

For this study of television input, we chose the show “Barney andFriends.” Earlier research by my colleagues had confirmed the popularperception that this show is extremely engaging to preschoolers; there-fore, the episodes could be counted on to keep the children’s attention(Singer, Singer, Sells, and Rosen 1995). Moreover, Singer and Singer(1997) found that preschoolers who watched specific episodes of“Barney and Friends” showed significant gains in the number of nouns(all used in the episodes) they could define while those who had notwatched the episodes showed no change. Thus we can conclude thatthe children were attending to at least some of the linguistic content ofthe episodes. Furthermore, Singer and Singer (1997) had already per-

256 Letitia R. Naigles

formed detailed codings of the social and cognitive content of 48“Barney and Friends” episodes, and their ten top-ranked episodes werefound to include numerous uses of our three target verbs, think (63 to-kens), know (106 tokens), and guess (22 tokens).

Our goal in this study was to see if providing “extra” input for theMSVs think, know, and guess would influence children’s understandingof these verbs. We assessed children’s current stage of mastery of thecertainty distinction between these verbs and then had half of the chil-dren watch these “Barney and Friends” episodes over the course of twoweeks. After the two week period, each child’s MSV understanding wasassessed again. A pure frequency account, à la Moore et al. (1994),would yield the prediction that the children who were exposed to theseten episodes of “Barney and Friends” would perform better on mentalverb comprehension tasks after exposure than before, and also betterthan the comparison group of children who received no special expo-sure. This is because simply hearing these verbs more frequently shouldpromote children’s better understanding. However, an account basedon the ways in which the verbs were used might yield a different predic-tion. When we coded the utterances containing the three verbs into thefollowing five categories:

(a) Certainty (e.g., “I know that I’m part of my neighborhood”)

(b) Uncertainty (e.g., “I think I’ve seen this napkin somewherebefore”)

(c) Opinion (e.g., “I think your African clothes are pretty”)

(d) Process (e.g., “You could think of a number for a guessinggame”)

(e) Accomplishment (e.g., “I know that song”)

we found that uncertain uses of “think” and “guess” were rare, com-prising only 8–9% of all tokens. In contrast, certain uses were muchmore prevalent, comprising almost 43% of “think” tokens, 27% of“guess” tokens, and 32% of “know” tokens (most of the other utter-ances containing these verbs invoked their process or accomplishmentsenses). Thus “think” and “guess” were used three to five times morefrequently in certain contexts than in uncertain ones. Moreover, the per-cent of certain utterances was roughly equivalent for all three verbs:“know” was not distinguished from “think” or “guess” by appearingmore often as pragmatically certain. If children are sensitive to the waysin which the verbs are used, rather than just their frequency, then thechildren watching these episodes of “Barney and Friends” might come

Manipulating the Input 257

away with the notion that all three mental verbs refer to certain mentalstates, in which case the children’s subsequent performance on a men-tal-verb-distinction task would not be expected to improve.

Method

Participants The final sample included 39 three-, four-, and five-year-oldchildren drawn from three local preschools. All of the children were na-tive speakers of American English; all but six were of European-American heritage. Twenty-two participated in the “Barney”-watchinggroup (10 boys, 12 girls; MA = 47.73 months [SD = 6.18]) and seventeenin the nonwatching group (8 boys, 11 girls; MA = 49.35 months [SD =7.57]). Because of their failure to reach criterion on the practice trials(see below), an additional 11 children were tested but then eliminated.

Materials, design, procedure Moore et al.’s (1989) assessment of MSV un-derstanding was used (see also Moore and Davidge 1989; Moore andFurrow 1991; Moore et al. 1994). The materials included two smallboxes, one blue and one white; two novel hand puppets, named Jazzand George; and one small toy. Experimenters told the children the fol-lowing: “We are going to play a hiding game. When you close youreyes, I will hide the toy in either the white or the blue box and you haveto find it. Lucky for you, Jazz and George will watch me hide it so theycan help you to find the toy. So if you want to find the toy, you need tolisten carefully to what Jazz and George tell you.”

During the practice trials, the puppets distinguished the boxes via theuse of the negative; that is, Jazz says, “It’s in the blue box” and Georgesays “It’s not in the white box.” When the children chose the correct boxduring these trials they were praised and given stickers, and if theychose the incorrect box they were corrected. To reach criterion, the chil-dren had to be correct on three practice trials in a row (out of six). Asmentioned earlier, 11 children did not reach criterion during the pretest,posttest, or both.

Once the practice trials were successfully completed, the test trialscommenced in much the same format. Here, the puppets distinguishedthe boxes on the basis of the verbs think, know, and guess. That is, if thetoy was in the white box, Jazz might say “I think it’s in the blue box,”while George would say “I know it’s in the white box.” Care was takennot to unduly emphasize the mental verbs; the experimenters main-tained an even prosody throughout each utterance. Then the experi-menter would ask, “Where is the toy?” Notice that in these test trials(and unlike the practice trials), the two puppets’ clues were at oddswith each other, so the children’s task was to determine which was the

258 Letitia R. Naigles

correct box. The children were not told whether or not they were correctafter each test trial; this was necessary to prevent the children from re-ceiving direct feedback as to the correctness of their choices throughoutthe session. When the test trials were completed, each child wasthanked for his or her participation and given some colorful stickers.

Each child received twelve test trials in which two of the three verbswere contrasted; thus there were four presentations of each verb con-trast (think/know, guess/know, think/guess). The particular puppet thatmade each statement, the order in which the puppets made their state-ments, and the box to which each referred were randomly variedthroughout all trials. The trials were videotaped and then coded fromthe videos.

The think/know and guess/know trials were coded for correctness. Thecorrect response was to choose the box referred to by the puppet whosaid “I know.” The think/guess trials were not coded because Moore et al.(1989; see also Moore and Davidge 1989) had found that even eight-year-olds did not distinguish these verbs, and in fact, it is not obviouswhich should be considered more certain (cf. Furrow and Moore 1991;Schwanenflugel et al. 1994, 1996).

Results and discussionOur first analysis compared the children’s percent of correct responsesdistinguishing think and guess from know for each age (three and fouryears), group (“Barney”-watchers and nonwatchers), and time (pretestand posttest). The results are shown in table 15.1. As the table shows,three-year-olds tended to perform more poorly than the four-year-oldsduring the pretest; across verb pairs, the three-year-olds chose correctly59.7% of the time whereas the four-year-olds chose correctly 64.7% ofthe time. These scores are comparable to, albeit a bit lower than, thosegenerated by the preschool-aged children of Moore et al. (1989).

Did watching “Barney” (or not) affect the children’s responses? A four-way repeated-measures ANOVA was performed, in which thebetween-subjects variables were age (three vs. four years) and group(watchers versus nonwatchers), and the within-subjects variables weretime (pretest vs. posttest) and verb pair (think/know vs. guess/know).Because of our substantial subject loss (often resulting in fewer than ten children per cell) and the exploratory nature of this study, we choseto designate an alpha level of 0.10 as our boundary of significance.Only the three-way interaction of age, group, and time reached signifi-cance (F(1,35) = 5.12, p < 0.05). Planned contrasts were performed foreach age and group from pretest to posttest, collapsing across verbpair; the results are highlighted in the two graphs in figure 15.1. The

Manipulating the Input 259

top panel shows that the three-year-old children in either group changedlittle from pretest to posttest, but the bottom panel shows somewhatgreater change within the four-year-olds. In essence, the watchers’ scoresworsened while the nonwatchers’ scores improved. However, only theplanned contrast involving the nonwatchers group was significant (t(8) =1.96, p < 0.10).

At the very least, these analyses suggest that watching these tenepisodes of “Barney” provided no enhancement to our child partici-pants, while not watching “Barney” facilitated those children’s im-proved mental verb understanding.5 However, the absence of an effectof watching “Barney” could be attributable to either of two factors:Either there really was no consistent effect, in that some children im-proved, some worsened, and some showed no change, or there reallywas a consistent effect, but it was fairly small and required a morehighly powered sample to reveal itself statistically. To distinguish thesepossibilities, we performed a second analysis of the data in which thenumber of children whose scores improved, worsened, or stayed thesame from pretest to posttest was tabulated. Because the previousanalysis found no difference between the verb pairs, the children’sthink/know and guess/know scores were averaged in this second analysis.The results are shown in figure 15.2.

260 Letitia R. Naigles

Table 15.1Mean percent correct (SD) on mental verb comprehension task

Age Group Time (n) Think/Know Guess/Know Both verbs

Three Watchers Pretest (10) 60.00 55.00 57.50(16.58) (15.00) (8.29)

Post-test (10) 70.00 52.50 61.25(24.49) (28.39) (19.72)

Nonwatchers Prestest (8) 64.63 60.38 62.50(17.06) (23.53) (16.54)

Post-test (8) 59.38 46.88 53.13(30.46) (29.15) (21.42)

Four Watchers Pretest (12) 68.75 75.00 71.88(27.24) (17.68) (19.18)

Post-test (12) 64.58 68.75 66.67(21.55) (29.09) (21.85)

Nonwatchers Pretest (9) 55.56 54.67 55.11(30.68) (32.43) (27.85)

Post-test (9) 72.22 67.56 69.89(18.43) (25.03) (14.02)

As with the percent correct analysis, our three-year-old participantsshowed little consistent change in either experimental group. In con-trast, the bottom graph of figure 15.2 shows that more “Barney”-watch-ers’ scores worsened than improved or stayed the same, from pretest toposttest, while more nonwatchers’ scores improved than worsened orstayed the same. A chi-square test revealed that these two distributionswere significantly different (X2 = 5.96, p < 0.06). More importantly, a signtest revealed that significantly more watcher four-year-olds’ scoresworsened (7) than improved (2; p = 0.07 using the binomial distribu-tion).

In summary, it appears that watching ten episodes of the TV show“Barney and Friends” did not affect three-year-olds’ understanding of

Manipulating the Input 261

Figure 15.1.Percent of correct responses distinguishing think and guess from know, at Time 1 and Time 2.

the certainty distinction between the mental verbs think and guess, andknow; however, such viewing did appear to affect the four-year-olds.Taken together, the percent correct and number who change analysesshowed that the four-year-old children in the nonwatcher condition im-proved their scores, whereas the scores of many of those in the watchercondition declined. Thus, watching “Barney” seems to have led morefour-year-olds to minimize the certainty distinction between think andguess, and know, whereas not watching “Barney” is associated with fur-ther progress on this distinction.

These results suggest that, indeed, young children are sensitive to theways mental verbs are used. It was not the case that simply presentingmore instances of think, know, and guess yielded improved performance;in fact, more children who heard additional MSVs (the watchers) per-formed more poorly after exposure. What seemed to be happening tothe watcher group was that the frequent certain uses of “think” and

262 Letitia R. Naigles

Figure 15.2.Number of children whose mental-verb-distinction scores improve, worsen, or stay thesame from Time 1 to Time 2.

“guess” highlighted one way in which these verbs were equivalent to“know,” and so reinforced their undifferentiated status with respect tothat verb. In other words, the “Barney” input could be viewed as tem-porarily shifting the balance of differentiating and nondifferentiatinginput the children received, so as to create a (one hopes) momentarydelay or decrement in the watchers’ progress on the think/know andguess/know distinctions.6

Why did the nonwatchers, who received no special input, improvetheir scores from pretest to posttest? This question is really part andparcel of the larger one with which I began: Why do most children im-prove in their understanding of the certainty distinction between think,know, and guess after age four? Earlier, I hypothesized that this im-provement could be attributed to a change in children’s input, specifi-cally, an increase in the proportion of uncertain think and guess uses byadults. The results of the “Barney” study give this hypothesis someplausibility, in that children this age were shown to be sensitive to theways MSVs are used; however, the study did not explain how thechange in input actually occurs. That is, what is it that instigates thischange? Do adults tap into some cognitive development that childrenhave made and adjust their usage accordingly? Or do the children needless directing of their interactions, thus “freeing” adults’ use of think tomanifest its other senses? Both of these factors might contribute, but it ishard to conceive of an entire population of parents deliberately alteringtheir speech to their children at just the same age in order to facilitatethis development. As a previous generation of Gleitman students hasshown, parents’ talk to their children is primarily for the purposes of so-cialization and care, not for language teaching (Newport et al. 1977).However, it is the case that many children—especially those who arelikely to be participants in developmental psychology studies—beginto receive a new form of input just around three to four years of age.This new input comes not from parents, but from preschool teachers.

Study 2: Does preschool experience influence mental verb understanding?A major social change occurs in many children’s lives at around three tofour years of age, in that they begin to attend preschool (or child careprograms that include a preschool component) for anywhere from fif-teen to forty hours per week. Before this time, most children are caredfor either at home or in small family child care settings (Hofferth 1996).The preschool experience may be very different from this earlier type ofcare, in that (a) there are more children with whom to interact, espe-cially more children close in age; (b) there is more structure to the day;and (c) teacher-child interactions tend to be more purposely instructivethan mother-child interactions (e.g., about colors, numbers, and letters).

Manipulating the Input 263

Some recent studies have shown that preschool interactions poten-tially relevant to MSV development are different in kind from interac-tions at home with parents. Overall, the linguistic input provided inpreschool by teachers has been found to be both more formal and morecomplex than that heard at home (Dickinson and Smith 1995). More-over, when Hall et al. (1987) coded adult usage of MSVs as a class (i.e.,not broken down by individual verb), they found that the typicalparental usage was different from the typical teacher usage (this alsovaried by social class). Brown, Donelan-McCall, and Dunn (1996) com-pared MSV usage in mothers, siblings, and friends in conversation withfour-year-olds, and found that friends’ (and siblings’) MSV use (again,not broken down by verb, although think and know were the most com-mon) included more modulations of assertion than did mothers’.Finally, there has emerged recently some evidence that the experienceof good quality child care or preschool matters in the pace of linguisticand cognitive development. Huttenlocher (1995) found that five-year-olds experience more growth in language comprehension over the partof the year that includes preschool attendance than over the part that in-cludes the summer vacation. And Shatz, Behrend, Gelman, and Ebeling(1996) have found that two-year-olds who attend child care show bettercolor-name understanding than their peers who are cared for at home.

My hypothesis, then, was that the preschool environment plays a sig-nificant role in the observed progression of MSV understanding fromage three to age four. It is possible that, for example, teachers of pre-schoolers may use think and guess in their uncertain senses more thanmothers do. Morever, children may hear more of such uses from theirpeers, as three-year-olds and four-year-olds are often in the same classin American preschools. My conjecture was that such preschool experi-ences may provide a partial account for four-year-olds’ enhanced un-derstanding and performance on MSV comprehension tasks relative tothree-year-olds. The current literature on MSV development (and alsoTOM development, for that matter) cannot speak to this hypothesis, be-cause all of the experimental studies that I know of have used preschoolattendees as participants. What this means, though, is that the literatureincludes a potential confound: Is the developmental pattern that hasbeen observed a factor of age, or of time spent in preschool? It was timefor a “true” deprivation study.

How could preschool input be manipulated, to see the extent towhich it accounts for this transition in mental verb understanding?Luckily, here I could take advantage of a “natural experiment” in theworld, because although most American three- and four-year-olds (par-ticularly the latter) do attend some kind of preschool, sizeable numbersexist whose parents have chosen to keep them at home. Comparisons of

264 Letitia R. Naigles

the MSV understanding of children who have and who have not at-tended preschool might reveal differences in the onset of their under-standing of MSV distinctions. My prediction was that children whoattend preschool would show enhanced understanding of the degree ofcertainty distinction among the verbs think, guess, and know, relative totheir agemates who have not yet attended preschool.

Method

Participants Twenty-four child subjects participated, twelve of whomwere drawn from local preschools (MA = 52.5 months (SD = 3.28)).These children were enrolled in preschool full-time (i.e., 40 hours perweek). The 12 home-reared children (MA = 53.7 months (SD = 4.66))were recruited from playgrounds, flyers in doctors’ offices, and muse-ums. These children had minimal experience with child care; what ex-perience they had was in family child care (M = 8.79 hours per week).All of the children were monolingual speakers of American English,and all belonged to middle SES families. An additional three preschoolchildren were eliminated because of their failure to reach criterion onthe practice trials.

The materials and procedure were the same as for Study 1. Thepreschool children were tested in their preschools and the home-rearedchildren were tested at home.

Results and discussionThe responses were again tabulated for percent correct; the children’sperformance on the think/know and guess/know distinctions were com-bined. The results are shown in figure 15.3. The performance of thepreschoolers (M = 71.87% correct, SD = 19.18) was in line with thatfound by previous studies (e.g., Moore et al. 1989), and was signifi-cantly better than would be expected by chance (p < 0.05). Nine of thetwelve children performed at 62.5% correct or better. The performanceof the home-reared children was much lower (M = 55.21% correct, SD =22.51), did not differ significantly from chance (p > 0.10), and was sig-nificantly worse than that of the preschoolers (t(22) = 1.95, p < 0.05).Only six of the twelve home-reared children performed at 62.5% corrector better.

These findings support my prediction that preschoolers would per-form better on MSV comprehension tasks than children of the same agewho had not attended preschool. These four-year-old preschoolers cor-rectly distinguished think and guess from know, in that they chose thebox designated by the puppet who said “I know,” rather than the pup-pet who said “I think” or “I guess,” significantly more often than would

Manipulating the Input 265

be expected by chance. In contrast, the home-reared four-year-olds’ per-formance resembled that of the three-year-old subjects seen in otherstudies (e.g., Moore et al. 1989; Johnson and Maratsos 1977): They wereequally likely to pick the boxes designated by puppets who used “know,”“think,” or “guess.” In other words, they did not distinguish these threeverbs on the degree of certainty dimension.

C. Discussion and Conclusions

Thus far, these hypotheses concerning a role for input in children’s ac-quisition of MSV distinctions have received some preliminary support:Both television input and preschool experience affected children’s per-formance on a test requiring them to distinguish between mental verbs.That is, television input that minimized the certainty distinction be-tween think, guess, and know evidently led more four-year-olds to treatthe three verbs as equivalent on this dimension. Moreover, preschoolinput—broadly defined as full-time experience in preschool—evidentlyresulted in the relevant children treating the verbs more distinctivelythan their non-preschool-attending peers. The notion is, then, that oneinstigating factor for children’s development of the certainty distinctionbetween think, guess, and know at age four is that their preschool-basedinput has gained some empirical as well as theoretical plausibility.

266 Letitia R. Naigles

Figure 15.3.Percent of correct responses distinguishing think and guess from know, for preschool-attending and home-reared four-year-olds.

Clearly, though, more research is needed to address some criticalmethodological and theoretical issues. For example, one methodologi-cal question concerns how well the two samples in Study 2, of preschoolattendees and home-reared children, were equated. That is, just becausethe children were closely matched in age did not necessarily mean theywere as closely matched in other aspects of development, be they social,linguistic, or cognitive. Of course, I could not randomly assign half ofthe children to attend preschool and the other half to stay at home; I wasconstrained by the parents’ decisions regarding whether to send theirchildren to preschool or not. Thus it is possible that the preschool atten-dees were already ahead of their home-reared peers in language devel-opment, and this was why they were attending preschool. In otherwords, the time course of the children’s development may have causedtheir preschool attendance rather than the other way around. My col-laborators and I are beginning to address this issue by conducting a lon-gitudinal study in which three-year-old preschool attendees andhome-reared children, now matched on language and cognitive devel-opment milestones as well as age, are being repeatedly assessed fortheir mental verb understanding over the course of 1.5 years. Ifpreschool experience is a key factor in beginning to understand the cer-tainty distinction between think, guess, and know, then preschool atten-dees should perform above chance on these tests at an earlier age thanhome-reared children. Our preliminary findings point in this direction(Marsland, Hohenstein, and Naigles 1997).

More theoretical questions concern how the preschool experience, ifreal, exerts its influence. What is it about preschool that may be facilitat-ing the acquisition of this MSV distinction? Any serious answer to thisquestion must include detailed comparisons of teacher-preschooler andparent-child interactions, thereby highlighting how the language usedby adults in preschool differs from that used at home. My collaboratorsand I have collected a corpus of such interactions and are in the processof performing such comparisons (see Hohenstein, Naigles, andMarsland 1998 for some preliminary findings). What we have uncov-ered so far are numerous interactions in the preschools, such as thosebelow, which have the potential to be facilitative.

(1) Teacher: What color is your ant?Child A: BlackChild B: BrownChild A: No, blackChild B: I said brownTeacher: Thank you. And I think there are brown ants, I’malmost positive!

Manipulating the Input 267

(2) Teacher: Well, here’s a page missing, but this is what I thinkthe page said.

(3) Teacher: Now let’s count up here, one, two, three, fourChild A: Four on oneTeacher: Are you reading behind my back? Let’s count here.Child B: Five on one.Teacher: Wait a minute, now you’re guessing. Don’t do that.

In the first two extracts, the teacher’s use of think seems explicitlymarked as less-than-certain because she is only “almost positive” in (1)and because a page in the storybook is missing in (2). In extract (3) theteacher is reading Bears on Wheels (Berenstain and Berenstain 1969) butthe child is talking about a page yet to be read. The teacher’s use of guessin this context may serve to highlight her sense that the child must beuncertain about what she is saying. We expect to see fewer of such in-teractions in our home recordings, although we have not yet analyzedenough of them to come to any conclusions. In addition, in line with thelinguistics tradition’s focus on MSV forms, we expect to find more syn-tactically distinctive uses of think, guess, and know—what Naigles andHoff-Ginsberg (1995) have termed “syntactic diversity”—in teachers’input than in mothers’ input.

With these additional studies, we will have a clearer picture of whenchildren learn what about mental state verbs, and how their input (as op-posed to other aspects of their development) contributes to this learn-ing. Notice again that I am proposing a very specific role for a veryspecific type of input here, namely, that preschool input, by virtue of itsformality and didactic context, enables the appropriate contexts for thedistinctive use of these mental verbs in a way that the usual maternalinput, with its focus on socialization and care, does not. One would notnecessarily expect that the preschool experience would matter for otheraspects of language acquisition, such as the acquisition of argumentstructure or of yes-no questions, because these aspects seem less sus-ceptible to the overly polite register often used with young children inthis culture. However, given the correlations observed between MSVacquisition and theory of mind development, it is possible that thepreschool experience may also facilitate children’s development of aTOM. Recent discussions of TOM development have begun to considerthe child’s environment in more detail, and researchers have pointed tosuch possible instigating factors as siblings in general, intersibling con-flict and trickery, pretense, and peer language use (Bartsch and Well-man 1995; Jenkins and Astington 1996; Perner et al. 1994; Brown et al.1996; Lillard 1993; Lewis et al. 1997). Surprisingly, none has specificallymentioned the preschool experience, in which all of these factors ap-

268 Letitia R. Naigles

pear in combination. And yet preschool may turn out to be an impor-tant catalyst for many of the cognitive achievements children have beenshown to make between the ages of three and five. In closing, the depri-vation paradigm pioneered by Lila and Henry Gleitman for research inlanguage learning has shown its worth once again, by highlighting andsuggesting how to weight the joint roles of input and endowment inchildren’s acquisition of language.

Acknowledgments

I am grateful to all of the teachers, parents and children who partici-pated in these studies. Much of this work was collaborative, performedwith Dorothy Singer, Jerome Singer, Betina Jean-Louis, David Sells, andCraig Rosen; moreover, I thank Abigail Heitler and Nancy McGraw fortheir assistance in data collection. This research has also benefitedgreatly from conversations with many colleagues, most especially JillHohenstein, Kate Marsland, Alice Carter, Jill deVilliers, Larry Horn,Bonnie Leadbeater, and Susan Rakowitz. This research was supportedby NIH FIRST Award HD26596 and a Yale University Social ScienceResearch Fund Fellowship. Correspondence should be sent to LetitiaNaigles, Department of Psychology, 406 Babbidge Road, U-20,University of Connecticut, Storrs, CT 06269-1020.

Notes

1. In neg-raising, the negative element in the main clause of a complex sentence reallyserves to negate the verb in the subordinate clause. The general idea is that the negatedelement can be “raised” from the subordinate clause to the main clause, but the nega-tion itself remains in the lower clause (see Horn 1978 for more discussion).

2. None of these studies has actually investigated whether children distinguish the factiv-ity and certainty dimensions from each other, although Moore and Davidge (1989)claim that the certainty dimension is primary in these initial mental state distinctions(see also Tager-Flusberg et al. 1997). Moreover, researchers have not yet investigatedthe process/product dimension with children in this age group.

3. This is with first-order false beliefs, which are distinguished from second-order falsebeliefs in that they are not embedded (Wimmer and Perner 1983; Wellman 1990;Astington 1998). Thus She thinks that the chocolate is in the cabinet, even though it is really inthe freezer is an example of a first-order false belief, whereas She thinks that he thinks thatthe chocolate is in the cabinet, even though it is really in the freezer is an example of a second-order false belief.

4. Analyzing MSVs as a class, Brown and Dunn (1991) noticed that mothers of two-year-olds tend to use them more in commentary talk than in didactic talk, and more in refer-ence to others than to the target child. This may result in the verbs being less salient tothe child and so contribute to their delay in acquisition relative to social/emotional andconcrete verbs.

5. The fact that the nonwatchers’ performance at pretest was considerably lower than that

Manipulating the Input 269

of the watchers’ raises the possibility that the former group’s improvement at posttestis attributable to regression to the mean. When we controlled for the children’s pretestscores with an ANCOVA, however, the interaction of age and group was still present,albeit at a lower level of significance. Furthermore, the estimated posttest scores for thenonwatchers were still higher than those for the watchers (72.8% vs. 62.9%). Thus it isunlikely that the nonwatchers’ improvement at posttest is solely a function of their de-pressed scores at pretest.

6. How can we be sure that it was the specific mental verb input of “Barney” that resultedin the decline in the watchers’ scores, and not just a general effect of watching “Barney”or any kind of television? One clue comes from the second language task these childrenparticipated in at pretest and posttest. They were asked to enact ungrammatical sen-tences in which transitive verbs were placed in intransitive frames and intransitiveverbs were placed in transitive frames (cf. Naigles, Gleitman, and Gleitman 1993).Their enactments were coded as to whether they followed the demands of the syntacticframe (the usual preschool-aged child response) or the demands of the verb (the usualgrade-schooler and adult response). On this task, the watcher group performed betterfrom pretest to posttest (i.e., adhered more to the demands of the verb) while the non-watchers showed no change (see Naigles, et al. 1995 and Naigles and Mayeux, in press,for more detail). At the very least, then, it is not the case that watching these tenepisodes of “Barney” depresses language abilities or performance overall.

References

Abbeduto, L. and Rosenberg, S. (1985) Children’s knowledge of the presuppositions of“know” and other cognitive verbs. Journal of Child Language 12:621–641.

Astington, J. (1998) Theory of mind, Humpty Dumpty, and the icebox. Human Develop-ment 41:30–39.

Bartsch, K. and Wellman, H. (1995) Children Talk About the Mind. Oxford: Oxford Uni-versity Press.

Berenstain, S. and Berenstain, J. (1969) Bears on Wheels. New York: Random House.Booth, J. and Hall, W. S. (1995) Development of the understanding of the polysemous

meanings of the mental-state verb know. Cognitive Development 10:529–549.Bretherton, I. and Beeghly, M. (1982) Talking about internal states: The acquisition of an

explicit theory of mind. Developmental Psychology 18:906–921.Brown, J. and Dunn, J. (1991) “You can cry, mum”: The social and developmental impli-

cations of talk about internal states. British Journal of Developmental Psychology9:237–256.

Brown, J., Donelan-McCall, N., and Dunn, J. (1996) Why talk abut mental states? The sig-nificance of children’s conversations with friends, siblings, and mothers. ChildDevelopment 67:836–849.

Clark, E. (1993) The Lexicon in Acquisition. Cambridge: Cambridge University Press.deVilliers, J. (1994) Questioning minds and answering machines. In Proceedings of the 1994

Boston University Conference on Language Development. Somerville, MA: CasadillaPress.

deVilliers, J. (1995) Steps in the mastery of sentence complements. Society for Research inChild Development, Indianapolis, IN.

Dickinson, D. and Smith, M. (1995) Effects of preschool lexical environment on low-in-come children’s language skill at the end of kindergarten. Paper presented at theBienniel Meeting of the Society for Research in Child Development, Indianapolis, IN.

270 Letitia R. Naigles

Dixon, R. M. W. (1991) A New Approach to English Grammar, on Semantic Principles. Oxford:Clarendon Press.

Feldman, H., Goldin-Meadow, S., and Gleitman, L. R. (1978) Beyond Herodotus: The cre-ation of language by linguistically deprived deaf children. In A. Locke (ed.), Action,Symbol, Gesture: The Emergence of Language. New York: Academic Press.

Fisher, C., Gleitman, H., and Gleitman, L. R. (1991) On the semantic content of subcatego-rization frames. Cognitive Psychology 23:331–392.

Fowler, A., Gelman, R., and Gleitman, L. R. (1994) The course of language learning in chil-drenwith Down Syndrome: Longitudinal and language level comparisons withyoung normally developing children. In H. Tager-Flusberg (ed.), Constraints onLanguage Acquisition: Studies of Atypical Children. Hillsdale, NJ: Erlbaum.

Frank, R. and Hall, W. S. (1991) Polysemy and the acquisition of the cognitive internalstate lexicon. Journal of Psycholinguistic Research 20:283–304.

Furrow, D., Moore, C., Davidge, J., and Chiasson, L. (1992) Mental terms in mothers’ andchildren’s speech: Similarities and relationships. Journal of Child Language19:617–631.

Gleitman, L. (1990) The structural sources of verb meanings. Language Acquisition 1:3–56.Gleitman, L. and Gleitman, H. (1992) A picture is worth a thousand words, but that’s the

problem: The role of syntax in vocabulary acquisition. Current Directions inPsychological Science 1:31–35.

Gleitman, L. and Gleitman, H. (1997) What is a language made of? Lingua 100:29–67.Hall, W. S. and Nagy, W. E. (1987) The semantic-pragmatic distinction in the investigation

of mental state words: The role of the situation. Discourse Processes 10:169–180.Hall, W. S., Nagy, W. E., and Linn, R. (1984) Spoken Words: Effects of Situation and Social

Group on Oral Word Usage and Frequency. Hillsdale, NJ: Erlbaum.Hall, W. S., Scholnick, E., and Hughes, A. (1987) Contextual constraints on usage of cog-

nitive words. Journal of Psycholinguistic Research 16:289–310.Hofferth, S. (1996) Child care in the United States today. The Future of Children: Financing

Child Care 6(2):41–61.Hogrefe, G., Wimmer, H., and Perner, J. (1986) Ignorance versus false belief: A develop-

mental lag in attribution of epistemic states. Child Development 57:567–582.Hohenstein, J., Naigles, L., and Marsland, K. (1998) Differences in mothers’ and preschool

teachers’ use of mental verbs. Presented at the Meeting of the Linguistic Society ofAmerica, New York City, January 1998.

Hooper, J. (1975) On assertive predicates. In J. Kimball (ed.), Syntax and Semantics, vol. 4.(pp. 91–124). New York: Academic Press.

Hopmann, M. and Maratsos, M. (1978) A developmental study of factivity and negationin complex syntax. Journal of Child Language 5:295–309.

Horn, L. (1978) Remarks on neg-raising. In P. Cole (ed.), Syntax and Semantics, vol. 9. (pp.129–220). New York: Academic Press.

Huttenlocher, J. (1995) Children’s language and relation to input. Paper presented at theBienniel Meeting of the Society for Research in Child Development, Indianapolis, IN.

Jenkins, J. and Astington, J.W. (1996) Cognitive factors and family structure associatedwith theory of mind development in young children. Developmental Psychology32:70–78.

Johnson, C. and Maratsos, M. (1977) Early comprehension of mental verbs: Think andKnow. Child Development 48:1743–1747.

Johnson, C. and Wellman, H. (1980) Children’s developing understanding of mentalverbs: Remember, know, and guess. Child Development 51:1095–1102.

Kiparsky, P. and Kiparsky, C. (1970) Fact. In M. Bierwisch and K. Heidolph (eds.),Progress in Linguistics (pp. 143–173). The Hague: Mouton.

Manipulating the Input 271

Kempson, R. (1975) Semantic Theory. Cambridge: Cambridge University Press.Landau, B. and Gleitman, L. (1985) Language and Experience. Cambridge: Harvard

University Press.Landau, B., Gleitman, H., and Spelke, E. (1981) Spatial knowledge and geometric repre-

sentation in a child blind from birth. Science 213:1275–1278.Leslie, A. (1991) The theory of mind impairment in autism: Evidence for a modular mech-

anism of development? In A. Whiten (ed.), Natural Theories of Mind. Blackwell.Lewis, C., Freeman, N., Kyriakidou, C., Maridaki-Kassotaki, K., and Berridge, D. (1996)

Social influences on false belief access: specific sibling influences or general ap-prenticeship? Child Development 67:2930–2947.

Lillard, A. (1993) Pretend play skills and the child’s theory of mind. Child Development64:348–371.

Limber, J. (1973) The genesis of complex sentences. In T. E. Moore (ed.), CognitiveDevelopment and the Acquisition of Language (pp. 169–185). New York: AcademicPress.

Lysvag, P. (1975) Verbs of hedging. In J. Kimball (ed.), Syntax and Semantics, vol. 4. (pp.125–154.) New York: Academic Press.

Macnamara, J., Baker, E., and Olson, C. (1976) Four-year-olds’ understanding of pretend,forget, and know: Evidence for propositional operations. Child Development47:62–70.

Marsland, K., Hohenstein, J., and Naigles, L. (1997) Learning that thinking is not knowing:The impact of preschool. Society for Research in Child Development, Washington,D.C., April, 1997.

Miscione, J., Marvin, R., O’Brien, R., and Greenberg, M. (1978) A developmental study ofpreschool chidren’s understanding of the words “know” and “guess.” Child Devel-opment 49:1107–1113.

Moore, C., Bryant, D., and Furrow, D. (1989) Mental terms and the development of cer-tainty. Child Development 60:167–171.

Moore, C. and Davidge, J. (1989) The development of mental terms: Pragmatics or seman-tics? Journal of Child Language 1:633–642.

Moore, C. and Furrow, D. (1991) The development of the language of belief: The expres-sion of relative certainty. In D. Frye and C. Moore (eds), Children’s Theories of Mind:Mental States and Social Understanding (pp. 173–193). Hillsdale, NJ: Erlbaum.

Moore, C., Furrow, D., Chiasson, L., and Patriquin, M. (1994) Developmental relation-ships between production and comprehension of mental terms. First Language14:1–17.

Naigles, L., Gleitman, H., and Gleitman, L. R. (1993). Children acquire word meaningcomponents from syntactic evidence. In E. Dromi (ed.), Language and Development(pp. 104–140). Norwood, NJ: Ablex.

Naigles, L. and Hoff-Ginsberg, E. (1995) Input to verb learning: Evidence for the plausi-bility of syntactic bootstrapping. Developmental Psychology 31:827–837.

Naigles, L. and Mayeux, L. (in press) Television as incident language teacher. To appearin D.G. Singer and J. Singer (eds.), Handbook of Children and the Media. Beverly Hills,CA: Sage.

Naigles, L., Singer, D., Singer, J., Jean-Louis, B., Sells, D., and Rosen, C. (1995) Barney says,“come, go, think, know”: Television influences specific aspects of language devel-opment. Presented at the American Psychological Society, New York, NY.

Newport, E., Gleitman, H., and Gleitman, L. (1977) Mother, I’d rather do it myself: Someeffects and noneffects of maternal speech style. In C. Snow and C. Ferguson (eds.),Talking to Children (pp. 109–150). Cambridge: Cambridge University Press.

272 Letitia R. Naigles

Perner, J., Ruffman, T., and Leekman, S. (1994) Theory of mind is contagious: You catch itfrom your sibs. Child Development 65:1228–1238.

Quine, W. v. O. (1960) Word and Object. Cambridge, MA: MIT Press.Rice, M., Huston, A., Truglio, R., and Wright, J. (1990) Words from “Sesame Street”:

Learning vocabulary while viewing. Developmental Psychology 20:421–428.Rice, M. and Woodsmall, L. (1988) Lessons from television: Children’s word learning

when viewing. Child Development 59:420–429.Richards, M. (1982) Empiricism and learning to mean. In S. Kuczaj (ed.), Language

Development Vol. 1, Syntax and Semantics (pp. 365–396). Hillsdale, NJ: ErlbaumAssociates.

Scholnick, E. (1987) The language of mind: Statements about mental states. DiscourseProcesses 10:181–192.

Schwanenflugel, P., Fabricus, W., and Noyes, C. (1996) Developing organization of men-tal verbs: Evidence for the development of a constructivist thoery of mind in mid-dle childhood. Cognitive Development 11:265–294.

Schwanenflugel, P., Fabricus, W., Noyes, C., Bigler, K., and Alexander, J. (1994) The orga-nization of mental verbs and folk theories of knowing. Journal of Memory andLanguage 33:376–395.

Shatz, M., Behrend, D., Gelman, S., and Ebeling, K. (1996) Color term knowledge in two-year-olds: Evidence for early competence. Journal of Child Language 23:177–200.

Shatz, M., Wellman, H. and Silber, S. (1983) The acquisition of mental verbs: A systematicinvestigation of the first reference to mental state. Cognition 14:301–321.

Singer, J. and Singer, D. (1997) “Barney and Friends” as entertainment and education:Evaluating the quality and effectiveness of a television series for preschool chil-dren. In W. K. Asamen and G. Berry (eds.), Research Paradigms in the Study ofTelevision and Social Behavior. Beverly Hills, CA: Sage.

Singer, J., Singer, D., Sells, D., and Rosen, C. (1995) “Barney and Friends” as education andentertainment: The comprehension study: Preschoolers’ cognitive responses immediatelyafter viewing a Barney episode. New Haven, CT: Yale University Family TelevisionResearch and Consulation Center.

Snow, C. and Ferguson, C. A. (1977) Talking to Children. Cambridge: CambridgeUniversity Press.

Sullivan, K. and Winner, E. (1991) When 3-year-olds understand ignorance, false belief,and representational change. British Journal of Developmental Psychology 9:159–171.

Tager-Flusberg, H. (1993) What language reveals about the understanding of minds inchildren with autism. In Baron-Cohen, S., Tager-Flusberg, H., and Cohen, D.Understanding Other Minds: Perspectives from Autism. Oxford: Oxford UniversityPress.

Tager-Flusberg, H., Sullivan, K., Barker, J., Harris, A., and Boshart, J. (1997) Theory ofmind and language acquisition: The development of cognitive verbs. Society forResearch in Child Development. Washington, D.C.

Urmson, J. (1963) Parenthetical verbs. In C. Caton (ed.), Philosophy and Ordinary Language(pp. 220–246). Urbana: University of Illinois Press.

Wellman, H. (1990) The Child’s Theory of Mind. Cambridge, MA: MIT Press.Wellman, H. and Estes, D. (1986) Early understanding of mental entities: A reexamination

of childhood realism. Child Development 57:910–923.Wellman, H. and Estes, D. (1987) Children’s early use of mental verbs and what they

mean. Discourse Processes 10:141–156.Wierzbicka, A. (1988) The Semantics of Grammar. Philadelphia: John Benjamins Publishing

Company.

Manipulating the Input 273

Wimmer, H. and Hartl, M. (1991) Against the Cartesian view on mind: Young children’sdifficulty with own false beliefs. British Journal of Developmental Psychology9:125–128.

Wimmer, H. and Perner, J. (1983) Beliefs about beliefs: Representation and constrainingfunction of wrong beliefs in young children’s understanding of deception.Cognition 13:103–128.

274 Letitia R. Naigles

Chapter 16

Partial Sentence Structure as an Early Constrainton Language Acquisition

Cynthia Fisher

For the jokes alone, the students of Lila and Henry Gleitman would beforever in their debt. But the true debt, of course, is even greater. Lilaand Henry, as teachers and scientists, encourage in their students both athorough respect for the great complexity and elegant systematicity ofhuman language, and an equal regard for the complexity and system-aticity of human learning. Together, these themes invite a series of ques-tions that characterize much of the research on language acquisitionthat has emerged from the group including the Gleitmans and their stu-dents. That is, what can the learner—a child who does not yet know thegrammar or the lexicon of English or Greek, or whatever language is tobe learned—begin with in learning any particular language? How willthe child identify and take in the relevant data provided in the environ-ment? How will the child analyze and interpret the data he or she canencode? These are fundamental questions about the acquisition of lan-guage, but they are also questions about how very young children per-ceive, remember, and learn from language experience.

The need to find a perceptible starting point, and to specify how thechild proceeds from this point, is unmistakable to all who turn theirthoughts to this matter, and is clearly stated in the following wordsfrom Chomsky. This quote is particularly appropriate in this contextsince it was recently pointed out to me by Lila as a plain statement of theproblem:

[O]ne has to try to find the set of primitives which have the empir-ical property that the child can more or less tell in the data itselfwhether they apply before it knows the grammar. . . . So now takegrammatical relations, say the notion subject. The question is: is itplausible to believe that in the flow of speech, in the noises that arepresented, it is possible to pick out something of which one cansay: here is the subject? That seems wildly implausible. Rather itseems that somehow you must be able to identify the subject onthe basis of other things you’ve identified, maybe configurationalnotions which are somehow constructed out of accessible materials

or maybe out of semantic notions, which are primitive for the lan-guage faculty. (Chomsky 1982, 118–119)

These primitives, whatever they turn out to be, are a part of what wehave come to call Universal Grammar (UG), broadly conceived as theset of capacities and limitations, mechanisms, and constraints that per-mit a child to soak up languages like a sponge, and guarantee that alllanguages, various and mutually incomprehensible as they are, share aset of core properties. It goes without saying that the charge embodiedin this quote is an extremely tall order. What I will do in this chapter ismerely review evidence and arguments for a few potential primitives.The story I hope to tell—with some but not all of the relevant data al-ready in—can be summarized as follows: Viewed in the way I will de-scribe, both configurational and semantic notions can be constructedout of materials ready to the child’s hand, and arguments can be madethat together they yield an appropriately constrained starting point forlinguistic understanding and syntax acquisition. The ideas summarizedhere have grown out of years of collaboration with Lila and Henry, andfollow directly from their previous and ongoing ground-breaking workon syntactic bootstrapping (e.g., Landau and Gleitman 1985; Gleitman1990; Gleitman and Gleitman 1997). To the extent that this makes anysense now, it is owing to their teaching, inspiration, innovation, andcontinued collaboration.

The Contribution of Sentence Structure to Meaning

It is a truism that the syntactic structure of a sentence affects its inter-pretation. This is what syntax is for: Brutus killed Caeser and Caeser killedBrutus differ in both sense and truth value, and languages’ various tech-niques for signaling the role of each noun phrase relative to the verbconstitute the basic grammatical relations of the clause. The contribu-tion of sentence structure to meaning can be seen in some often-described phenomena: First, the same verbs occurring in differentstructures have different meanings (see, e.g., Goldberg 1996; Rappaportand Levin 1988; Ritter and Rosen 1993, among many others). For exam-ple, sentences (1) through (3) below all use the main verb have. But Janeowns something in (1), causes an event in (2), and experiences a misfor-tune in (3) (examples adapted from Ritter and Rosen 1993). Not much ofthese various senses belongs directly to have. Second, adults readily andlawfully interpret novel uses of verbs like the ones in (4), adapted fromGoldberg (1996; see also Fisher 1994). Presumably, to understand orproduce these, we need not already know that laugh or frown can conveytransfer of possession or position. Instead, the three-argument structure,

276 Cynthia Fisher

in combination with the words in the sentence, gives it that meaning.Children produce (see, e.g., Bowerman 1982) and understand (Naigles,Fowler, and Helm 1992) these novel uses as well; some of Bowerman’sexamples are shown in (5). Ritter and Rosen (1993) argue that the sur-face structure and lexical content of a sentence must always be con-sulted to interpret the verb in that sentence. However this knowledgemay best be modeled in adult grammars, the contribution of sentencestructure to sentence meaning is clear. Some set of links between syntaxand semantics permits adults to infer aspects of a sentence’s meaningfrom its structure.

(1) Jane had a brown dog.

(2) Jane had her secretary get her a cup of coffee.

(3) Jane had her dog get run over by a car.

(4) The panel laughed the proposal off the table.

Her father frowned away the compliment.

(5) Don’t say me that or you’ll make me cry.

Why didn’t you want to go your head under?

Syntactic Bootstrapping: The Basic Claim

The view known as syntactic bootstrapping (Gleitman 1990; Landauand Gleitman 1985) proposes that young children use precursors of thesame links between sentence structure and meaning, in concert with ob-servations of world events, to understand sentences and therefore to ac-quire the meanings of verbs. If part of the relational meaning of a verbin a sentence is predictable from the sentence structure itself, then achild who hears a sentence containing a novel verb could gain some in-formation about the meaning of the sentence from its structure. Thisclaim is supported by evidence that children from about two to fiveyears of age take novel verbs in different sentence structures to meandifferent things (see, e.g., Fisher 1996; Fisher, Hall, Rakowitz, and Gleit-man 1994; Naigles 1990; Naigles and Kako 1993).

The semantic information gleaned from syntax will necessarily bevery abstract. After all, many verbs with widely varying meaningsoccur in each syntactic structure: Transitive verbs include break and like,intransitive verbs include dance and sleep. The interpretive informationthat could be inferred from a sentence structure could be described as relevant to a sentence’s semantic structure—for example, how manyparticipants are involved in the sentence?—rather than event-dependent

Partial Sentence Structure as an Early Constraint 277

semantic content (see, e.g., Grimshaw 1993). Dance and sleep are similar,not in the specifics of the activities or states they describe, but in theirformal structure: Both require only one participant.

Moreover, as mentioned above, most verbs occur in more than onesentence frame. This information could further constrain interpreta-tions of each verb, much as subcategorization frame set information hasplayed such a powerful role in linguistic characterizations of semanticsin the verb lexicon (see, e.g., Levin and Rappaport-Hovav 1995). That is,while explain in (6) shares an abstract semantic structure with otherthree-place predicates, explain also occurs with sentence complements(as in 7), and shares semantic structural properties with other sentence-complement-taking verbs. This combination of frames more sharplylimits the possible interpretations that are consistent with both sentenceframes (Fisher, Gleitman, and Gleitman 1991; Gleitman and Gleitman1997). Recent evidence suggests that young children differently inter-pret a novel verb that appears in the two related frames shown in (8), asopposed to the two frames shown in (9) (Naigles 1996).

(6) Mary explained the program to John.

(7) Mary explained that her computer had eaten her paper.

(8) The ducki is pilking the bunny.The ducki is pilking.

(9) The duck is pilking the bunnyi.The bunnyi is pilking.

We have argued that such abstract hints from the syntax could help tosolve some serious problems for verb learning (see, e.g., Fisher 1994;Gleitman 1990). For example, a verb in a sentence does not simply labelan event, but instead describes a speaker’s perspective on that event.Thus sentences (10) and (11) could accompany the same events. The dif-ference between them lies not in whether the event (in the world) has acause, but in whether the speaker chooses to mention it. This is whyeven adults who already know the vocabulary of English cannot guesswhich verb a speaker utters when shown a set of events in which theverb was used, though they can reasonably accurately guess what nounwas uttered given the same kind of information (Gillette, Gleitman,Gleitman, and Lederer, 1999). Observations of events alone do not pro-vide the right kind of information to interpret a sentence. Sentencestructure cues, on the other hand, bearing principled relations to thesentence’s semantic structure, could provide information directly rele-vant to the speaker’s intent.

278 Cynthia Fisher

(10) The block goes in here.

(11) I’m putting the block in here.

How Does the Child Obtain Syntactic Evidence?

But, as Lila and Henry might say, not so fast (Gleitman and Gleitman1997). How could syntactic bootstrapping begin? A sentence structure isa complex object, constructed of elements that are quite implausible asprimitives to the language acquisition system—notions like argumentas opposed to adjunct noun phrase, and subject as opposed to object oroblique argument. In considering the possible role of sentence structurein the earliest comprehension of sentences, we must also keep in mindthe need to seek plausible presyntactic primitives, and mechanisms bywhich these might influence comprehension before a true syntactic de-scription of a sentence can be attained (Fisher et al. 1994).

Recent evidence for presyntactic structural cues to verb meaningA recent series of experiments was designed to isolate features of asentence’s structure, testing what aspects of sentence structure influ-enced young children’s interpretations of a novel verb. These studiesprovide evidence that a plausibly early description of the structure ofa sentence—its number of noun phrases—is meaningful to youngpreschoolers. In several studies, three- and five-year-olds (Fisher1996) and two-and-a-half- and three-year-olds (Fisher, in press) weretaught novel transitive or intransitive verbs for unfamiliar agent-pa-tient events. On each of four trials, children watched an event inwhich one participant moved another participant in some novel way.These events were described by a novel verb presented in a sentencecontext: One group of children heard intransitive sentences, while theother group heard transitive sentences. The key feature of this studywas that the identity of the subject and object of these sentences washidden by using ambiguous pronouns, yielding sentences that dif-fered only in their number of noun phrases. An example is shown in(12). The critical sentence frame was repeated several times (in appro-priate tenses) before, during, and after three repetitions of the sameevent.

(12) Event: One person rolls another on a wheeled dolly bypulling with a crowbar.

Transitive: She’s pilking her over there.Intransitive: She’s pilking over there.

Partial Sentence Structure as an Early Constraint 279

Following this introduction, on each trial the children’s interpreta-tions of the novel verb in its sentence context were assessed by askingthem to point to the participant, in a still display of the midpoint of theevent, whose role the verb described (e.g., “Which one was pilking theother one over there?” vs. “Which one was pilking over there?”). Bothadults and children 2.5 and 3 years old were more likely to choosecausal agents as the subjects of transitive than intransitive verbs,though neither sentence identified one participant in the event as thesubject. A subsequent study replicated this finding for the 2.5-year-oldgroup alone (28–32 months), finding that even this youngest groupchose agents as the participant whose actions the verbs described signif-icantly more often for transitive than intransitive verbs (Fisher, in press).

In previous studies of the role of syntax in verb learning, the linguis-tic contexts of novel verbs have always specified the identity of theverbs’ arguments (as in “The duck is blicking the bunny,” describing ascene in which these characters participated; Fisher et al. 1994; Naigles1990; Naigles and Kako 1993). Given this information, children mightachieve structure-sensitive interpretations of verbs by relying on as-sumptions about the class of semantic roles associated with each gram-matical position: Children could infer that the verb referred to theactivities of the participant mentioned in subject position, on the groundsthat grammatical subjects tend to be semantic agents. Such a procedureis plausible, and has sometimes been assumed in discussions of syntac-tic bootstrapping, in part for lack of any explicit alternative. Innate linksbetween thematic roles (abstract relational concepts like agent andtheme) and grammatical functions (like subject and direct object) havebeen proposed to explain cross-linguistic regularities in the assign-ments of semantic roles to sentence positions. Though various treat-ments of thematic roles differ significantly in their inventory of rolesand in how they map onto syntax, some system of thematic roles consti-tutes a primary device in linguistic theory for expressing relations be-tween verb syntax and semantics (see, e.g., Baker 1997; Dowty 1991;Grimshaw 1990; Jackendoff 1990; Rappaport and Levin 1988).

In the studies described above, however, the entire structure of thesentence, the configuration of arguments itself, was shown to be mean-ingful to quite young children. Even 2.5-year-olds interpret the subjectreferent to “mean” different things—play different roles—in the sameevent, depending on the overall structure of the sentence. Subjects arenot preferentially causal agents unless a verb has two noun phrase ar-guments. This finding gives strong support to the notion that sentencestructures per se are meaningful, to adults and to children as young as2.5 years, in a way not reducible to links between event roles like agentand patient or theme, and grammatical functions like subject or object.

280 Cynthia Fisher

Sentence Interpretation Based on Partial Sentence Representations

How could sentence structures provide information about the mean-ings of verbs in sentences, without the aid of links between thematicroles and particular grammatical positions? The approach taken byFisher et al. (1994), and further supported by the findings describedabove (Fisher 1996, in press), capitalizes on the intrinsically relational orstructural nature of sentences, conceptual representations, and verbmeanings (see, e.g., Bloom 1970; Braine 1992; Fisher 1996; Fisher et al.1994; Gentner 1982; Gleitman 1990; Grimshaw 1993; Jackendoff 1990).Given the following set of assumptions, the gross similarities amongthese structures could permit sentence structure to influence interpreta-tion.

Conceptual structuresFirst, in common with most recent work in verb semantics, based onJackendoff’s (1990) research, we assume that semantic structures ofverbs are essentially of the same kind as the nonlinguistic conceptualstructures by which humans represent events. Both verb semanticstructures and conceptual representations of events demand a divisionbetween predicates and arguments, and thus between relations and theentities they relate (see Bierwisch and Schreuder 1992; Bloom 1970;Braine 1992; Fodor 1979). Even otherwise divergent views of languageacquisition strongly assume that structured conceptual representationsof events, fundamentally like linguistic semantic structures, are a dri-ving force in language acquisition (see, e.g., Bloom 1970; Pinker 1989;and many others). The current view and any form of syntactic boot-strapping share this assumption (see, e.g., Fisher 1996; Gleitman 1990).

Sentence structuresSecond, we assume that children learning their first verbs can (a) iden-tify some familiar nouns in fluent speech, and (b) represent these asgrouped within a larger utterance structure. Whenever a child managesto do this, she will have what we have called a partial sentence repre-sentation (PSR; Fisher et al. 1994). The early appearance of nouns inchildren’s productive vocabularies has long been noted (see, e.g.,Gentner 1982). More to the point for present purposes is that evidencefor the comprehension of object names (see, e.g., Waxman and Markow1995) precedes comprehension of relational terms by a considerablemargin (e.g., Hirsh-Pasek and Golinkoff 1996), and there is strong evi-dence that at least some concrete noun meanings can be acquired fromobservation of word/world contingencies alone (Gillette et al. 1999).The grouping of words into utterances has also typically been assumed

Partial Sentence Structure as an Early Constraint 281

as a prerequisite to syntax acquisition. Recent explorations of utteranceprosody have begun to cash out this assumption, suggesting that chil-dren could hear utterances as cohesive based on the familiar prosodicmelodies of their language (see, e.g., Fisher and Tokura 1996; Jusczyk1997; Morgan 1986).

The influence of sentence structure on selection of a conceptual structureThese two sets of assumptions have consequences for early sentencecomprehension. When children interpret a sentence they link one struc-ture with another. To the extent that these distinct representations—sen-tence and conceptual—have similar structures, a sentence couldprovide a rough structural analogy for its interpretation in conceptualterms (see, e.g., Gentner 1983). Assuming that conceptual and semanticstructures are of like kind, the result of their alignment will be, againroughly, a semantic structure for the sentence.

To illustrate, even prior to the identification of subject and object, sen-tences still contain some number of noun phrases. This simple struc-tural fact could be informative. Once children can identify some nouns,they could assign different meanings to transitive and intransitive verbsby linking a sentence containing two noun phrases with a conceptualrelation between the two named entities in the current scene, and a sen-tence containing one noun phrase with a conceptual predicate charac-terizing the single named entity in the current scene. The result wouldbe a rough semantic structure for the sentence, with semantic contentderived from the specifics of the observed situation.

Structural alignment would allow children to map entire sentencestructures onto possible semantic structures derived from observationof events, without requiring prior identification of the subject referentas a grammatical subject, and thus could account for the findings fromthe pronoun-disambiguation task described above (Fisher 1996; inpress). Via structural alignment, merely identifying the set of nounswithin a representation of a sentence could give the hearer a clue as tothe speaker’s perspective on an event. This inference need not dependon true syntactic representations; thus if this description of the phe-nomenon is correct, it represents a potential presyntactic route wherebysimple aspects of the structure of a sentence could influence interpreta-tion.

Structure-sensitivity of this simple kind could presumably be imple-mented in a working model in many ways. For example, Siskind’s(1996) model of the role of cross-situational observation in vocabularylearning relies on the constraints that (a) the input is processed one ut-terance (rather than one word) at a time, and (b) any previously ac-quired elements of the meanings of words in an utterance must be

282 Cynthia Fisher

included in the interpretation selected from candidates available fromworld observation. As Brent (1996) points out, this pair of assumptionsmakes sentence interpretation a presyntactic mapping of sentence toworld rather than word to world, much as suggested by work in syn-tactic bootstrapping (Gleitman 1990).

A presyntactic division of the linguistic dataA presyntactic structure-to-meaning mapping constitutes only a veryrough take on argument linking, and leaves the child considerableroom to maneuver in interpreting sentences. However, the structuralalignment of sentence and scene representation as described abovewould permit a useful distinction between transitive and intransitivesentences, giving the child a significantly better chance of interpretingsentences as their speaker intended. If we assume that working outlinks between something like thematic roles and grammatical positionsplays a key role in syntax acquisition (see, e.g., Bloom 1970; Grimshaw1981; Pinker 1989), at least a rough presyntactic distinction betweentransitive and intransitive sentences may be essential. Discussions oflinking regularities assume, either explicitly or implicitly, that a predi-cate’s number of arguments is known from the start, often by limitingdiscussion to either two-place or three-place predicates (see, e.g., Baker1997; Dowty 1991). Without assuming a fixed number of arguments,links between thematic and grammatical roles are much less regular: Asour 2.5-year-old subjects showed that they knew, causal agents are themost likely subjects only of predicates with at least two arguments. Apresyntactic division of the linguistic data into (roughly) one-argumentand two-argument sentences could allow the child to begin with the do-mains within which semantic/syntactic mappings will be most regular.

Number of Nouns as a Presyntactic Primitive

But again, not so fast. How could the child know—before learning thegrammar of English—that these sentences contain one- versus two-argument predicates? Nouns in the sentence and arguments of a verb inthe sentence are not the same thing. In (13) and (14), dance has one argu-ment position but two nouns. Via conjunction in subject position in (13),and the addition of an adjunct prepositional phrase in (14), these sen-tences display more nouns than arguments. If children align a two-noun sentence with the most salient conceptual representation thatrelates the referents of those two nouns, then they should systematicallyerr in interpreting such sentences. That is, before a child has learnedwhat “with” and “and” mean, or that English transitive sentences can-not appear in NNV order, (13) and (14) should both yield the same in-terpretation as a transitive sentence.

Partial Sentence Structure as an Early Constraint 283

(13) Fred and Ginger danced.

(14) Ginger danced with Fred.

Previous research has explored these sentence types extensively, andthe overall pattern of results provides some preliminary evidence forthe predicted errors in children just at or under 2 years. At 25 months,children can interpret sentences like (13) correctly: Naigles (1990) intro-duced 25-month-olds to causal and noncausal versions of the sameevent (e.g., two characters moving in some manner under their ownpower versus one causing another to move in the same manner). Shefound that the children looked longer at the causal version when theyheard a novel transitive verb, as in (15), and looked longer at the non-causal version when they heard a novel intransitive verb, as in (16). Theintransitive sentence (16) is of the problematic type alluded to above, anintransitive verb appearing with two nouns conjoined in subject posi-tion. Successful interpretation of both sentences tells us that, by 25months, the children had learned enough about the word order andfunctional morphology of English to interpret this as an intransitivesentence despite its two nouns. Hirsh-Pasek and Golinkoff (1996), how-ever, found that children at 19, 24, and 28 months did not interpret sim-ilar sentences correctly when not given redundant morphological cuesto help identify the structure. An example is shown in (17): The subjectnoun phrase contains “and,” which should signal the conjoined subjectstructure to wise listeners, but does not also have the plural copula“are.” Apparently, without multiple clues that the unfamiliar verb is in-transitive, even 28-month-olds can be fooled by a mismatch betweennumber of argument positions and number of nouns. This suggests thatnumber of nouns is a strong early cue for structure-guided interpreta-tion, and also provides a tantalizing glimpse of young children’s grow-ing use of language-specific morphological evidence to differentiatesentence structures. Similarly, at 24 months, boys (but not girls) system-atically misinterpreted sentences like (18) as naming causal acts (Hirsh-Pasek and Golinkoff 1996). These two-year-olds, presumably unawareof the meaning of “with,” assume that a two-noun sentence is transitive.

(15) The duck is gorping the bunny.

(16) The duck and the bunny are gorping.

(17) Find Big Bird and Cookie Monster gorping!

(18) Find Big Bird gorping with Cookie Monster!

In summary, the presyntactic mechanism for syntactic bootstrappingproposed above makes a unique prediction. Before children acquire

284 Cynthia Fisher

much of the syntax and function morphology of a particular language,they should systematically misinterpret sentences that have morenouns than verb argument positions. Further research is needed to ex-plore these errors more fully. However, as described above, prior re-search gives some preliminary evidence for this prediction.

An Appropriately Constrained Starting Point

Thus far, I have suggested that a basic, presyntactic distinction betweentransitive and intransitive sentences could be achieved simply by iden-tifying the nouns in a sentence and representing them as parts of alarger utterance structure. This constitutes a partial sentence represen-tation, which shares gross structural properties of the conceptual struc-tures the sentence could convey. It is important to note that within thisview, a great deal of work in the selection of an interpretation remains tobe done by the child’s preferences in constructing conceptual represen-tations. In principle, the set of referents named in a sentence could be in-volved in indefinitely many different conceptual representations. Thus,like virtually all other views of language acquisition, syntactic and pre-syntactic bootstrapping depend on the young language learner to sharesignificant biases in the conceptualization of events with older humans.The addition proposed by syntactic and presyntactic bootstrapping issimply that sentence structures, however the child can represent them,can play an interesting role in language acquisition as well. In the re-maining space I will briefly address one kind of objection to the pro-posed view, and argue that, contrary to the objection, this accountprovides a useful initial constraint on the alignment of sentence andconceptual representations.

What about subjects?At first glance the separate treatment of transitive and intransitive sen-tences based on a presyntactic representation of their structures mayseem to stand in the way of an important syntactic generalization—thenotion of syntactic subject, encompassing both transitive and intransi-tive subjects. Subjects, after all, are picked out by a large constellation oflinguistic generalizations, including subject-verb agreement, case mark-ings, deletion in imperatives, special control properties, the affinity ofdiscourse topics for subject position, and so on (see, e.g., Keenan 1976).This would suggest that even though the proposed presyntactic infer-ence provides only a gross constraint on interpretation, it is nonethelesstoo specific to permit an important syntactic generalization.

However, as already mentioned, a category of grammatical subjectgeneral enough to encompass both transitive and intransitive subjects is

Partial Sentence Structure as an Early Constraint 285

not very useful for linking grammatical and semantic/conceptualstructures. The purport of the experimental evidence described above isthat quite young children (just like linguists) link grammatical and se-mantic roles within the constraints of the number of arguments pro-vided in a sentence. On the presyntactic structural alignment viewdescribed above, the child need not initially assume that either argu-ment of a transitive verb plays the same semantic role as the single ar-gument of an intransitive verb.

Moreover, it is not so clear that a category “subject” broad enough tospan all sentence structures should be considered a unitary primitivecategory. It has long been noted that the constellation of syntactic sub-ject properties alluded to above coheres imperfectly within and acrosslanguages (see, e.g., Keenan 1976). A particularly troublesome type ofcross-linguistic variation concerns the phenomenon of so-called erga-tive languages (see, e.g., Dixon 1994). A majority of languages, includ-ing English, have nominative-accusative syntax: The subject of anintransitive sentence is treated, morphologically and syntactically, likethe agent argument of a prototypical transitive sentence. This groupingof arguments defines the familiar category subject, with the set of spe-cial within- and across-sentence subject properties listed above: As in(19), the underlined elements are in nominative case, agree in numberwith the verb, control null subjects in conjunctions as shown in (20), andso on. But the ergative pattern is quite different. The agent argument ofa prototypical transitive receives its own case (ergative), whereas the in-transitive subject and the patient argument of a prototypical transitivereceive the same case. A few strongly syntactically ergative languageseven reverse the pattern shown in (20): Coreference across conjoinedverb phrases mirrors the morphologically ergative pattern, producingthe pattern glossed in (21), unimaginable in English (Baker 1997; Dixon1994).

(19) They see him.They flee.

(20) Theyi see him and Øi flee*They see himj and Øj flees

(21) *Theyi see him and Øi fleeThey see himj and Øj flees

Some (e.g., Marantz 1984) have suggested that the subjects of intran-sitives and the object arguments of transitive sentences together consti-tute the syntactic subject for those languages. This solution maintains

286 Cynthia Fisher

one syntactic definition of subjecthood—having the same case as thesubject of an intransitive—while dropping the cross-linguisticallywidespread link between subjects and agents. This approach raisesgrave problems for the project of defining regular links between syntac-tic and semantic relations as a starting point for language acquisition.However, by other accounts the claim that ergative languages have “pa-tient subjects” does not describe the linguistic phenomena very well.Not all of the syntactic properties associated cross-linguistically withthe subject category exhibit the reversal predicted by the patient-subjecthypothesis, even in the most strongly ergative languages (see, e.g.,Dixon 1994). Such data cast doubt not on the linking of agents with sub-jects in two-argument predicates, but on the existence of a single, prim-itive category “subject” that applies to both transitive and intransitivesentences across languages. Recent accounts that encompass these factspropose two senses in which a constituent can be the subject of a clause(Baker 1997; Dixon 1994), only one of which maintains the traditionallink between subject and agent.

What is the significance of these phenomena for the current dis-cussion? I have argued above that, given the polysemy of the subjectcategory, an early presyntactic distinction between transitive and in-transitive sentences is essential, giving the child the division of thedata within which linking regularities will work out. Now it seemsthat the same presyntactic division of the linguistic data could be essential for syntax acquisition more generally. Languages differ inhow they distribute important syntactic phenomena over the possiblecombinations of argument positions in transitive and intransitive sen-tences. If children can make a roughly accurate presyntactic distinctionbetween transitive and intransitive sentences based on their numberof nouns, then they could begin learning about the syntactic choices oftheir language without assuming that either argument of a transitivesentence should be treated syntactically like the single argument of anintransitive sentence. To establish the basic morphological and syntac-tic typology of a language, learners may have to begin with at leastthree basic structural positions in sentences (the two transitive argu-ment positions and the intransitive subject), rather than two (subjectand object) (see Dixon 1994). The developmental facts are at leastroughly consistent with this more flexible view of the starting pointfor syntax acquisition: Children seem to have no special difficulty ac-quiring languages with the ergative pattern, or with a combination ofergative and nominative morphology and syntax (see, e.g., Rispoli1991; Schieffelin 1985).

Partial Sentence Structure as an Early Constraint 287

Concluding Remarks

This proposal for presyntactic structural guidance in sentence interpre-tation is intended as a first example of what I believe will be a fruitfulline to pursue in discovering the earliest integration of sentence-structural and event information in verb learning. Lila’s original pro-posal for syntactic bootstrapping, developed with Barbara Landau,presented the strikingly innovative idea that “verb learning, whilepartly a function of the ostensive evidence provided, feeds upon theconceptual representation of predicate-argument logic in the syntacticformat of the sentence” (Landau and Gleitman 1985, p. 121). In laterwork (Fisher, Hall, Rakowitz, and Gleitman 1994), we proposed thatone could think of sentences as having structure even before the learnerknows enough about a particular grammar to build a true syntacticstructure. This partial or presyntactic structure shares some nearly in-escapable similarity with the range of conceptual structures that thatsentence could convey. In the work reviewed here, I have argued that ifwe endow the learner with some very simple alignment biases, thenthis primitive structure will influence interpretation as soon as the childcan identify some nouns and represent them as grouped within a largerutterance. The alignment of sentence and conceptual structure wouldprovide a (rough) presyntactic distinction between transitive and in-transitive sentences. This distinction is demonstrably helpful to youngchildren in sentence interpretation, and I have suggested that it mightbe needed for syntax acquisition as well. To acquire a grammar the childmust have some way to represent linguistic data presyntactically. Theintuition explored here is that even these initial representations couldhelp to constrain acquisition. By exploring the potential uses of partialinformation in each linguistic domain, we can move toward a morecomplete view of the information sources, constraints, and biases re-quired to get the child started in the acquisition of language.

Acknowledgment

The research described in this paper was partially supported by NSFgrant DBC 9113580, and by the University of Illinois.

References

Baker, M. C. (1997) Thematic Roles and Syntactic Structure. In L. Haegeman (ed.), Ele-ments of Grammar (pp. 73–137). Boston: Kluwer.

Bierwisch, M. and Schreuder, R. (1992) From concepts to lexical items. Cognition 42:23–60.Bloom, L. (1970) Language Development: Form and Function in Emerging Grammars.

Cambridge, MA: MIT Press.Bowerman, M. (1982) Reorganizational processes in lexical and syntactic development. In

E. Wanner and L. R. Gleitman (eds.), Language Acquisition: The State of the Art (pp.319–346). New York: Cambridge University Press.

288 Cynthia Fisher

Braine, M. D. S. (1992) What sort of innate structure is needed to “bootstrap” into syntax?Cognition 45:77–100.

Brent, M. R. (1996) Advances in the computational study of language acquisition.Cognition 61:1–38.

Chomsky, N. (1982) Noam Chomsky on the Generative Enterprise: A Discussion with R.Huybregts and H. van Riemsdijk. Dordrecht, Holland: Foris Publications.

Dixon, R. M. W. (1994) Ergativity. Cambridge: Cambridge University Press.Dowty, D. (1991) Thematic proto-roles and argument selection. Language 67(3): 547–619.Fisher, C. (1994) Structure and meaning in the verb lexicon: Input for a syntax-aided verb

learning procedure. Language and Cognitive Processes 9:473–518.Fisher, C. (1996) Structural limits on verb mapping: The role of analogy in children’s in-

terpretation of sentences. Cognitive Psychology 31:41–81.Fisher, C. (in press) Simple structural guides for verb learning: On starting with next to

nothing. In E. V. Clark (ed.), Proceedings of the 30th Stanford Child Language ResearchForum. Stanford, CA: CSLI Publications.

Fisher, C., Gleitman, H., and Gleitman, L. R. (1991) On the semantic content of subcatego-rization frames. Cognitive Psychology 23:331–392.

Fisher, C., Hall, D. G., Rakowitz, S., and Gleitman, L. R. (1994) When it is better to receivethan to give: Syntactic and conceptual constraints on vocabulary growth. Lingua92:333–375.

Fisher, C. and Tokura, H. (1996) Acoustic cues to grammatical structure in infant-directedspeech: Cross-linguistic evidence. Child Development 67:3192–3218.

Fodor, J. A. (1979) The Language of Thought. Cambridge, MA: Harvard University Press.Gentner, D. (1982) Why nouns are learned before verbs: Linguistic relativity versus nat-

ural partitioning. In K. Bean (ed.), Language, Thought, and Culture (pp. 301–334).Hillsdale, NJ: Erlbaum.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. CognitiveScience 7:155–170.

Gillette, J., Gleitman, H., Gleitman, L. R., and Lederer, A. (1999) Human Simulations ofLexical Acquisition. Cognition 73:135–176.

Gleitman, L. R. (1990) The structural sources of verb meanings. Language Acquisition 1(1):3–55.

Gleitman, L. and Gleitman, H. (1997) What is a language made out of? Lingua 100:29–55.Goldberg, A. (1996) Constructions: A Construction Grammar Approach to Argument Struc-

ture. Chicago: The University of Chicago Press.Grimshaw, J. (1981) Form, function, and the language acquisition device. In C. L. Baker

and J. J. McCarthy (eds.), The Logical Problem of Language Acquisition (pp. 165–182).Cambridge, MA: The MIT Press.

Grimshaw, J. (1990) Argument Structure. Cambridge, MA: MIT Press.Grimshaw, J. (1993) Semantic structure and semantic content: A preliminary note. Paper

presented at conference on Early Cognition and the Transition to Language.University of Texas at Austin.

Hirsh-Pasek, K. and Golinkoff, R. (1996) The Origins of Grammar. Cambridge, MA: MITPress.

Jackendoff, R. (1990) Semantic Structures. Cambridge, MA: MIT Press.Jusczyk, P. W. (1997) The Discovery of Spoken Language. Cambridge, MA: MIT Press.Keenan, E. L. (1976) Toward a universal definition of “subject.” In C. N. Li (ed.), Subject

and Topic (pp. 303–334). New York: Academic Press.Landau, B. and Gleitman, L. R. (1985). Language and Experience: Evidence from the Blind

Child. Cambridge, MA: Harvard University Press.Levin, B. and Rappaport-Hovav, M. (1995) Unaccusativity: At the Syntax-Lexical Semantics

Interface. Cambridge, MA: MIT Press.

Partial Sentence Structure as an Early Constraint 289

Marantz, A. (1984) On the Nature of Grammatical Relations. Cambridge, MA: MIT Press.Morgan, J. L. (1986) From Simple Input to Complex Grammar. Cambridge, MA: MIT Press.Naigles, L. (1990) Children use syntax to learn verb meanings. Journal of Child Language

17:357–374.Naigles, L. (1996) The use of multiple frames in verb learning via syntactic bootstrapping.

Cognition 58:221–251.Naigles, L. G. and Kako, E. T. (1993) First contact in verb acquisition: defining a role for

syntax. Child Development 64(6):1665–1687.Naigles, L., Fowler, A., and Helm, A. (1992) Developmental shifts in the construction of

verb meanings. Cognitive Development 7:403–427.Pinker, S. (1989) Learnability and Cognition. Cambridge, MA: MIT Press.Rappaport, M. and Levin, B. (1988) What to do with theta-roles. In W. Wilkins (ed.),

Syntax and Semantics, volume 21: Thematic Relations. New York: Academic Press.Rispoli, M. (1991) The mosaic acquisition of grammatical relations. Journal of Child

Language 18:517–551.Ritter, E. and Rosen, S. T. (1993) Deriving causation. Natural Language and Linguistic

Theory 11:519–555.Schieffelin, B. (1985) The acquisition of Kaluli. In D. Slobin (ed.), The Cross-Linguistic Study

of Language Acquisition: The Data. Hillsdale, NJ: Erlbaum.Siskind, J. (1996) A computational study of cross-situational techniques for learning

word-to-meaning mappings. Cognition 61:39–91.Waxman, S. R. and Markow, D. B. (1995) Words as invitations to form categories:

Evidence from 12- to 13-month-old infants. Cognitive Psychology 29:257–302.

290 Cynthia Fisher

Chapter 17Perception of Persistence: Stability and ChangeThomas F. Shipley

Psychology has come to be seen by many as a fragmented disciplinewith apparently few core concepts that span the field. However, oneconcept that seems to show up at many levels is identity. Throughouttheir careers both Henry and Lila Gleitman have grappled with theproblem of defining when an organism will treat two things as thesame—psychologically identical. In research ranging from rats runningaround in mazes to children learning a language, these two psycholo-gists (with a little help from their friends) have sought to provide ac-counts of psychological identities.

To illustrate the pervasive nature of identity problems in psychology,consider some research from Henry and Lila’s past. In studies of howrats find their way around an environment, Henry has shown thatbeing passively moved through a maze allows animals to successfullyrun through a maze with the same shape (Gleitman 1955). Successfulperformance in the second maze requires that the rat treats the twomazes as identical, despite the differing motor behavior. In the samevein, categorization is basically a problem in establishing identities;Henry and Lila have worried about how one decides whether or nottwo objects belong in the same category (Armstrong, Gleitman, andGleitman 1983), and which of two categories (number or letter) will beused for one object (an “0”) (Jonides and Gleitman 1972). Henry andLila have also addressed one of the central problems of language learn-ing: How does a child identify words and phrases that mean the samething? To learn the meaning of a word, a child must solve the matchingproblem—what words go with what events in the world. Their work onverb frames, which shows that children can infer causal properties of anovel verb when the verb is presented in a familiar sentence frame, of-fers an important clue into how children solve this identity problem(Gleitman and Gleitman 1992; Naigles, Gleitman, and Gleitman 1993).Finally, my own dissertation research with Henry on perceptual unitformation also addressed an identity problem. I was interested in howtwo objects could appear to have the same shape when one was fully

visible and the other partially occluded. In this chapter I review somerecent work on the perception of identity over time that grew out of thisquestion.

Identity and Perception

In perception the two most familiar examples of identity problems arerecognition—how we decide we are looking at something we have seenbefore—and the perceptual constancies (e.g., size, distance, and light-ness constancy). When viewing a scene over time both processes are ev-ident; the size, shape, and color of most objects appear unchanging overtime, and an object will be recognized as the same one that occupiedthat location several seconds ago. These impressions of stability holdeven as we move through the environment. When we drive and lookout upon the road we are approaching (or in Henry’s case, with hispropensity for talking to whoever is in the back seat, the road where hehas just been), the size, shape, and spacing of objects remain the same,despite a changing viewpoint. How could we possibly see stable quali-ties given the massive changes that occur in the retinal image wheneverwe move? Stability is achieved by taking advantage of the fact that thechanges are not random in nature, and using the regularities in the pat-tern of change to identify that which remains unchanged (Gibson 1979).

Some of the earliest work on perceiving stable qualities in changingarrays focused on how dynamic changes provide information for three-dimensional spatial relations. Hans Wallach, one of Henry’s colleaguesat Swarthmore, described the aspects of dynamic two-dimensional dis-plays that are necessary to perceive the three-dimensional shape of ob-jects (Wallach and O’Connell 1953). Each static image from a motionsequence may look quite different since the two-dimensional distancesbetween an object’s parts vary considerably in projected images, butwhen animated, a moving object with a stable shape is seen. For exam-ple, in biomechanical motion displays (e.g., point-light walkers), likethose used by Johansson (1973), the appearance of human forms doesnot occur until the elements move. The pattern of element motions al-lows the global form (a human) to be seen. The visual processes respon-sible for perceiving structure from motion may be present whenever wemove through the environment and thus play a central role in the ap-parent stability of the world.

The dynamic information does not need to be continuously availablefor a stable three-dimensional form to be seen (Michotte, Thines, andCrabbe 1964). Brief periods of occlusion do not affect the apparent sta-bility of an object. Henry’s driving illustrates this quite clearly, and hiscalm, while facing rearward, reveals the compelling and potentially er-

292 Thomas F. Shipley

roneous nature of this impression of stability. In this situation, Henry’slack of concern about not being able to see where he is going does notreflect an absence of imagination, but rather the impression (or, onemight say the conviction) that the world does not change simply be-cause one has changed one’s view. Objects don’t cease to exist simplybecause they are not visible. Two general classes of explanations havebeen offered to account for this stability over time. The first and morewidely accepted is based on representations in memory, the other onpatterns of change that indicate stability.

Internal representationsThe phenomenal persistence of objects, even when they are momen-tarily out of view, has led many researchers to propose, explicitly orimplicitly, a memory that contains representations of all objects in ascene, for example, object files (Treisman and Gelade 1980) and visualbuffers (McConkie and Rayner 1976). Stability is achieved by match-ing the present image of the world, with its various visible pieces, tothe objects in memory. This type of approach has found broad sup-port, perhaps because it is consistent with our phenomenal experienceof the visual world extending all around us, even in regions where wehave few or no receptors. If objects appear stable and continuouslypresent despite their sensory absence, something inside the organism(i.e., the representation in memory) must be stable and continuouslypresent.

Change as information for stabilityAlthough it might be tempting to believe the visual system maintainsrepresentations of all aspects of the environment, this is not necessarysince the environment changes in lawful ways. The visual system isconstructed to operate in a world where objects don’t change as a func-tion of the viewer’s direction of gaze, or with the presence of interven-ing objects. The visual system does not need to store a copy of an objectif the object will be there to reexamine when necessary. A representationof an object is not needed to perceive the object as stable over time ifthere is information that that object persists even when not in sight.

Theoretical alternatives to a representation-based approach havebeen offered by both Gibson and Michotte. Michotte et al. (1964) arguedthat stability was a perceptual phenomenon—the experience of stabilitywas a consequence not of a memory for the object, but of some aspect of the stimulus. For the display illustrated in figure 17.1, observers al-most uniformly experience a circle changing visibility, although inprinciple one could see a form changing shape. An additional aspect ofthis display that may be relevant for understanding stability is that a

Perception of Persistence 293

second boundary, an edge that hides the circle, is seen. This edge has aphenomenal quality similar to the ones seen in illusory figure dis-plays.

Gibson, Kaplan, Reynolds, and Wheeler (1969) identified the charac-teristic pattern of change that occurs whenever an object disappearsfrom view as the aspect of the stimulus responsible for the appearanceof continued existence. The pattern associated with occlusion differsfrom the pattern observed with changes in existence (such as drying up,exploding, or corroding). Distinguishing between a circle changing vis-ibility and one changing shape requires only that the visual system beable to distinguish between the patterns of change that occur in the twocases.

Evidence against Internal Representations

Aside from the enormous burden a memory-based scheme seems toplace on the visual system, this approach has difficulty explaining someof the recent work on the perception of persistence. There are a numberof observations that suggest humans are much less sensitive to changethan one might think. In each case some aspects of a scene are remem-bered; however, the finding of particular interest is that these represen-tations appear to be quite impoverished. Substantial changes can bemade in a scene in such a way that the scene appears stable—the phe-nomenal stability in each case is illusory.

294 Thomas F. Shipley

Figure 17.1.An illustration of Michotte’s kinetic screen effect (figure similar to figure 1 in Shipley andKellman 1994).

Changes that occur during saccadesA wide variety of changes can occur in text while the eyes are in motion(e.g., changes in case such as replacing “eStUaRiEs” with “ EsTuArIeS”)with little effect on reading, and the reader is generally unaware thatany change has occurred (see, e.g., McConkie and Zola 1979). Using asimple procedure, moving a picture to produce eye movements,Blackmore, Brelstaff, Nelson, and Troscianko (1995) have shown thatpeople are similarly unable to detect changes in natural scenes thatoccur during saccades. When a picture is shown, removed, and then analtered version of the picture displayed next to the original position,subjects fail to detect the alteration. For example a chair might be re-moved from a scene with three chairs. The change in spatial position ofthe picture—and the resulting saccade—were necessary for this effect.Alterations were readily detected when the picture did not shift loca-tion.

Luminance maskingSubjects’ ability to report changes in large arrays of familiar elements(e.g., letters) is also quite limited (Pashler 1988). Subjects’ ability to re-port which letter changes in an array of ten letters is close to the level ex-pected on the basis of full report studies when the altered array appearsmore than 150 msec after the original array disappears. Accuracy levelsfor detecting a single change in a ten-item array was consistent withsubjects remembering about four items, and using those four to com-pare the old and new array. A similar inability to detect changes wasfound at shorter intervals when a luminance mask was inserted be-tween the target and altered array. Recently Rensink, O’Regan, andClark (1996) reported a similar finding for natural scenes. They usedpairs of pictures in which some aspect of the picture was altered (e.g.,the engine of an airplane was present in one picture and not in theother). Subjects were very slow to detect the differences between thepictures when a luminance mask (a grey field) was presented betweenthe first picture’s offset and the second picture’s onset. Subjects ap-peared to be serially searching the picture for the change, since the timeto detect the change was directly related to the order that the changeditem showed up in a verbal description of the image.

Continuity errorsWhen movies are filmed, scenes that will immediately follow eachother in the final movie are often filmed at different times. Such a prac-tice can result in a “continuity error,” when some detail in a scenechanges across a cut. A classic continuity error is the disappearance of

Perception of Persistence 295

Noah around the forty-second minute of “The Grapes of Wrath.” Oneminute he is part of the party traveling to California, and the next he isgone, never to return. This was not noticed by most viewers, and in gen-eral, continuity errors are not noticed by audiences (Levin and Simons1997). Simons has brought this phenomenon into the laboratory. Sub-jects shown a brief film in which objects in the scene change across cuts(e.g., a two-liter soda bottle was replaced by a box) consistently fail tonotice anything wrong (Simons 1996). It is even possible to change thecentral character in a story, and if the change occurs between cuts, sub-jects will fail to note the change in identity of the actor in their descrip-tions of the story (Levin and Simons 1997). Recently Simons and Levin(1997) extended this work to real-world interactions. They found thatchanges in the identity of a person are detected less than half the timewhen the change occurs during occlusion (e.g., by an object passing be-tween two people engaged in a conversation).

“The world as visual memory”The phenomenal experience of an extended visual field in which theboundaries of objects appear clearly defined and surface characteristicsare clear have led theorists to assume that perception depends on repre-sentations that capture all of the apparent richness of a scene. Illusorystability presents a problem for such accounts: Why can’t the visual sys-tem use its representations to detect changes by comparing the presentvisual image with the past image? A number of researchers and philoso-phers have used these finding to argue that models requiring detailedrepresentations of the visual world must be abandoned (e.g., Dennett1991; O’Regan 1992). Massive representational edifices are not neededsince the world is always there available to be consulted as needed. Ifthe world serves as the visual store then only minimal representationsneed be maintained by the perceiver.

If, as Gibson and Michotte claim, change and stability can be discrim-inated on the basis of stimulus properties, then observers may rely onthe fact that they can detect changes as they occur (e.g., changes maystimulate motion detectors), and representations of the previous state ofthe world are not required. On such an account, stability is not the resultof a psychological process, but a consequence of the way the system isconstructed. In the absence of perceptual evidence for change, stabilityis the default; we do not actively perceive that the world is unchanging.

On such an account, perception is an active process in which atten-tion guides the pick-up of whatever information is needed for the taskat hand. Any role for representations in perception is then limited toguiding attention and the ongoing task. However, attention cannot becontrolled solely by the observer. As noted by Neisser (1976), any model

296 Thomas F. Shipley

of perception that relies substantially on internal guidance would besusceptible to problems inherent in too much assimilation: If perceptionis guided by the organism alone, how can it detect and process unex-pected events? Furthermore, Yantis and colleagues have found thatabrupt appearances of new objects attract attention (Yantis and Jonides1984; Yantis 1993). Attention must be guided by an interaction betweenthe organism and the environment; the pick-up of information will bedetermined by the observer’s intentions and expectations, as well as bysome events in the world (e.g., abrupt changes in luminance and thesudden appearance of objects).

Illusions of stability are also problematic for theories of perceptionbased on patterns of change. These theories must provide some accountof why the change evident in all of the examples of illusory stabilitycited above is not picked up. Note that the important question here isnot why the world appears stable in each case, but how these cases dif-fer from everyday experience where we reliably distinguish changefrom stability.

One thing common to all the illusory-stability cases is that massivemotion signals are present when the undetected change occurs. Motionsignals occur when the eye moves, and when luminance levels changeabruptly (as would occur whenever one image is replaced by a differentimage). Perhaps these motion signals interfere with detecting the pat-tern of changes that would normally be experienced as changes in theworld. In support of such a hypothesis, consider one more example of afailure to detect change.

Occlusion and Object Constancy

Even very young children appear to treat an object that has disappearedfrom view as continuing to exist (Baillargeon 1987). They are surprisedif an object is hidden and does not reappear when the occluding surfaceis removed. However, if the object is not the focus of attention, objectconstancy may not be seen. Douglas Cunningham and I created a videotape in which five objects moved back and forth five times, and halfwaythrough the tape, one of the objects did not return after a brief period ofocclusion. Figure 17.2 shows four frames from this video. When sixtysubjects were shown the tape, introduced as an example of motion par-allax, none of the subjects spontaneously reported that one of the ob-jects disappeared. When asked if they noticed anything odd about thevideo, only one subject noted the change.

Unlike the other examples of illusory stability, this example does notcontain motion signals spread over the entire visual field. Here, the mo-tion signals that do occur appear in a pattern that is consistent with

Perception of Persistence 297

298 Thomas F. Shipley

Figure 17.2.Four frames from a video sequence of five objects moving back and forth. The most dis-tant object (a small cardboard box) disappears and reappears (images 1, 2, and 3) initially,but then disappears and does not reappear (image 4).

occlusion. As a result, they are not treated as information for change.The pattern of local changes in this display are all consistent with a sta-ble world, so the disappearance of an object is not detected.

To investigate further the role of change in the perception of occlusionand stability, we employed displays in which an occluding form is dy-namically specified. A moving form with well-defined boundaries isseen in displays in which the elements of a sparse texture field changein a systematic manner (Shipley and Kellman 1994). For example, anopaque surface will be seen if elements disappear along the leadingedge of a moving form that is the same color as the background, andthen reappear at its trailing edge (see figure 17.3). No bounded form isseen in static frames of such displays.

An important aspect of these displays is that phenomenally, only theforward, occluding surface appears to move; the small background ele-ments appeared stable. From the perspective of understanding percep-tual stability, this is notable since the spatial and temporal relationshipsbetween the appearance and disappearance of the background ele-ments would, if elements were presented in isolation, result in apparent

Perception of Persistence 299

Figure 17.3.Three frame sequence illustrating dynamic occlusion. The dotted square represents an in-visible form moving over the array of elements. Elements are only visible (black) whenthey are outside the form; they are invisible (gray) inside the form (figure similar to figure2 in Shipley and Kellman 1994).

motion—when one element disappears and another appears, motionbetween the two locations is normally experienced. If the apparent sta-bility in dynamic occlusion displays is a result of the same perceptualprocesses that result in illusory stability, then we may understand per-ceptual stability by understanding the perceptual processes responsiblefor dynamic unit formation.

To test for illusory stability in dynamic occlusion displays, we em-ployed a free report procedure (Shipley, Cunningham, and Kellman1994). We asked subjects to describe what they saw in displays in whichthe background elements either changed position, or returned to theiroriginal position, following occlusion. In one type of display, simulat-ing an opaque form, elements were invisible while inside a moving cir-cular region. We also included displays that simulated a wire circle(elements were invisible for only a brief period of time—66 ms) and dis-plays that simulated transparency (elements changed to red inside thecircle). In half of the displays, elements reappeared where they disap-peared (in the case of transparency they did not change location whenthey changed color), as they would if an occluder had actually passedover them. In the other half of the displays, elements reappeared in anew location following occlusion (in the case of transparency, elementschange location when they changed color). None of the ten subjects re-ported any difference between the displays where elements reappearedin their original location and displays where elements reappeared innew locations (even in the wire and transparency displays where thetemporal gap between old and new locations was minimal). The ele-ments in both sets of displays appeared stable.

To test that occlusion was critical for this illusory stability we asked anew set of subjects to describe six control displays where an occluderwas not seen and elements either changed location or stayed in thesame location. Each control was created using two intermediate framesfrom each of the previous displays. For the two occlusion controls, ele-ments within a circular region disappeared for 167 ms (the average timeelements were invisible in the dynamic occlusion display), and then appeared, either in the same location or in a new location. For the trans-parency controls there was no temporal gap between changes—ele-ments changed to red, appearing either in the same or in a new locationfor 167 ms, and then returned to their original color. For the wire figurecontrols, elements in a circular ring disappeared for 66 ms before reap-pearing in either an old or new location. Subjects could detect changesin element location in these displays, where occluders are not seen.They had no difficulty discriminating displays in which elementsstayed in the same location from displays in which elements changedlocation. Eight out of ten subjects reported that the elements appeared

300 Thomas F. Shipley

to move in at least one of the displays in which elements changed loca-tion.

A Motion-Based Model of Stability and Change

Philip Kellman and I recently developed a model of boundary forma-tion in dynamic displays that may help account for the apparent stabil-ity of dynamic occlusion displays (Shipley and Kellman 1997). Themodel is based, in part, on principles developed in our model of staticunit formation (discussed by Kellman in his chapter for this volume).The dynamic unit formation model uses the pattern of motion signalsthat occur over time to define the boundaries of moving objects. As aconsequence, it can offer a description of the pattern of motion thatidentifies changes in visibility. Below I review some of our recent workthat indicates the visual system uses motion signals defined by sequen-tial occlusion events to perceive a moving surface.

Motion signals as information for boundariesThe background elements in dynamic occlusion displays lose their phe-nomenal stability as frame duration increases. At short frame durationsthe elements appear stable, while at longer durations the elements ap-pear to move. The clarity and phenomenal presence of a moving surfacealso decreases as the duration of each frame increases (Shipley andKellman 1994). We used accuracy in a ten-alternative shape identifica-tion task to access figural clarity; the effect of varying frame duration onboundary formation is shown in figure 17.4. In earlier work on apparentmotion Sigman and Rock (1974) and Petersik and McDill (1981) noted asimilar relationship between appearance of an occluding edge and theapparent stability of elements: In all cases, when a moving form is seen,the background appears stable, and when no form is seen, the elementsappear to move. This suggests that the visual processes responsible forseeing the edges may incorporate local motion signals that occur at theoccluding edge. As a consequence, motion signals are not consciouslyexperienced as motion in the world when they define a boundary, butwhen no boundary is formed we see the individual motion signals.

To test the hypothesis that motion signals are used to perceive a mov-ing boundary, we asked a fairly simple question: What happens to theperception of boundaries when additional motion signals that do not fitthe pattern produced by the moving form are added (Shipley andKellman 1997)? Displays consisted of a form translating over an array ofstationary elements while eighteen elements rotated around the centerof the screen. The motion signals generated by the rotating elementsproved to be very effective at disrupting shape perception; subjects’

Perception of Persistence 301

accuracies in identifying the translating form were much lower whenthe additional motion signals were present than when they were absent(figure 17.5). Furthermore, the effect of the additional motion signalsdid not depend on their global organization: Coherent motions inwhich all elements rotated in the same direction were as effective as ran-dom local motions. This suggests that the local motion signals them-selves were the cause of the disruption.

Motion signals are invariably present whenever one object occludesanother because abrupt changes in the visibility of elements along theedges of a moving opaque object will always result in local motion sig-nals. Local motion signals alone, however, are not sufficient to identifyocclusion since motions signals also occur when objects change shapeor location in the world. These two cases may be discriminated on thebasis of the pattern of local motion signals.

How do motion signals define a boundary?The pattern of motion that results from dynamic occlusion can be char-acterized by the pattern produced by the local occlusion of only three el-

302 Thomas F. Shipley

Figure 17.4.Shape identification accuracy plotted as a function of frame duration for three back-ground element densities. As density increases the number of changes per frame in-creases, and accuracy increases. As frame duration increases the number of changesthatoccur within a given temporal window decreases, and accuracy decreases. There was nointeraction between spatial and temporal density suggesting a fixed temporal integrationwindow (figure similar to figure 7 in Shipley and Kellman 1994).

ements (illustrated in figure 17.6a). Each pair of disappearances resultsin a motion signal. The magnitude and direction of that signal will be afunction of the spatial and temporal separation of changes. Thus localmotion signals combine spatial and temporal information about ele-ment changes. If the two vectors representing the motion signal have acommon origin, their tips define the orientation of the occludingboundary (Shipley and Kellman 1997). Thus the pattern of local motionsignals provides information about the local orientation of an edge, andthat the elements that disappeared were occluded.

To find out if observers are sensitive to the sequential pattern of mo-tion signals, we developed displays that were consistent with dynamicocclusion but contained degenerate motion patterns (Shipley and Kell-man 1997). In these displays, elements were arrayed so that the localmotion signals were sequentially similar in direction and magnitude(figure 17.7a illustrates a local edge segment approaching elements thatwhen covered would produce similar motion signals). Such a pattern isdegenerate because the orientation solution outlined in figure 17.6would be very sensitive to small errors or noise when the vectors have asimilar direction. Therefore the edge should be unstable and formrecognition should be compromised. Indeed subjects’ ability to identifythe shape of the form defined by sequentially similar motion signals

Perception of Persistence 303

Figure 17.5.Shape identification accuracy plotted as a function of background element density forfour conditions: No Motion, elements rotating in the Same direction as the target form, theOpposite direction, or in Random directions (figure similar to figure 4 in Shipley andKellman 1997).

304 Thomas F. Shipley

Figure 17.6.An illustration of sequential occlusion by a local edge segment. a) As an edge moves fromleft to right, it sequentially covers three elements. The local motion signals, v12 and v23,are defined by the sequential disappearance of elements 1 and then 2, and 2 and then 3,respectively. b) The orientation of the occluding edge is defined by the length and orien-tation of the local motion signals (figure similar to figure 5 in Shipley and Kellman 1997).

was severely impaired relative to displays with the usual randomly ori-ented motion signals (figure 17.7b illustrates a local edge segment ap-proaching a set of elements that is identical in their spatial arrangementrelative to the edge, but the sequence in which they will be occluded israndom). The phenomenal appearance of these displays was also con-sistent with our hypothesis that local motion signals will be experi-enced when not incorporated into a boundary. Occlusion was not seenin the sequentially similar motion displays; instead motion of the ele-ments was seen.

In addition to providing information about the continued existenceof an occluded surface and the shape of the occluding surface, motionsignal patterns may also provide information about the opacity of mov-ing surfaces. The pattern of motion signals produced by the movementof a partially transparent surface will resemble the one produced by anopaque surface. It will differ only in the magnitude of the temporal con-trast modulation, relative to the average contrast. As an initial test to seeif opacity could be dynamically specified, subjects were asked to de-scribe what they saw in dynamic occlusion displays from which all sta-tic information for surfaces was removed (Cunningham, Shipley, andKellman 1998). To remove static information while retaining the patternof change over time, we added a large number of unchanging elementsto a display in which elements disappeared as a form moved around thescreen. One might conceive of such a display as two fields of elementswith a form moving between the two. From the point of view of the ob-server only some of the elements disappear and reappear (the ones inthe more distant field), so elements are seen both inside and outside themoving surface. This effectively masks static information for a surfacehiding background elements. When the form moved, subjects reportedseeing a dark surface with well-defined boundaries, and a specific loca-tion in depth (the form appeared to move between two layers of ele-ments). The pattern of changes over time must have been responsiblefor the perception of shape, opacity, and depth. We are currently inves-tigating whether subjects are sensitive to dynamic specification of de-gree of opacity.

In sum, the pattern of motion signals that result from elements ap-pearing and disappearing provides information about the occludingedge and about the continued existence of the elements. The standardaccretion and deletion displays appear to have stable backgrounds be-cause the local motion signals are integrated into the motion of the oc-cluding figure. When this does not occur (e.g., when there is a longpause between frames) or when edges are unstable, then the local mo-tions are experienced and no edge is seen. In occlusion-based displaysof illusory stability, the background appears stable because the motion

Perception of Persistence 305

pattern is consistent with a moving boundary—the motion signals areintegrated into the moving edge and are not interpreted as motion ofthe elements. Only when a boundary is not seen are changes in elementlocations noticed.

The Role of Attention

Finally, although subjects in our experiments do not seem to be sensi-tive to changes in location following occlusion, there are a number of perceptual phenomena for which observers do appear to maintainsome representation over time that includes spatial position. Michotte’sdemonstration of tunneling is one example (Michotte et al. 1964). InMichotte’s displays a moving dot disappeared behind an occluder andthen reappeared on the other side. If the dot reappeared at a time and lo-cation consistent with a smooth continuous path behind the occluder, itappeared to continue to exist while out of sight. In contrast if the dot ap-peared at some other location or after a very long (or short) interval,subjects reported seeing two dots—one that disappeared and one thatappeared.

How does tunneling, in which the percept is sensitive to spatialchanges during occlusion, differ from the displays presented here? Onepossibility is that the difference lies in the number of elements. Alterna-tively, as suggested earlier, attention may play an important role in

306 Thomas F. Shipley

Figure 17.7.An illustration of a set of elements that when occluded will produce (a) similar and (b)random motion signals. a) Each element is shifted off the line defined by the previoustwo elements by 6 degrees. When a moving edge (indicated by a grey dashed line) oc-cludes these elements the sequential motion signals will be similar in magnitude and ori-entation. b) These elements have the same location, relative to the occluding edge, as inFigure 7a but order of occlusion has been randomized so sequential motion signals willdiffer in orientation.

detecting change when the pattern of motion signals cannot be used.Object permanence in occlusion displays may require attention to a par-ticular object (or set of objects). Indeed, it is possible to see the changesin illusory-stability displays if the object that changes is the focus of at-tention. However, as noted by Levin and Simons (1997), attention to theobject that changes is not sufficient—not all aspects of an object may berepresented. The appearance of illusory stability in the examples dis-cussed previously does appear to change with experience. For example,once the changes in Simons’s and Rensink et al.’s displays have beenseen, they are almost immediately noticed when shown a second time.

Conclusion

Recent interest in illusory stability seems to reflect a hope that it willhelp us with a long-standing problem: How does perception relate toour conscious experience of the world? These particular illusions mayhave captured attention because the mismatch between reality and con-scious experience is large and (for many accounts of perception) shouldbe noticed. I have argued here that the psychological identity of objectsover time is based on local motion information. However, the relation-ship between local motion signals and the conscious experience ofsomething in the world going out of sight or changing shape is not di-rect. Information for stability and change are, to use Köhler’s term,“Ehrenfels qualities” (Köhler 1947). It is not motion per se that distin-guishes persistence from change, but rather the pattern of motion sig-nals: One pattern tells us about how things are changing in the world,and another pattern tells us that things are stable. So, the perception ofthe here and now depends on both the way things appear at the mo-ment, and how things are changing over time.

Acknowledgments

The research and preparation of this manuscript were supported byNSF Research Grant BNS 93–96309. I would like to thank John Jonidesand Daniel Reisberg for their extensive feedback on an earlier version ofthis chapter.

ReferencesArmstrong, S. L., Gleitman, L. R., and Gleitman, H. (1983) What some concepts might not

be. Cognition 13:263–308.Baillargeon, R. (1987) Object permanence in 3 1/2- and 4 1/2-month-old infants.

Developmental Psychology 23:655–664.

Perception of Persistence 307

Blackmore, S. J., Brelstaff, G., Nelson, K., and Troscianko, T. (1995) Is the richness of ourvisual world an illusion? Transsaccadic memory for complex scenes. Perception24:1075–1081.

Cunningham, D. W., Shipley, T. F., and Kellman, P. J. (1998) The dynamic specification ofsurfaces and boundaries. Perception 27:403–415.

Dennett, D. C. (1991) Consciousness Explained. Boston: Little, Brown.Gibson, J. J., Kaplan, G., Reynolds, H., and Wheeler, K. (1969) The change from visible to

invisible: A study of optical transitions. Perception and Psychophysics 5(2):113–116.Gibson, J. J. (1979) The Ecological Approach to Visual Perception. Hillsdale, NJ: LEA.Gleitman, H. (1955) Place learning without prior reinforcement. Journal of Comparative and

Physiological Psychology 48:77–89.Gleitman, L. R. and Gleitman, H. (1992) A picture is worth a thousand words, but that is

the problem: The role of syntax in vocabulary acquisition. Current Directions inPsychological Science 1(1):31–35.

Johansson, G. (1973) Visual perception of biological motion and a model for its analysis.Perception and Psychophysics 14:201–211.

Jonides, J. and Gleitman, H. (1972) A conceptual category effect in visual search: O as a let-ter or a digit. Perception and Psychophysics 12:457–460.

Köhler, W. (1947) Gestalt Psychology. New York: Liveright PublishingLevin, D. T. and Simons, D. J. (1997) Failure to detect changes to attended objects in mo-

tion pictures. Psychological Bulletin and Review 4(4):501–506.McConkie, G. W. and Rayner, K. (1976) Identifying the span of the effective stimulus in

reading: Literature review and theories of reading. In Theoretical Models andProcessing in Reading, ed. H. Singer and R. B. Ruddell. Newark, Del.: InternationalReading Association, 137–162.

McConkie, G. W. and Zola, D. (1979) Is visual information integrated across successivefixations in reading. Perception and Psychophysics 25(3): 221–224.

Michotte, A., Thines, G., and Crabbe, G. (1964) Les complements amodaux des structuresperceptives. Studia Psycologica. Louvain: Publications Universitaires de Louvain.(English translation given in: Michotte, A. [1991] Michotte’s experimental phenome-nology of perception, ed. and trans. by Thines, G., Ccostall, A., and Butterworth, G.,pp. 140–169. Mahwah, NJ: Erlbaum.)

Naigles, L., Gleitman, H., and Gleitman, L. R. (1993) Children acquire word meaningcomponents from syntactic evidence. In Language and Cognition: A DevelopmentalPerspective, ed. E. Dromi. Norwood, NJ: Ablex, 104–140.

Neisser, U. (1976) Cognition and Reality. New York: Freeman.O’Regan, J. K. (1992) Solving the “real” mysteries of visual perception: The world as an

outside memory. Canadian Journal of Psychology 46(3):461–488.Pashler, H. (1988) Familiarity and visual change detection. Perception and Psychophysics

44:369–378.Petersik, J. T. and McDill, M. (1981) A new bistable motion illusion based upon “kinetic

optical occlusion.” Perception 10:563–572.Rensink, R. A., O’Regan, J. K., and Clark, J. J. (1996) To see or not to see: The need for at-

tention to perceive change in scenes. Investigative Ophthalmology and Visual ScienceSupplement 37(3):S978.

Shipley, T. F. and Kellman, P. J. (1994) Spatiotemporal boundary formation: Boundary,form, and motion perception from transformations of surface elements. Journal ofExperimental Psychology: General 123(1):3–20.

Shipley, T. F. and Kellman, P. J. (1997) Spatiotemporal boundary formation: The role oflocal motion signals in boundary perception. Vision Research 37(10):1281–1293.

308 Thomas F. Shipley

Shipley, T. F., Cunningham, D. W., and Kellman, P. J. (1994) Perception of stability in dy-namic scenes. Paper presented at the 35th Annual Meeting of The PsychonomicSociety, St. Louis, November 1994.

Sigman, E. and Rock, I. (1974) Stroboscopic movement based on perceptual intelligence.Perception 3:9–28.

Simons, D. J. (1996) In sight, out of mind: When object representations fail. PsychologicalScience 7(5): 301–305.

Simons, D. J. and Levin, D. T. (1997) Failure to detect changes to attended objects.Investigative Ophthalmology and Visual Science Supplement 38(4):S707.

Treisman, A. M. and Gelade, G., (1980) A feature-integration theory of attention. CognitivePsychology 12:97–136.

Wallach, H. and O’Connell, D. (1953) The kinetic depth effect. Journal of ExperimentalPsychology 45(4):205–217.

Yantis, S. and Jonides, J. (1984) Abrupt visual onsets and selective attention: Voluntaryversus automatic allocation. Journal of Experimental Psychology: Human Perceptionand Performance 10:601–621.

Yantis, S. (1993) Stimulus-driven attentional capture. Current Directions in PsychologicalScience 2(5):156–161.

Perception of Persistence 309

Chapter 18

Putting some Oberon into Cognitive Science

Michael Kelly

In directing A Midsummer Night’s Dream a few years ago, HenryGleitman cast the same actor in the roles of Theseus and Oberon. Theformer represents rationality distilled to its essence, the scientist Apollowho grasps as much as “cool reason ever comprehends,” but no further.The latter is the artist Dionysus, imaginative beyond reason, but conse-quently self-indulgent, undisciplined, and lazy in the natural luxury ofhis forest realm. These two figures, reason and imagination, are failures,dead ends that cast a shadow of pessimism over the celebration at theend of the play. I never understood this impression fully until I experi-enced Henry’s version. Throughout the play, Lysander, Demetrius,Hermia, and Helena fret and scheme and argue and moan about whomthey love and who should love them. In the end, though, all the piecesseem in place and the lovers twitter happily. However, Henry’s insight-ful casting made clear that one couple is still divorced: Theseus andOberon.

In the world of the play, Theseus and Oberon will never be united.After all, they don’t even seriously acknowledge each other’s existence.In contrast, Henry Gleitman has consistently rejected a fundamentalopposition between science and art. In Henry’s educational philosophy,one student might enroll in college as a premed Theseus and another asan actor Oberon. Given the proper environment, each student shouldnonetheless graduate as a “Theseron,” and be better doctors and actorsas a consequence.

Henry has practiced this philosophy throughout his teaching career.In his text Psychology, he frequently uses art to illustrate psychologicalprinciples. However, as might be expected from someone who hasworked on the concept of symmetry (Gleitman, Gleitman, Miller, andOstrin 1996), Henry is aiming for reciprocal effects here by encouragingstudents to think about artwork in a novel way. In his seminar on thepsychology of drama, Henry brings together psychology majors andstudents of theater and literature. As the students struggle to communi-cate and understand their diverse perspectives on Hamlet and Othello,

they develop an appreciation of human achievement that is bothbroader and deeper than could have been attained in a class that sepa-rated the arts from the sciences.

When you see Henry carving out a complex ANOVA as though it’ssome kind of classical sculpture, you realize that the distinction be-tween art and science is as meaningless to him in research as in teach-ing. Though she differs from Henry on many other issues, like theworthiness of various activities to be deemed sports, Lila Gleitman hasthe same attitude. Indeed, they both live that view to the hilt. It’s hard tothink of two people who merge so much passion for their objet d’art—language—with analytical talents that are relentless in determininghow it’s learned, and then put to use in both work and play.

In keeping with this theme of science and art . . . well, if not united, atleast aligned in “fearful symmetry,” I will in this chapter present someexamples of how cognitive principles can illuminate certain aspects ofcreative language use. The examples are far from exhaustive; they aremore like a sampling of cheeses at the Gleitman research seminar.However, the topics do correspond roughly with aspects of languagethat Henry and Lila have examined over the years, such as lexical andphrasal stress (Gleitman and Gleitman 1970; Gleitman, Gleitman,Landau, and Wanner 1988), orthography (Gleitman and Rozin 1977),phrasal conjuncts (Gleitman 1965), and associative learning (Meier andGleitman 1967). I hope to show through these case studies that basic re-search in cognitive science can be applied productively to language in-novation, and might even be worthy of discussion in future versions ofHenry’s psychology of drama course.

The Rhythmic Structure of Verse

Like Shakespeare’s other poetry, the verse portions of A MidsummerNight’s Dream are generally written in iambic pentameter. This meterhas had a distinguished history in English literature because it formsthe rhythmic basis for much of our poetry, including the greatest worksof Chaucer, Shakespeare, and Milton. Given the prominence and preva-lence of iambic pentameter in English verse, poeticists have placed highpriority on understanding its structure.

A canonical line in iambic pentameter consists of five disyllabic feet,with each foot beginning with a weak beat and ending with a strongbeat. However, few lines actually fit this pattern perfectly. For example,in (1) the adjective “wise” appears in a weak position even though, asan open-class word, it should be prosodically salient.

(1) And, after that wise prince, Henry V (3HVI.III.iii.)

312 Michael Kelly

However, given its context, this positioning is understandable. In partic-ular, phrases like “wise prince” and “black bird” generally have aniambic rhythm in speech, and this rhythm is respected in poetry byaligning such phrases in weak-strong position. In contrast, compoundwords like “blackbird” are pronounced with a trochaic rhythm, and con-sequently are set in strong-weak position in verse (Kiparksy 1975, 1977).

This analysis assumes that the rhythmic structure of verse generallyrespects the prosodic principles of speech. This link could provide apowerful heuristic for proposing and testing hypotheses about poeticmeter. For example, spoken stress is associated with information value(see Levelt 1989, for summary). If this relationship is preserved in po-etry, then relatively informative words should appear in strong positionmore often than less informative words. For instance, marked adjec-tives like “short” are more informative than unmarked adjectives like“tall” in that they pick out a particular region of a dimension such asheight whereas the unmarked adjective often refers to the dimension asa whole. Thus a question like “How tall is Theseus?” does not presup-pose that Theseus is especially tall. However, the use of “short” wouldimply that Theseus is low on the height dimension (relative to some cat-egory, such as predemocracy Athenians). Given this information differ-ence between marked and unmarked adjectives, one would predict thatthe former should be more likely to appear in stressed position in po-etry. In Kelly (1989), I tested this hypothesis by examining where 17 di-mensional adjective pairs like short-tall, cold-hot, and smooth-roughappeared in the Shakespeare selections printed in Bartlett’s Quotations.Overall 70% of the uses of marked adjectives appeared in stressed posi-tion compared with 49% of unmarked adjectives. Furthermore, in 14 ofthe 17 pairs, the marked member was more likely to appear in stressedposition.

As another example of how informativeness might influence thealignment of words with poetic meter, consider (2). Theseus’s openinglines in A Midsummer Night’s Dream contain two instances of “moon,”with the first appearing in a stressed position and the second appearingin an unstressed position.

(2) Now, fair Hippolyta, our nuptial hourDraws on apace; four happy days bring inAnother moon: but, O, methinks, how slowThis old moon wanes! she lingers my desires. (MSD, I.i.1–4)

This difference might reflect prosodic effects of the given-new distinc-tion. In particular, Fowler and Housum (1987) found that the first oc-currence of a word in speech, corresponding with new information,

Putting some Oberon into Cognitive Science 313

receives more stress than the second occurrence, corresponding withgiven information. If this relationship between givenness and stress op-erates in poetry as well as spoken prose, then one might expect patternslike that shown in (2). A detailed test of this hypothesis remains to beperformed, but it further illustrates the manner in which our knowl-edge of prosody can be applied to verse.

Spelling and Stress

Proper names like “Claire” are often padded with extra letters that donot affect pronunciation but, like word-initial capitalization, provide adistinguishing mark for names (Carney 1994). This phenomenon is il-lustrated most clearly in homophones that involve proper and commonnouns such as /web/ and /faks/. In contrast with the common nouns“web” and “fox,” the surnames “Webb” and “Foxx” double the final let-ter. This distinction exploits creatively the oft-derided variability inEnglish orthography. In particular, when properly manipulated, spell-ings like “Penn” can make orthographic distinctions between homo-phones and mark certain words as particularly salient while at the sametime preserving the correct phonemic structure.

There are many distinctions that could be represented in the orthog-raphy by systematically selecting different spellings of a phoneme orphonemic sequence. Although the choice between single and doubleletters might be the most obvious method, others are available. For in-stance, word-final /k/ could be represented by “k” as in “kiosk” or “que”as in “burlesque.” Word-final /m/ can be spelled “m” as in “velum” or“mb” as in “succumb.” My students and I have recently argued that thelonger versions of such alternatives are used to represent lexical stress.Analyses of the English vocabulary have revealed that syllables endingin spellings like “que,” “mb,” and various letter doublings are morelikely to be stressed than syllables ending in “k,” “m,” and various lettersingletons (Verrekia 1996; Verrekia and Kelly 1996). Subsequent experi-ments documented that literate English speakers have learned these re-lationships and might use them in reading. For example, subjects aremore likely to pronounce disyllabic pseudowords with iambic stress if they are spelled “fofvesque,” “zertumb,” or “filrass” rather than“fofvesk,” “zertum,” or “filras” (Verrekia and Kelly 1996). Furthermore,disyllabic real words whose spelling patterns are consistent with theirstress patterns show advantages in naming over words that have incon-sistent relations between these domains. Thus trochaic words like “pel-let” and iambic words like “dinette” are named more quickly andaccurately than trochaic words like “palette” and iambic words like“duet” (Kelly, Morris, and Verrekia 1998).

314 Michael Kelly

Although we have claimed from such results that English spellingcan directly encode lexical stress, Lila Gleitman has often countered thatmany of the spelling patterns that we have studied correspond withmorphemes. Since the morphemic structure of a word has clear andwell-documented effects on stress (see Gleitman and Rozin 1977 for re-view), one might say that English spelling only affects stress indirectlythrough its representation of morphemes. For example, “ette” repre-sents a morpheme meaning small or diminutive. Furthermore, thismorpheme is usually stressed. Hence when readers encounter a pseu-doword like “rinvette,” the actual morpheme nested within it is recog-nized and its typical stress level assigned. No direct link betweenorthography and stress needs to be proposed.

Although the morphemic account works well for spelling patternslike word-final “ette” and “ee,” it has difficulty with other cases. For ex-ample, word-final /o/ is typically stressed when it is spelled as “eau”rather than “o,” but “eau” is not a morpheme. Consider also the mor-phemes /™bl/, meaning capable of a specified action, and /Ins/,meaning a state or condition. The former can be spelled using “able” or“ible” whereas the latter can be spelled with “ance” or “ence.” There isno known difference in meaning associated with the spelling alterna-tives, and yet Verrekia (1996) has shown that they do have conse-quences for stress. For example, she found in a dictionary analysis that65% of trisyllabic words ending in “ance” had stress on the second syl-lable whereas 67% of trisyllabic words ending in “ence” had stress onthe first syllable.

I could cite other evidence for a direct link between spelling andstress in English, but in many ways the clearest and most interesting ex-ample can be found in early editions of Milton’s Paradise Lost. Englishspelling in the seventeenth century was still far from standardized(Brengelman 1980), and hence texts from this and earlier periods oftencontain multiple spellings of a particular word. Milton’s works are noexception, and so early editions of Paradise Lost have alternations like“he-hee,” “me-mee” and “star-starr.” However, the variability in spell-ing choice is not random. Rather, the longer version of each pair is morelikely to appear in stressed positions in Milton’s verse (Darbishire1952). For example, I surveyed all instances of “he” and “hee” in anelectronic version of the first edition of Paradise Lost.1 Since the poemwas written in iambic pentameter, the pronoun was considered stressedif it appeared in even syllable positions and unstressed if it appeared inodd syllable positions. Whereas “hee” appeared in stressed positions61% of the time, “he” occurred in such positions only 27% of the time.Similar patterns can be found in other alternations. Thus “mee” and

Putting some Oberon into Cognitive Science 315

“starr” occurred in stressed positions 77% and 95% of the time respec-tively. In contrast, their shorter versions “me” and “star” occurred instressed positions 41% and 67% of the time. These spelling differencesclearly do not reflect morphemic differences but creatively link spellingto metrically strong positions in verse.

This systematic relation between stress and spelling could be used toexamine more fine-grained aspects of Milton’s meter. In general, how-ever, literature scholars have not performed detailed analyses ofMilton’s spelling variations because it is possible that their source is notthe poet himself, but his printers. Although Darbishire emphasizes themeticulous care with which Milton handled the publication of hisworks, Adams (1954) responds sarcastically, “This hypothesis [thatMilton was involved intimately in selecting between spelling options]puts blind Milton, his amanuenses, and his manuscript in the middle ofa busy printshop, adding and subtracting e’s, changing small letters tocaps and vice versa, altering spellings, correcting type fonts, and break-ing in upon the sweaty printers as the sheets were being run off, toloosen the forms and drag out or insert tiny bits of inky lead” (p. 87).More generally, authors in Milton’s time simply did not follow the type-setting of their manuscripts with much diligence or even concern.Furthermore, given Milton’s blindness at the time, he is more likely tohave proofheard rather than proofread Paradise Lost.

In considering the spellings of words in Paradise Lost, we should notbecome excessively distracted by who precisely added an “e” or dou-bled an “r.” Suppose, for the sake of argument, that the printers and notthe author were responsible for the spelling variants in Paradise Lost.One could still argue that Milton was their ultimate source. In particu-lar, after reading thousands of lines of Milton’s verse, the printers mayhave abstracted schematic knowledge of his meter. This knowledgemight then have subtly influenced spelling choices. If so, then we couldstill use spelling variability to infer characteristics of iambic pentameterin general and Milton’s use of it in particular. For example, when “hee”does appear in unstressed positions, its distribution is not random.Instead, it occurs most often in the first syllable of a line. This positionmakes sense given that the opening beat in iambic meter is more likelyto be stressed than other odd locations (Newton 1975). As another ex-ample, consider (3):

(3) Thine shall submit, hee over thee shall rule. (PL IX.196)

Even though “hee” occurs in a position that is typically unstressed iniambic pentameter, the longer spelling may have been chosen becauseof the contrast with “thee,” and such contrasting situations are associ-ated with prosodic prominence (Selkirk 1984).

316 Michael Kelly

In sum, spelling variability should not necessarily be judged deroga-tively, as a sign of sloppiness in the orthography or its users. English or-thography can and does encode more than phonemic information.Indeed, its flexibility allows one to represent morphology, stress,salience, gender,2 and perhaps other factors without sacrificing its abil-ity to represent segmental phonology. Consequently, systematic variabil-ity (i.e., creativity) in spelling, both synchronically and diachronically,could be a rich source of evidence for testing diverse hypotheses aboutlanguage structure and use.

A Verb by Any Other Name?

Toward the end of an especially festive affair in Lower Merion, I over-heard a guest say, “They sure out-gleitmaned themselves this time,”meaning that the hosts had surpassed their own benchmark standardsfor throwing parties that illustrate every chapter in a psych 1 text:Sensation, learning, social cognition, maybe even psychopathology andits treatment with food, wine, and engaging company. These eventshave also spawned a large catch of linguistic novelties, such as the useof “Gleitman” as a verb.

Extending the usage of a word into another grammatical class is acommon form of lexical innovation in English, as Clark and Clark(1979) documented in their classic study of denominal verbs. For exam-ple, nine of the top twenty animal nouns in Battig and Montague’s(1969) category dominance norms have verb uses listed in The AmericanHeritage Electronic Dictionary. However, mere frequency is not necessar-ily a sign of unprincipled promiscuity. As Clark and Clark first showed,many factors can influence the likelihood with which a word will joinanother grammatical class. For instance, nouns seem to be blocked fromdeveloping verb uses if their new meaning would be synonymous withan existing verb. Thus many vehicle terms are used as verbs to mean “totravel by X,” where X is the vehicle. However, despite its high noun fre-quency, “car” has not acquired a verb usage. Clark and Clark arguedthat “car” has been kept out of the verb category because its moststraightforward verb meaning would be synonymous with “drive,”and speakers have a bias against the existence of synonyms.

Most investigations of grammatical category extensions have focusedon semantic and pragmatic factors that constrain their use (e.g., Clarkand Clark 1979; Kelly 1998). This orientation is consistent with moregeneral work on nouns and verbs that emphasize their semantic differ-ences (e.g., Langacker 1987; Pinker 1989). However, analyses of theEnglish lexicon have shown that these classes can also contrast phono-logically. Thus English nouns and verbs differ in stress patterns, vowel

Putting some Oberon into Cognitive Science 317

distributions, and the number of syllables they contain (see Kelly 1992,for review). These distinctions are so informative that formal classifica-tion models can learn to assign words to the noun and verb categorieswith high accuracy using only phonological information (Kelly, in prep-aration).

I will focus here on a stress difference between English nouns andverbs and examine its implications for denominal verb and deverbalnoun formation. Whereas the vast majority of disyllabic English nounshave first-syllable, or trochaic, stress, most verbs have second-syllable,or iambic, stress. This contrast can be best illustrated by contrasting thestress patterns of certain noun-verb homographs like “record,” “con-test,” and “permit.” In all cases where noun and verb homographs dif-fer in stress, the noun version has a trochaic pattern and the verbversion has an iambic pattern (Sherman 1975).

Many studies have shown that native speakers (and for that matter,nonnatives; Davis and Kelly 1997) have implicitly learned the noun-verb stress difference. Most relevant here is a study in which subjectslistened to a series of disyllabic pseudowords that varied in stress (Kelly1988). After hearing each word, the subjects were asked to use it in asentence. The stress patterns of the pseudowords affected the grammat-ical roles to which they were assigned in the sentences. In particular,iambic words were more likely to be used as verbs rather than nouns.Thus the phonological structure of a word draws it toward a particulargrammatical class.

When applied to grammatical category extensions, this conclusionleads to the prediction that a word should be more likely to develop ause in a new grammatical class if it has phonological properties typicalof that class. In terms of the noun-verb stress difference, one would pre-dict that iambic nouns should be more likely than trochaic nouns to de-velop verb uses. In contrast, trochaic verbs should be more likely thaniambic verbs to develop noun uses. Both predictions were confirmed inan historical analysis of English denominal verb and deverbal noun for-mation (Kelly 1988). Furthermore, the diachronic survey was translatedinto an experiment with current English speakers. Subjects were pre-sented with pairs of disyllabic nouns that lacked verb uses in Englishand disyllabic verbs that lacked noun uses. One member of each pairhad trochaic stress and one had iambic stress, with some other factorscontrolled. For example, the noun pairs were drawn from the same cat-egory (e.g., universities) and did not differ in prototypicality or wordfrequency. Subjects were asked to select one member of each noun pairand use it as a verb in a sentence and one member of each verb pair anduse it as a noun in a sentence. Knowledge of the noun-verb stress differ-ence affected their choices, as iambic nouns and trochaic verbs were se-

318 Michael Kelly

lected for grammatical transfers more often than trochaic nouns andiambic verbs. For instance, subjects were more likely to say “I cornelledfor my degree” rather than “I dartmouthed for my degree,” and “I did agrovel for a grade” rather than “I did a beseech for a grade.” Based onsuch findings, I would predict that “gleitman” should not sound partic-ularly melodious as a verb, however apt in meaning.

Word Blends

In 1911, a cartoonist for the Minneapolis Tribune created a new word“donkephant” by combining parts of “donkey” and “elephant.” How-ever amusing the thought might be, this wordsmith was not referring tothe offspring of a probably uncomfortable liaison. No, the reference wasto that dreaded and all too real chimera: A politician whose views don’tseem to distinguish between the Democratic and Republican Parties(Pound 1914; aka “republicrat”). English contains hundreds of blendwords like “donkephant,” such as “smog” (“smoke” + “fog”), “Jacob-ethan” (“Jacobean” + “Elizabethan”) and, newly minted for this occa-sion, “Gleitschrift” (“Gleitman” + “festschrift”).3 However, linguistshave had little to say about factors that might influence blend structure.For example, one could just as well say “eledonk” instead of “donke-phant” or “foke” instead of “smog.” Idiosyncratic aspects of blendscould certainly be relevant to their structure. Thus Lewis Carroll mayhave chosen “mimsy” rather than “flimserable” because this blend of“miserable” and “flimsy” created a more euphonic rhythm for the line“All mimsy were the borogoves.” However, one could still ask whetherany general principles could explain why existing forms won out overother alternatives. Bauer (1983), for example, recognized that someblends are probably blocked because they would be homophonouswith existing words. Thus “damn” and “hang” combined to form“dang” rather than “hamn” because the latter could be confused with“ham.” Other than this general bias against making confusions with ex-isting words, however, Bauer (p. 235) states that blend formations are“random” and “fairly arbitrary.”

In this section, I will present evidence that certain patterns in blendscan be predicted if we think of them as contractions of conjunctivephrases. Thus “fratority” and “jazzercise” are contracted forms of “fra-ternity and sorority” and “jazz and exercise.” On first inspection, thestructure of conjunctive phrases seems as arbitrary as that of blends. Inparticular, from the standpoint of grammar, word order in conjunctscan vary freely. Thus both “Henry and Lila” and “Lila and Henry” areequally grammatical (Gleitman 1965). However, analyses of large cor-pora of conjuncts have revealed that certain word order patterns are

Putting some Oberon into Cognitive Science 319

more common than others (Cooper and Ross 1975; Kelly 1986). In par-ticular, words with certain phonological and semantic characteristicstend to appear first in conjuncts. For example, the first elements of con-juncts tend to contain fewer syllables and denote more prototypical ob-jects than the second elements of conjuncts. Thus phrases like “salt andpepper” and “apple and lemon” are more common than phrases like“pepper and salt” and “lemon and apple.” Bock (1982) has induced thefollowing generalization from these patterns: The first elements in con-juncts tend to be more accessible in memory than the second elements.This difference reflects a speech production strategy to produce wordsin the order in which they are retrieved from memory, within the con-straints imposed by grammar. Since grammar imposes few constraintson word order in conjuncts, it is fairly easy to see the effects of memoryaccessibility here. However, accessibility can also affect more complexstructures, like the choice of active over passive voice and prepositionalover double object datives (Bock and Warren 1985).

This analysis could be extended to the order of elements in blends.Thus “smog” may have had an advantage over “foke” because “smoke”is a more frequent word than “fog,” and frequency is directly related toaccessibility. Similarly, “donkephant” may have won out over “ele-donk” because “donkey” contains fewer syllables than “elephant.” Inorder to examine the relation between these accessibility variables andblend structure, I supplemented Pound’s (1914) collection of blendswith a set obtained by searching the electronic version of the OxfordEnglish Dictionary. This search was conducted by retrieving all wordsthat contained “blend” or “portmanteau” in their definitions. Note thatthe resulting list was not exhaustive because many blends did not havethese search words in their entries, but there was no other systematicway to sift these remaining blends out from other words. Blends wereexcluded from the corpus if they involved more than two words (e.g.,“compushity” is composed of “compulsion,” “push,” and “necessity”)or if they could not be sensibly expanded into conjunctive phrases. Forinstance, “Westralia” is based on the adjective-noun phrase “WestAustralia,” and the early appearance of the adjective in the blend wasmore likely driven by grammatical constraints than frequency or sylla-ble number.

The words that composed each of the remaining 320 blends werescored for their syllable numbers, word frequencies (Francis andKucera 1982), and whether they appeared first or second in their respec-tive blends. Based on the analogy with word order in phrases, I predictthat shorter and more frequent words should be cannibalized for thefirst part of the blends. Both predictions were supported as the wordsrepresented early in the blends averaged 2.2 syllables and 40.1 occur-

320 Michael Kelly

rences per million words whereas the words represented later averaged2.7 syllables and 14.8 occurrences per million (syllable number: t(319) =–8.33, Word frequency: t(319) = 3.99, with raw frequencies converted tonatural log values; both ps < 0.0001 two-tailed).

One problem with this initial analysis is that syllable number andword frequency are not independent in that shorter words tend to havehigher frequencies (Zipf 1935). In order to examine word frequency sep-arately from syllable number, blends were only included if their con-stituent words contained the same number of syllables. The firstelements of blends were still more frequent than the second elements(t(116) = 2.34, p < .03 two-tailed). Syllable number could not be exam-ined by using blends whose constituents were equal in frequency be-cause there were very few blends of this type. So, syllable number wasseparated from frequency by analyzing blends if the frequency of thesecond element was greater than or equal to the frequency of the first el-ement. Even with word frequency controlled in this way, blends typi-cally placed the shorter word before the longer word (t(148) = –4.48, p <.001, two-tailed).

In sum, this analysis demonstrates that general aspects of blendstructure can indeed be predicted by psycholinguistic principles thatare broad enough to affect other aspects of language, such as wordorder. However, it will be difficult to test more detailed hypothesesusing naturally occurring blends because of likely confounds betweenvariables of interest. One could imagine, however, taking blend forma-tion into the laboratory by asking subjects to construct blends fromproperly controlled words or pseudowords, such as “Theseus” and“Oberon” or “Claire” and “Ellen.”

Rhyme Patterns in Child Verse

Throughout the world, children chant little poems while they jumprope or choose who’s “it” in games like kick-the-can and tag (seeAbrahams and Rankin 1980; Opie and Opie 1959, for review). A well-known example of the latter class of “counting-out” verse is (4):

(4) One potato, two potato, three potato, four;Five potato, six potato, seven potato, more.

One of the most interesting aspects of these poems is that they are partof an oral tradition, and hence must be recited from memory. Onewould therefore expect such poems to be structured in ways that wouldease recall. For example, there are many historical and geographicalvariants of “eeny, meeny, miney, mo,” which is the first line of the mostcommon counting-out poem among English speaking children around

Putting some Oberon into Cognitive Science 321

the world. However, all of these variants preserve the line’s regularrhythmic pattern, assonance, and alliteration. Thus versions include“eena, deena, dina, doe” but not “eeny deena miney moe” (Rubin 1995).Owing to its greater use of poetic devices, the former line has a morepredictable structure, which should aid recall. Indeed, the most com-mon form of the entire poem makes the greatest use of poetic devices(Kelly and Rubin 1988).

In this section, I will exploit our knowledge of human memory to pro-pose hypotheses about the rhyme patterns in jump rope and counting-out poems. In particular, I will assume that rhyming words in oralpoetry share some properties with paired associates in that the success-ful retrieval of the first word in a rhyme pair cues recall for the secondword. Under this description, the first word can be considered a stimu-lus for retrieval of the response word. If so, then factors that increase theeffectiveness of recall cues should cluster primarily on the first word ina rhyme pair. Ideally, such factors should also increase the intrinsicmemorability of the first word since, after all, a cue is useless if it is notavailable.

To illustrate this idea in a relatively pure form of paired associatelearning, consider an experiment by Paivio, Smythe, and Yuille (1968).Subjects first studied a set of word pairs and then, in the recall phase,had to provide the “response” member of a pair when prompted withthe “stimulus” member. The stimulus and response words could eitherbe high or low in rated imagery. Recall was best for the condition inwhich both stimulus and response words were highly imageable andworst for the condition in which both words were poorly imageable.This finding replicates many experiments that show memory advan-tages for words rated high in imagery. Of most relevance here, however,are the mixed conditions in which one word of the paired associate washigh imagery and the other low imagery. Recall scores were signifi-cantly better when the stimulus word was high imagery and the re-sponse word was low imagery than vice versa. High imagery words aretherefore better recall cues than low imagery words.

When applied to counting out and jump rope poems, these findingslead to the prediction that the first member of rhyme pairs should behigher in imagery than the second member, as in (5).

(5) As I went up the brandy hillI met my father with good will.

More generally, first rhymes should have characteristics that increasememory accessibility. This hypothesis was tested by examining twosuch variables: Imagery and syllable number. As discussed in the sec-tion on word blends, syllable number is inversely related to accessibil-

322 Michael Kelly

ity. Hence, the first word in a rhyme pair should tend to contain fewersyllables than the second word in the pair, as in (6):

(6) A bottle of pop, big bananaWe’re from southern Louisiana.

All rhymes consisting of noun pairs like “rat-cat” or “boat-petticoat”were recorded from corpora of jump rope (Abrahams 1969) and counting-out poems (Abrahams and Rankin 1980). The analysis was restricted tonoun pairs because of the definition of imagery given below. Also, sincevariables like syllable number (Cassidy and Kelly 1991) are associatedwith grammatical class, the use of mixed grammar pairs like “meadow-grow” could introduce undesirable confounds into the results for thesyllable variable. The overall survey consisted of 231 jump rope and 221counting-out rhyme pairs. The analyses combined results from both cor-pora to increase statistical power. However, the same patterns of resultsappeared in both the jump rope and counting-out samples.

Since only a small proportion of the words were listed in imagerynorms (e.g., Paivio, Yuille, and Madigan 1968), a very general, binarydefinition of imagery was used to classify each word into either a highor low imagery category. In particular, if physical object predicates like“is red” could be applied to a particular word sensibly, then that wordwas classified as high imagery. If such predicates could not be applied,then the word was considered low imagery. Note that “sensibly” doesnot mean “truthfully.” Thus the statement “Milk is red” is literally falsefor unadulterated milk, but the attribution is sensible since milk doeshave a color. Examples of words that fit the criterion for high imageryare “milk,” “fork,” “door,” and “belly.” Examples of low imagery wordsare “truth,” “duty,” “prayers,” and “noise.” The words in most rhymepairs had the same imagery value, namely high. However, when thewords differed in imagery, the first word was high imagery and the sec-ond low imagery 62% of the time (53 out of 86 cases), which was signif-icantly greater than chance (z = 2.12, p < 0.05).

The results with the syllable number variable also supported thememory accessibility hypothesis. As in the case with imagery, thewords in the rhyme pairs generally contained the same number of syl-lables, as in “bed-head” and “tomato-potato.” However, when rhymepairs contained words that differed in syllable length, the shorter wordtended to be first, as in “melon-persimmon” and “wine-turpentine.”This pattern of short word before long word occurred 58 times whereasthe reverse occurred only 21 times (z = 3.98, p < 0.01).

In sum, oral traditions of poetry and storytelling offer a rich domainfor studying memory in a naturalistic setting and for examining how

Putting some Oberon into Cognitive Science 323

memory requirements could affect the structure of such forms of cre-ative cognition (see Rubin 1995, for more details). Indeed, analyses ofsuch traditions have been well represented in volumes that examinememory at work outside the laboratory (e.g., Neisser 1980). However,these analyses have focused almost exclusively on adult traditions,such as oral poetry in the Balkans (Lord 1960) or oral history in Liberia(D’Azevedo 1962). Child verse, such as counting-out poetry, has beenrelatively ignored even though these poetic forms are apparently uni-versal, part of oral traditions, and, most importantly for research pur-poses, well documented by anthropologists. Large corpora of thesepoems are available for analysis, and as this section and other research(Rubin 1995; Kelly and Rubin 1988) show, specific hypotheses abouttheir structure can be motivated by psychological principles and tested.

Conclusion

My concluding remark is simply to thank Henry and Lila Gleitman forthe wealth of helpful contributions they have made to my research and,more importantly, to that of my students over the years. They exemplifythe honored goals of life in the Academy: to learn and to teach with de-voted reason and passion. The Greeks have a word for their tempera-ment: arete.

Notes

1. I conducted my own counts because Darbishire did not provide detailed results of herinvestigation.

2. For example, word final /i/ is sometimes spelled “y” in male names but “ie” in femalenames. Thus English has contrasts like “Billy” and “Billie.”

3. It is not entirely clear that “gleitschrift” involves blending of whole words or mor-pheme compounding at the sublexical level. In particular, the frequent use of wordslike “Gleitpeople” and “Gleitfest” in certain circles may have led to the extraction of anew morpheme “gleit” just as “scape” was extracted from the original Dutch, borrow-ing “landscape” to form “cityscape” and “seascape” (Algeo 1977). So, is the word“Gleitscape,” meaning the intellectual world from the Gleitman perspective, a blend of“Gleitman” and “landscape” or a concatenation of the morphemes “Gleit” and“scape?” Unfortunately, the issue cannot be decided in a short gleitnote.

References

Abrahams, R. D. (1969) Jump-Rope Rhymes. Austin: University of Texas Press.Abrahams, R. D. and Rankin, L. (1980) Counting-Out Rhymes: A Dictionary. Austin:

University of Texas Press.Adams, R. M. (1954) The text of Paradise Lost: Emphatic and unemphatic spellings. Modern

Philology 52:84–91.Algeo, J. (1977) Blends, a structural and systemic view. American Speech 52:47–64.

324 Michael Kelly

Battig, W. F. and Montague, W. E. (1969) Category norms for verbal items in 56 categories:A replication of the Connecticut category norms. Journal of Experimental Psychology80(3):1–46.

Bauer, L. (1983) English Word-Formation. Cambridge: Cambridge University Press.Bock, J. K. (1982) Toward a cognitive psychology of syntax: Information processing con-

tributions to sentence formulation. Psychological Review 89:1–47.Bock, J. K. and Warren, R. K. (1985) Conceptual accessibility and syntactic structure in

sentence formulation. Cognition 21:47–67.Brengelman, F. H. (1980) Orthoepists, printers, and the rationalization of English spelling.

Journal of English and German Philology 79:332–354.Carney, E. (1994) A Survey of English Spelling. London: Routledge.Cassidy, K. W. and Kelly, M. H. (1991) Phonological information for grammatical cate-

gory assignments. Journal of Memory and Language 30:348–369.Clark, E. V. and Clark, H. H. (1979) When nouns surface as verbs. Language 55:767–811.Cooper, W. E. and Ross, J. R. (1975) World order. In Papers from the parasession on function-

alism, ed. R. E. Grossman, L. J. San, and T. J. Vance. Chicago: Chicago LinguisticSociety, 63–111.

Darbishire, H. (1952) Milton’s Poetical Works. Oxford: Oxford University Press.Davis, S. M. and Kelly, M. H. (1997) Knowledge of the English noun-verb stress difference

by native and nonnative speakers. Journal of Memory and Language 36:445–460.D’Azevedo, W. L. (1962) Uses of the past in Gola discourse. Journal of African History

3:11–34.Francis, W. N. and Kucera, H. (1982) Frequency Analysis of English Usage: Lexicon and

Grammar. Boston: Houghton-Mifflin.Fowler, C. A. and Housum, J. (1987) Talkers signaling of “new” and “old” words in

speech and listeners’ perception and use of the distinction. Journal of Memory andLanguage 26:489–504.

Gleitman, L.R. (1965) Coordinating conjunctions in English. Language 41:260–293.Gleitman, L. R. and Gleitman, H. (1970) Phrase and Paraphrase. New York: Norton.Gleitman, L. R., Gleitman, H., Miller, C., and Ostrin, R. (1996) Similar, and similar con-

cepts. Cognition 58:321–376.Gleitman, L. R., Gleitman, H., Landau, B., and Wanner, E. (1988) Where learning begins:

Initial representations for language learning. In Linguistics: The Cambridge survey.Vol. 3: Language: Psychological and biological aspects, ed. F. Newmeyer. Cambridge:Cambridge University Press.

Gleitman, L. R. and Rozin, P. (1977) The structure and acquisition of reading I: Relationsbetween orthographies and the structure of language. In Toward a Psychology ofReading: The Proceedings of the CUNY Conferences, ed. A. S. Reber and D. L.Scarborough. Hillsdale, NJ: Erlbaum.

Kelly, M. H. (1986) On the selection of linguistic options. Unpublished doctoral disserta-tion, Cornell University.

Kelly, M. H. (1988) Phonological biases in grammatical category shifts. Journal of Memoryand Language 27:343–358.

Kelly, M. H. (1989) Review of Phonetics and Phonology: Volume 1: Rhythm and Meter.Language and Speech 32:171–178.

Kelly, M. H. (1992) Using sound to solve syntactic problems: The role of phonology ingrammatical category assignments. Psychological Review 99:349–364.

Kelly, M. H. (1998) Rule and idiosyncratically derived denominal verbs: Effects on lan-guage production and comprehension. Memory and Cognition 26:369–381.

Kelly, M. H., Morris, J., and Verrekia, L. (1998) Orthographic cues to lexical stress: Effectson naming and lexical decision. Memory and Cognition 26:822–832.

Putting some Oberon into Cognitive Science 325

Kelly, M. H. and Rubin, D. C. (1988) Natural rhythmic patterns in English verse: Evidencefrom child counting-out rhymes. Journal of Memory and Language 27:718–840.

Kiparsky, P. (1975) Stress, syntax, and meter. Language 51:576–616.Kiparsky, P. (1977) The rhythmic structure of English verse. Linguistic Inquiry 8:189–247.Langacker, R. W. (1987) Nouns and verbs. Language 63:53–94.Levelt, W. J. M. (1989) Speaking: From Intention to Articulation. Cambridge, MA: MIT Press.Lord, A. B. (1960) The Singer of Tales. Cambridge, MA: Harvard University Press.Meier, S. F. and Gleitman, H. (1967) Proactive interference in rats. Psychonomic Science

7:25–26.Neisser, U. (1982) Memory Observed: Remembering in Natural Contexts. San Francisco: W. H.

Freeman.Newton, R. P. (1975) Trochaic and iambic. Language and Style 8:127–156.Opie, I. and Opie, P. (1959) The Lore and Language of Schoolchildren. London: Oxford

University Press.Paivio, A., Smythe, P. C., and Yuille, J. C. (1968) Imagery versus meaningfulness of norms

in paired-associate learning. Canadian Journal of Psychology 22:427–441.Paivio, A., Yuille, J. C., and Madigan, S. A. (1968) Concreteness, imagery, and meaning-

fulness values for 925 nouns. Journal of Experimental Psychology MonographSupplement 76, part 2. 1–25.

Pinker, S. (1989) Learnability and Cognition. Cambridge, MA: MIT Press.Pound, L. (1914) Blends: Their Relation to English Word Formation. Heidelberg: Carl

Winter’s Universitätsbuchhandlung.Rubin, D. C. (1995) Memory in Oral Traditions: The Cognitive Psychology of Epic, Ballads, and

Counting-Out Rhymes. New York: Oxford University Press.Selkirk, E. O. (1984) Phonology and Syntax. Cambridge, MA: MIT Press.Sherman, D. (1975) Noun-verb stress alternation: An example of lexical diffusion of

sound change. Linguistics 159:43–81.Verrekia, L. (1996) Orthography and English stress. Unpublished doctoral dissertation,

University of Pennsylvania.Verrekia, L. and Kelly, M. H. (1996) Orthographic information for lexical stress in English.

Unpublished manuscript.Zipf, G. K. (1935) The Psycho-Biology of Language: An Introduction to Dynamic Philology.

Boston: Houghton-Mifflin.

326 Michael Kelly

Chapter 19

The Organization and Use of the Lexicon forLanguage Comprehension

John C. Trueswell

Our intuitions tell us that language comprehension is an incrementaland integrative process. As we read or listen to a sentence, we have thestrong sense that we are constantly updating our estimation of the in-tended meaning of the utterance, perhaps on a word-by-word basis. Inaddition, we make these rapid decisions by integrating a wide range ofknowledge, including grammatical knowledge of the language, “refer-ential” knowledge about what the expressions refer to in the world, andeven pragmatic and semantic knowledge about what is plausible orlikely given the situation.

One of the best illustrations of the incremental nature of languagecomprehension comes from the so-called garden-path effect, which cansometimes occur when a reader or listener is faced with a temporarilyambiguous phrase. For instance, temporary syntactic ambiguities canbe found in the following sentence fragments, which are highlighted byexamples of possible continuations.

(1) Henry forgot Lila . . .1

(a) . . . at her office. (direct object interpretation)(b) . . . was almost always right. (sentence complement interpre-tation)

(2) The man awarded the prize . . .(a) . . . to his friend and colleague of many years. (main clauseinterpretation)(b) . . . was deeply moved by the honor. (reduced relative clauseinterpretation)

In the first example, the noun phrase “Lila” could be the direct object ofthe verb, as in (1a), or the subject of an embedded sentence, as in (1b). Inthe second example, the entire fragment could make up a main clause,as in (2a), in which case the man is doing the awarding. Or, the phrase“awarded the prize” could be modifying “The man” as a reduced rela-tive clause, in which case the man is being awarded (2b). When facedwith syntactic ambiguities like these, readers and listeners show clear

signs of incremental interpretation in that they tend to pick a single in-terpretation at the point of ambiguity. Evidence for this comes from thefact that readers and listeners show systematic preferences, which needto be revised when incorrect (see, e.g., Bever 1970; Frazier and Fodor1978). This revision (or garden-path) effect is revealed by increases inprocessing difficulty, such as long fixation times and regressive eyemovements in reading (Frazier and Rayner 1982). For instance, readersprefer the direct object interpretation in examples like (1), resulting indifficulty with (1b). And, readers prefer the main clause interpretationin examples like (2), resulting in difficulty with (2b).

Although garden-path effects illustrate the incremental nature of in-terpretation, there has been considerable debate over whether readers’and listeners’ initial decisions about ambiguous phrases are the resultof integrative processes. For instance, one could argue that these deci-sions need to happen so quickly that only a subset of the most highlyrelevant information is initially consulted. Knowledge about the detailsof how particular words combine together (e.g., verb argument struc-ture), as well as semantic and pragmatic knowledge, may either be tooslow to access or too difficult to deal with during the rapid flow of in-coming speech or text. Advocates of this approach have proposed thatonly basic syntactic knowledge (e.g., major category information andphrase structure rules) is used to structure the input, and that a decisionmetric of some type is used to select among ambiguous structures, forexample, pick the simplest structure (see, e.g., Frazier 1989), or pick themost common structure (see, e.g., Mitchell, Cuetos, Corley, and Brys-baert 1995). Support for an encapsulated syntactic processor of this typehas come from studies suggesting the existence of garden-path struc-tures (e.g., a more complex or a less common syntactic alternative),which, when presented, always cause a garden path, regardless of thepresence of biasing lexical or contextual information (see, e.g., Ferreiraand Clifton 1986; Rayner, Carlson, and Frazier 1983). These studies havebeen appealing to those who support modular approaches to languageand cognition, especially given the existence of neurological data indi-cating a dissociation between syntactic and semantic processing (see,e.g., Levy 1996; Schwartz, Marin, and Saffran 1979; Hodges, Patterson,and Tyler 1994; but cf. Bates, Harris, Marchman, Wulfeck, and Krit-chevsky 1995).

Alternatives to Encapsulated Parsing

A number of recent experimental findings have, however, drawn intoquestion the basic assumptions behind an encapsulated structural stageof processing (e.g., Juliano and Tanenhaus 1994; Pearlmutter and

328 John C. Trueswell

MacDonald 1995; Taraban and McClelland 1988; Trueswell, Tanenhaus,and Garnsey 1994; Trueswell, Tanenhaus, and Kello 1993). Much of thiswork has focused on the use of lexical information, demonstrating thatdetailed syntactic and semantic information about individual wordscan have a rapid impact on parsing decisions. While space precludes afull description of these findings, it is important for this chapter to con-sider briefly two prior studies that I have conducted on this issue—oneon lexically specific syntactic information, and the other on lexicallyspecific semantic information. First, Trueswell, Tanenhaus and Kello(1993) looked at lexically specific syntactic constraints by examininghow people dealt with the direct object / sentence complement ambi-guity, as in example (1) above. We had people read ambiguous sen-tences that resolved toward the sentence complement alternative (e.g.,“Henry forgot Lila was almost always right”). In this research, we com-pared two groups of verbs: DO-bias and SC-bias verbs, which differ intheir tendency to be used with a direct object or sentence complement.DO-bias verbs permit a sentence complement, but have a strong ten-dency to be used with a direct object (e.g., “forgot”). SC-bias verbs tendto be used with a sentence complement and rarely use a direct object(e.g., “realized”). These tendencies were determined by syntacticallyanalyzing how a separate group of participants used these verbs in asentence production study. In the reading experiments, sentences withDO-bias verbs (e.g., “. . . forgot Lila was almost always right”) showedthe typical garden-path effect (i.e., long fixations and regressive eyemovements in the “disambiguating” region, “was almost always . . .”),suggesting that readers had incorrectly taken the noun as the direct ob-ject and were revising their commitment. Sentences with SC-bias verbs(e.g., “. . . realized Lila was almost always right”) showed no signs ofdifficulty in this region, suggesting that the noun was initially taken asthe subject of a sentence complement. Thus specific syntactic knowl-edge about verbs was used quite rapidly to inform the decision aboutan ambiguous phrase.

Likewise, Trueswell, Tanenhaus, and Garnsey (1994) found rapid useof lexically specific semantic information. This research examined thereading of ambiguous reduced relative clauses, like the second exampleabove. It was found that the usual garden path associated with reducedrelative clauses (e.g., “The defendant examined by the lawyer was un-reliable”) could be eliminated when the initial noun was a poor subjectand good object of the verb (e.g., “The evidence examined by the lawyerwas unreliable”). What little difficulty that was observed with theseitems correlated with ratings of how plausible the noun was as the ob-ject (theme role) of the verb. Thus semantic information about what

The Organization and Use of the Lexicon 329

makes a good subject or object of a verb can also be used to inform theearly stages of syntactic ambiguity resolution.

These and other findings have helped to develop a “lexicalist” theoryof sentence processing that emphasizes the integrative nature of inter-pretation (the constraint-based lexicalist theory; MacDonald, Pearl-mutter, and Seidenberg 1994; Trueswell and Tanenhaus 1994). Theframework assumes a constraint-based approach to ambiguity resolu-tion (Marslen-Wilson and Tyler 1987; McClelland 1987), in which multi-ple sources of information can be used to converge on a singleinterpretation. The central claim of this approach is that word recogni-tion includes the activation of rich lexical structures, including the par-allel activation of lexically specific syntactic and semantic information(e.g., verb argument structure). Syntactic ambiguities hinge upon oneor more of these lexical ambiguities, which define the initial set of possi-ble interpretations. Frequency of usage determines the initial availabil-ity of information. Thus the grammatical information computed duringword recognition determines the initial set of possible alternatives thatcontextual cues can support.

To make this more concrete, consider the account for the DO/S ambi-guity. When readers or listeners encounter a verb like “forgot,” the di-rect object (NP complement) and sentence complement structureswould become active based on frequency. Just like an ambiguous wordwith multiple meanings can have dominant and subordinate senses, anambiguous word can also have dominant and subordinate syntactic ar-gument structures. If we estimate structural frequencies from the sen-tence production data of Trueswell et al. (1993), we can assume that thedominant structure for “forgot” is the NP complement, and the domi-nant structure for “realized” is the sentence complement. This asymme-try in availability of argument structure is the proposed source of theprocessing preferences observed in the reading study, in which readersprefer the DO interpretation for “forgot” and the SC interpretation for“realized.”

The process of recognizing a verb also includes the activation of se-mantic information about the event denoted by the verb, including itsthematic/conceptual roles. What is meant by this is that the semanticrepresentation of an event includes knowledge about the possible par-ticipants of the event, as well as a mapping to the syntactic constituentsof the verb (see, e.g., Carlson and Tanenhaus 1988). This type of struc-ture permits an explanation of various semantic effects on parsing, likethose found for the reduced relative clause (“The defendant/evidenceexamined . . .”). A verb like “examined” has two roles associated withit, the agent, who is doing the examining, and the theme, which is beingexamined. In active argument structures (like the main clause), the

330 John C. Trueswell

agent maps onto the NP preceding the verb, and the theme maps ontothe NP following the verb. In passive structures (like the relative clause)the opposite pattern holds. If this information is available when recog-nizing a verb, it could serve as a mechanism for explaining the initialpreference for the reduced relative over the main clause when the firstnoun is a good theme and poor agent (“The evidence examined . . .”).Thus the thematic information of a verb can play a central role in inte-grating conceptual and syntactic constraints on interpretation.

Although the lexicalist theory is consistent with the findings de-scribed above, many of its central predictions have so far gone untested.For instance, there is little work that has demonstrated in a direct man-ner that the initial stages of recognizing a word include the activation ofargument structure. Until quite recently, most studies examining thepresence of verb argument structure during word recognition have re-lied upon secondary measures of processing load (e.g., Shapiro, Zurif,and Grimshaw 1987, 1989), and have found conflicting results(Schmauder 1991; Schmauder, Kennison, and Clifton 1991; Shapiro et al. 1987, 1989). In addition, these results have been inconclusive aboutwhether the activation of argument structure, if it occurs during wordrecognition, is frequency based, showing signs of subordinate anddominant structures. Finally, others have suggested that rapid lexicaleffects on syntactic ambiguity, like those described above, may in fact beconsistent with a structurally based system that permits extremelyrapid revision of an initial, lexically blind stage of processing (Frazier1995; Mitchell et al. 1995).

In the remainder of this chapter, I will present experimental evidencethat addresses these issues. Two different groups of results will be pre-sented, both of which explore the relationship between lexical and syn-tactic ambiguity. In the first section, I’ll describe experiments that revealhow effects of lexically specific argument preferences proliferate in syn-tactic ambiguity resolution and interact with semantic constraints. Inthe second section, I will turn my attention to effects of word recogni-tion on syntactic ambiguity resolution. I will present results that use anew lexical priming technique to examine whether the argument pref-erences of briefly displayed prime words (displayed for less than 40msec) can have an impact on a reader’s syntactic decisions about tem-porarily ambiguous sentences.

Lexical Frequency and Semantic Constraints

According to the lexicalist theory, the initial availability of a word’s syn-tactic alternatives depends upon how often the reader or listener hasencountered the word in each syntactic context. In addition, semantic/

The Organization and Use of the Lexicon 331

contextual information can come into play quite rapidly to help resolvepossible ambiguities. The theory also predicts that these two sets of con-straints interact in particular ways. For instance, processing difficultyshould arise when these constraints are in conflict, as when semantic in-formation supports a subordinate (less common) structure. Such an ef-fect has already been observed for words with multiple senses (the“subordinate bias” effect; Rayner and Frazier 1989; Rayner, Pacht, andDuffy 1994; Sereno, Pacht, and Rayner 1992). In these studies, the leftcontext of an ambiguous word supported the intended meaning of theword (as determined by the upcoming right context). Local increases inreading time occurred only when the context supported a subordinatemeaning of a word. No increases were found when the context sup-ported the dominant meaning of a word, or when the context supportedone meaning of a “balanced” word that has two equally frequent mean-ings (Rayner and Frazier 1989; Rayner et al. 1994; Sereno, Pacht, andRayner 1992).

Similar effects of context interacting with lexical preference are ex-pected for syntactic ambiguities. Consider again the semantic effects forthe ambiguous reduced relative clause (“The defendant/evidence ex-amined by the lawyer. . . ,” Trueswell et al. 1994), in which processingdifficulty was eliminated when the noun was a poor agent (“evi-dence”). One might conclude from this finding alone that the presenceof strongly biasing semantic information is sufficient for establishing aninitial preference for the relative clause. However, the lexicalist accountwould expect that the effectiveness of a semantic constraint dependsupon the availability of the appropriate structural alternative. It is wellknown that the reduced relative hinges upon an ambiguity involvingthe tense of the verb (“examined”). The “-ed” marker for most Englishverbs can indicate a past-tense verb in an active structure, such as themain clause, or a passive participle verb in a passive structure, such asthe relative clause. (Compare with unambiguous verbs like “showed/shown.”) Reading an ambiguous verb would provide partial activationfor both the past-tense and participle forms of the verb. These alterna-tives would also activate corresponding argument structures (in thiscase, the main clause and relative clause) that are consistent with thesyntactic context of a noun phrase followed by a verb. Thus there aretwo different types of frequency information predicted to play a role inthis ambiguity. One is the overall frequency of the relative clause andmain clause structures. This would result in an overwhelming prefer-ence for the main clause because a noun phrase followed by averb+”ed” is almost always a main clause structure (Bever 1970 cap-tured this in the NVN strategy). However, if structural informationhinges upon the lexical properties of verbs, this overwhelming struc-

332 John C. Trueswell

tural frequency asymmetry should be moderated for verbs with highparticiple frequency. As participle frequency increases, there is likely tobe an increase in the availability of the otherwise subordinate relativeclause alternative. For example, in Francis and Kucera (1982) frequencycounts reveal that “searched” is hardly ever used in a participle formwhereas “accused” is frequently used in a participle form. So one mightexpect to find that semantic support for the relative clause would bemore effective at eliminating difficulty when the relative clause con-tains a verb like “accused” than when it contains a verb like “searched.”

To test these predictions, I reexamined the reduced relative eyetrack-ing data reported in Trueswell, Tanenhaus, and Garnsey (1994; seeTrueswell 1996) for effects of participle frequency. Indeed, on average,verbs used in the study had relatively high participle frequencies, per-haps explaining why semantic support for the relative clause (e.g., “Theevidence examined . . .”) was in general so effective at eliminating pro-cessing difficulty (see also MacDonald et al. 1994). In addition, I foundevidence that some of the variation in processing difficulty betweenitems in this condition was predicted by variation in participle fre-quency. Regression analyses revealed that the initial processing diffi-culty for reduced relatives (as measured by first-pass reading times)negatively correlated with each verb’s participle frequency (r2 = 0.41, p< 0.05). In other words, contexts supporting the relative clause weremuch more effective at eliminating processing difficulty when the ambiguous verb was high in participle frequency. I have recently con-firmed these findings in a series of reading studies that directly com-pared verbs with high and low participle frequency (Trueswell 1996).These studies held semantic support for the relative clause constant,while manipulating participle frequency. As expected, reduced relativeclauses were more difficult to read when the verb was low in participlefrequency than when the verb was high in participle frequency (a “sub-ordinate bias” effect; see figure 19.1).

Although the relative clause data are consistent with the lexicalistpredictions for ambiguity resolution, one could argue that the findingsonly provide indirect evidence in support of this view. Specifically, onewould expect that the frequency of a verb’s argument structures, notnecessarily tense, determines the availability of syntactic forms. (Tenseonly indirectly estimates argument structure frequencies—see Trues-well 1996, for further discussion.) To address this issue, I examined howargument frequency affects the resolution of an ambiguity that does notdepend upon tense (Trueswell, Kim, and Shapiro 1997). These experi-ments took advantage of Penn’s syntactically analyzed corpora of En-glish Text (the Penn Treebank, Marcus, Santorini, and Marcinkiewicz1993) to estimate a verb’s probability of appearing with particular

The Organization and Use of the Lexicon 333

arguments. These probabilities were then used to predict processingpreferences in readers and listeners. The experiments examined a struc-tural ambiguity that arises when an alternating dative verb is placed ina passive frame (e.g., “The woman was sent . . .”). The verb “sent” canallow a second noun-phrase argument, as in “The woman was sentsome flowers,” in which case the woman is the recipient of the event.“Sent” can also allow a prepositional argument, as in “The woman wassent to the bank,” in which case the woman is the theme of the event.The ambiguity arises because “sent” is among a class of verbs called al-ternating datives, which have two competing syntactic structures fordenoting the theme and recipient roles. The verbs can be used in thedouble object construction (as in the active sentence “Bill sent Susan themoney,” or the passive sentence “Susan was sent the money”), in whichthere are two noun phrases as syntactic arguments of the verb. Theverbs can also be used in prepositional dative constructions (e.g., “Billsent the money to Susan,” “The money was sent to Susan”).

Given this observation, one might expect that knowing how often“sent” takes a second noun-phrase argument or a prepositional argu-ment could be very useful in determining the preferred interpretationof “The woman” when the verb is initially encountered in sentences like“The woman was sent . . . ”. In one experiment (Trueswell, Kim, andShapiro 1997), a cross-modal integration technique was used to exam-

334 John C. Trueswell

Figure 19.1.Ambiguity effect for the reduced relative (Trueswell 1996; copyright by Academic Press).

ine parsing commitments for the alternating dative. Participants heardauditory fragments that contained a noun that was a good recipient andpoor theme (“The boy was mailed . . .”) or a good theme and poor re-cipient (“The letter was mailed . . .”). Good recipients semantically sup-port the double object construction, whereas good themes support theprepositional dative. Immediately after hearing the fragment, the par-ticipants were visually presented with the word “the” or “to” to namealoud. The target word “the” is highly consistent with the double objectconstruction, whereas the word “to” is highly consistent with a preposi-tional phrase argument. Prior research using this technique has dem-onstrated that naming latencies are longer to target words that areungrammatical, or grammatically unexpected, continuations of thecontext (Cowart 1987; Tyler and Marlsen-Wilson 1977; Trueswell et al.1993; West and Stanovich 1986). Naming latencies (shown in table 19.1)were consistent with the rapid use of semantic information, mediatedby the initial availability of the argument structures. A reliable interac-tion was found between type of thematic fit (recipient, theme) and typeof target (“to,” “the”). When the noun was a good recipient of the verb,a double object construction should be expected, and indeed, naminglatencies in this condition were longer for “to” as compared to “the.”When the noun is a good theme of the verb, a double object constructionshould not be expected, and naming latencies in this condition shouldbe longer for “the” as compared with “to.”

Crucially, we expected these effects to depend upon the frequency ofthe verb argument structures. Again, keeping track of how often a verbappears in the double object construction could be quite useful in deter-mining the appropriate thematic assignment of the initial noun phrase.A corpus analysis was therefore conducted to determine the frequencywith which each verb appeared in the double object construction. Theanalysis revealed that double object frequency is in fact relatively lowfor verbs used in this study. Indeed, as seen in table 19.1, semantic sup-port for the recipient role (recipient-biasing nouns) is not completely ef-fective at reversing preferences for “to” over “the.” This is because thesemantic constraint in this condition supports the subordinate syntactic

The Organization and Use of the Lexicon 335

Table 19.1Mean Naming Latency to Target Word in Milliseconds

Type of Target

Context Auditory Fragment THE TO

Recipient-biasing “The boy was mailed . . .” 586 604Theme-biasing “The card was mailed . . .” 625 556

alternative (a subordinate bias effect). It was expected that the effective-ness of the semantic support for the double object (the recipient-biasedcontext) would vary continuously across verbs, with the most effectiveitems being associated with verbs that have relatively high double ob-ject frequency. This was confirmed in a regression analysis, whichpaired naming latencies in this condition with each verb’s double objectfrequency. As expected, a reliable negative correlation was found be-tween frequency and naming latencies (r2 = 0.22; p<0.05).

A second experiment (also in Trueswell, Kim, and Shapiro 1997)found that similar patterns hold for ambiguity resolution during read-ing. Eye movements were monitored as subjects read sentences like“The woman was mailed the letter . . . ”. The first noun was always agood recipient and poor theme. In this study, two classes of verbs weredirectly compared: verbs that are high in double object frequency andverbs that are low in double object frequency. As expected, processingdifficulty was found immediately after encountering a verb with lowdouble object frequency, despite the presence of semantic informationin support of this alternative.

These results complement recent findings examining the comprehen-sion of long-distance dependencies (e.g., “Which man/baseball did Billtoss . . .”), which find similar syntactic and semantic preference effectsfor alternating dative verbs (Boland 1997; Boland, Tanenhaus, Garnsey,and Carlson 1995). Taken together, the results suggest that both the-matic and syntactic information associated with a verb is accessed andused quite rapidly during interpretation. Indeed, it seems likely that re-trieval of this information during word recognition is needed to accountfor data indicating the early commitment to long-distance dependen-cies when the verb is first encountered (Boland et al. 1995; Boland 1997).

Thus it appears that for at least three ambiguities, the DO/S ambigu-ity, the relative clause ambiguity, and the alternating dative ambiguity,clear signs of verb argument preference emerge. The availability of thesyntactic properties of lexical items predicts processing difficulty andthe initial effectiveness of semantic constraints. As with other lexicalambiguities, semantic support for an alternative is less effective whenthis information supports a subordinate alternative.

Fast Lexical Priming of Argument Structure

This section turns to research that provides perhaps the most com-pelling evidence to date that word recognition itself includes the paral-lel activation of possible argument structures, and that it is thisinformation that determines initial availability of syntactic alternativesduring syntactic ambiguity resolution. These studies take advantage of

336 John C. Trueswell

a new lexical priming technique, fast lexical priming, first introducedby Sereno, Rayner, and colleagues (Rayner et al. 1995; Sereno andRayner 1992; Sereno 1995). The technique permits the examination oflexical priming during uninterrupted silent reading. In the eye-trackingversion of this technique, fixation patterns are recorded as participantssilently read text. When the eye lands on a critical target word, a primeword (of equal number of characters) appears in place of the target. Theprime is displayed for a brief amount of time (the first 30–40 msec of theinitial fixation), and is immediately replaced by the target word. This se-quence appears as a “flicker” to the subject, with subjects rarely beingable to identify a prime word. Analyses of fixation times have revealedreliable effects of the prime word’s orthographic, phonological, and se-mantic properties (Rayner et al. 1995; Sereno and Rayner 1992; Sereno1995). For instance, fixations on a target word are faster when the targetis preceded by a semantically related prime, as compared to a semanti-cally unrelated prime (Sereno 1995). Similar patterns have been ob-served for orthographically and phonologically related prime words(see, e.g., Rayner et al. 1995). Taken together, these data are highly con-sistent with theories of word recognition that allow for the parallel acti-vation of the orthographic, phonological, and semantic informationassociated with a letter string.

A central prediction of lexicalist approaches to parsing is that wordrecognition also includes the parallel activation of rich grammatical in-formation, in the form of possible syntactic complements for a word. Ifthis is the case, the syntactic preferences associated with a briefly pre-sented prime word ought to have a direct impact on a reader’s parsingpreferences of a syntactically ambiguous phrase. To test these predic-tions, we have examined fast lexical priming effects for the direct objectcomplement / sentence complement (DO/SC) ambiguity, as illustratedin the following example (Trueswell and Kim 1998).

(3) The man accepted (that) the fire could not be put out.obtained (DO-prime)realized (SC-prime)

Target sentences contained a main verb (e.g., “accepted”) followed by asentence complement. Unambiguous sentence complements alwaysbegan with the optional complementizer “that.” Ambiguous sentencecomplements did not contain the complementizer “that,” making thenoun phrase “the fire” a potential direct object of the verb. The mainverb (e.g., “accepted”) was always a verb that permits a sentence com-plement, but strongly prefers to appear with a direct object as its argu-ment (i.e., DO-biased verbs, as confirmed by sentence productionnorms). The noun phrase (e.g., “the fire”) was always a poor object of

The Organization and Use of the Lexicon 337

the verb. Several reading studies (not involving fast-priming) have ex-amined the reading of materials like these (e.g., Holmes, Stowe, andCupples 1989; Garnsey, Pearlmutter, Myers, and Lotocky 1997;Trueswell et al. 1993). All of these studies have found large garden-patheffects for DO-biased verbs when reading the ambiguous forms of thesematerials—consistent with the notion that readers initially pursued adirect object analysis of the noun phrase. For instance, Garnsey et al.(1997) found that when the optional complementizer “that” was absent,readers were surprised by the poor object “fire,” resulting in long read-ing times. Long reading times were also observed in the verb-phrase re-gion (“could not be . . .”), suggesting that readers had difficultyretrieving the subordinate sentence complement argument structure.

In the present study, a self-paced reading version of fast-priming wasused. Prior to reading each sentence, the participant saw groups ofequal signs (“=”) covering each character in the sentence. Each press ofa button uncovered a word and replaced the previous word with equalsigns. When participants reached the target verb, a prime word was dis-played in the verb position, for exactly three screen cycles (39 msec).The prime word was then replaced by the target word, which remainedon the screen until the next button press. This event was typically per-ceived as a flicker on the screen, with participants rarely identifying theprime word.

Two different types of prime words were compared. DO-primes (e.g.,“obtained”) were verbs that strongly prefer a direct object and do notpermit a sentence complement. SC-primes were verbs that strongly pre-fer a sentence complement and rarely use a direct object. (Primes werematched for string length, overall frequency, and letter overlap with thetarget verb.) If the initial stages of word recognition include the activa-tion of verb argument structures, one might expect that the subcatego-rization preferences of the “flicker” (the prime verb) would have adirect impact on the size of the garden path observed for these sen-tences. In particular, prime verbs that prefer direct objects (DO-primes)should induce a large garden-path effect, whereas prime verbs that pre-fer sentence complements (SC-primes) should reduce or eliminate thegarden-path effect.

Data from twenty-eight subjects were collected, and the magnitude ofthe garden-path effect is shown in figure 19.2.2 The differences betweenthe ambiguous and unambiguous sentences are plotted, with positivenumbers indicating increased reading times for ambiguous items. Ascan be seen in the figure, lexicalist parsing predictions were confirmed.The magnitude of the garden-path effect was much greater for DO-primes than SC-primes, resulting in a reliable interaction between am-biguity and prime type at the noun “fire,” and a marginal interaction at

338 John C. Trueswell

the disambiguating verb “could.” (Because differences are graphed, it isimportant to note that the effect of prime type is carried by ambiguousrather than unambiguous items, with a reliable effect of prime type oc-curring only for ambiguous items.)

Thus there were robust effects of lexical priming on syntactic ambigu-ity resolution. DO-primes showed much larger garden-path effects thanSC-primes. What makes this finding even more striking is that the ex-periment is comparing reading times to perceptually identical sentencesacross conditions. The only difference is whether a DO-prime or SC-prime was flashed on the screen. Thus it is the subcategorization prefer-ences of the “flicker” that are determining readers’ parsing preferences.

To analyze in more detail the contribution of prime and target subcat-egorization preferences, corpus analyses were also conducted on allprime and target verbs, from the parsed text files of the Penn Treebank.We estimated the probability that each verb uses either a direct object,sentence complement, or some other argument structure. As can beseen in table 19.2, the probabilities confirm the various classifications ofverbs. Space precludes a full discussion of the corpus data. However,note that evidence was found that the subcategorization preferences ofboth the prime and target verbs combined to predict the variation ingarden-path effects between items. For instance, a simple averaging of aprime’s DO probability and a target’s DO probability reliably predictedgarden-path effects at the disambiguating word “could,” with garden-pathing being largest for targets and primes that more strongly preferdirect objects (r=0.44, p<0.05).

The Organization and Use of the Lexicon 339

Figure 19.2. Ambiguity effect for the sentence complement (Trueswell and Kim 1998; copyright byAcademic Press).

Finally, note that given this experimental design, we do not know ifDO-primes increased the garden-path effect, or SC-primes decreasedthe garden-path effect, or whether both contributed to the pattern. Asecond self-paced study was completed to answer this question. Thesame design was used; however, nonword primes (random letterstrings) were also included in the design. In addition, we doubled thenumber of experimental items. The overall pattern was replicated: DO-primes showed large garden-path effects, and SC-primes showed littleor no garden-path effects. Critically, nonword primes showed garden-path effects that were in between these two classes of verbs, suggestingthat DO-primes increased the garden-pathing and SC-primes decreasedthe garden-pathing.

Summary

A clear picture is emerging about the syntactic aspects of word recogni-tion and their impact on incremental interpretation. The earliest stagesof word recognition include the parallel activation of possible argumentstructures, and it is this information that determines initial availabilityof syntactic alternatives during ambiguity resolution. Perhaps more im-portantly, the data begin to explain how linguistic information is orga-nized to provide rapid integration of different classes of information.Lexical information is arranged along partially independent stimulusdimensions (phonological, orthographic, semantic, and syntactic), whichare relevant for the various ways that we use language. Each type ofrepresentation adheres to the same general processing principles, thatis, information is made available in a probabilistic fashion and can beconstrained by correlated information from other dimensions.

The rapid effects of semantic constraints on syntactic ambiguity reso-lution are consistent with interactive processing mechanisms. How-ever, the results regarding the availability and priming of argumentstructure suggest partially independent representation of syntax andsemantics. This theoretical description highlights a distinction between“modular representation” and “modular processing” (Garnham 1985;

340 John C. Trueswell

Table 19.2Probability of the Direct Object (DO), Sentence Complement (S) Structures

Type of Verb DO-Comp S-Comp Other

DO-biased Target Verbs 0.55 0.23 0.22SC-Prime 0.12 0.41 0.47DO-Prime 0.84 0.00 0.16

Trueswell et al. 1994; Trueswell and Tanenhaus 1994). A modular encod-ing scheme can emerge when a system is faced with complex stimulithat contain partially independent regularities (i.e., information thatcan sometimes vary independently). For instance, the visual systemhas, to a first approximation, adopted this approach. Color and motioninformation can vary independently (red things can move up anddown, for instance). It is therefore not surprising that this information isencoded along partially independent stimulus dimensions (color, motion, etc.). Within language comprehension, lexical items are also as-sociated with distinct classes of information, which can vary indepen-dently and are only partially correlated. For instance, although it is clearthat structures imply certain meanings, differences in meaning arisewhen these structures appear in particular contexts and with particularlexical items. It is therefore not surprising to find that the system has or-ganized information along several dimensions. However, modular rep-resentation does not require modular processing. A system needs todevelop consistent solutions across stimulus dimensions, and one effi-cient approach is to be highly sensitive to the correlations that exist andallow them to constrain ambiguous representations.

Finally, it is interesting that these findings, which emphasize a closerelationship between the grammar and the lexicon, tie in nicely with re-cent developments in computational linguistics, and in particular workhere at Penn. As is the case in psycholinguistics, computational linguis-tics has seen an increased interest in lexicalized syntactic accounts (e.g.,Bresnan and Kaplan 1982; Joshi, Vijay-Shanker, and Weir 1991; Pollardand Sag 1994; Steedman 1996) and a reemergence of statistical ap-proaches to parsing (see Church and Mercer 1993; Marcus 1995).Lexicalized grammar formalisms include combinatory categorial gram-mars (CCGs), head-driven phrase-structure grammars (HPSGs), andlexicalized tree-adjoining grammars (LTAGs). It will be important in theupcoming years to bridge the (relatively small) gap between these lin-guistic formalisms and the current psycholinguistic theories of sentenceprocessing.

Closing Remarks

It is not surprising that this body of research has required the coordina-tion of several different disciplines: psychology, linguistics, and maybeeven a little bit of computer science. Interdisciplinary research has be-come the norm in the study of language—an everyday thing. Indeed,Lila and Henry Gleitman have been at the forefront of developing thisinterdisciplinary approach. But, over the last two decades, we haveseen essentially all the subdisciplines of psychology move in this

The Organization and Use of the Lexicon 341

direction. As Henry is fond of saying (quite dramatically of course),“These are changing times.” Psychology is becoming more interdisci-plinary, more “biological,” more “computational.” Many people areconcerned about whether the broadening of psychology and the blur-ring of its boundaries will have a helpful or detrimental effect on thefield. However, attending this tribute to both Henry and Lila, and lis-tening to their students, has taught me a valuable lesson about this. Wecan worry a little bit less about these changes, if we know that there arepeople involved in this process who care deeply about their students,their colleagues, the exchange of a good idea, and the exchange of agood joke, for that matter—people who think and care. Henry and Lilahave done the field a great service by promoting these ideals in theirstudents. The changing field called psychology is in better hands be-cause they have made a difference in so many lives. Henry once told methat he didn’t feel that he had really earned his Ph.D. until a few yearsafter he had received it. I know exactly what he means. I have had theopportunity of a lifetime by beginning my career at Penn. And, bywatching Henry and Lila at work with their students, I have learned agreat deal about what it means to be a teacher and a researcher. I thankthem both for their hospitality, advice, and encouragement.

Acknowledgments

This work was partially supported by National Science FoundationGrant SBR-96-16833; the University of Pennsylvania Research Foun-dation; and the Institute for Research in Cognitive Science at theUniversity of Pennsylvania (NSF-STC Cooperative Agreement numberSBR-89-20230). I am grateful to Michael Kelly and Albert Kim for help-ful comments on earlier drafts of this paper.

Notes

1. Example sentences tend to use the name “John”—a practice that I have grown tired of.I have therefore developed a program that randomly selects from a list of two names,with the only constraint being that the names appear in alphabetical order in the sen-tence. Similarities to actual people and situations are purely accidental.

2. Four subjects were excluded from this analysis, because postexperiment interviews re-vealed that they could identify the majority of the prime words. Interestingly, thesesubjects show inhibitory effects of the prime’s argument structure (see Trueswell andKim 1998).

References

Bates, E., Harris, C., Marchman, V., Wulfeck, B., and Kritchevsky, M. (1995) Production ofcomplex syntax in normal aging and Alzheimer’s disease. Language and CognitiveProcesses 10:487–544.

342 John C. Trueswell

Bever, T. G. (1970) The cognitive basis for linguistic structures. In Cognition and theDevelopment of Language, ed. J. R. Hayes. New York: John Wiley and Sons.

Boland, J. E. (1997) The relationship between syntactic and semantic processes in sentenceprocessing. Language and Cognitive Processes 12:423–484.

Boland, J. E., Tanenhaus, M. K., Garnsey, S. M., and Carlson, G. (1995) Verb argumentstructure in parsing and interpretation: Evidence from wh-questions. Journal ofMemory and Language 34:774–806.

Bresnan, J. and Kaplan, R. (1982) Lexical functional grammar: A formal system of gram-matical representation. In Mental Representation of Grammatical Relations, ed. J.Bresnan. Cambridge, MA: MIT Press.

Carlson, G. N. and Tanenhaus, M. K. (1988) Thematic roles and language comprehension.In Syntax and Semantics, volume 21, ed. W. Wilkens. New York: Academic Press.

Church, K. and Mercer, R.(1993) Introduction to the special issue on computational lin-guistics using large corpora. Computational Linguistics 19:1–24.

Cowart, W. (1987) Evidence for an anaphoric mechanism within syntactic processing:some reference relations defy semantic and pragmatic considerations. Memory andCognition 15:318–331.

Ferreira, F. and Clifton, C. (1986) The independence of syntactic processing. Journal ofMemory and Language 25:348–368.

Francis, W. N. and Kucera, H. (1982) Frequency Analysis of English Usage: Lexicon andGrammar. Boston: Houghton-Mifflin.

Frazier, L. (1989) Against lexical projection of syntax. In Lexical Representation and Process,ed. W. Marslen-Wilson. Cambridge, MA: MIT Press, 505–528.

Frazier, L. (1995) Constraint satisfaction as a theory of sentence processing. Journal ofPsycholinguistic Research 24:437–468.

Frazier, L. and Fodor, J. D. (1978) The sausage machine: a new two-stage parsing model.Cognition 6:291–325.

Frazier, L. and Rayner, K. (1982) Making and correcting errors during sentence compre-hension: Eye movements in the analysis of structurally ambiguous sentences.Cognitive Psychology 14:178–210.

Garnham, A. (1985) Psycholinguistics: Central Topics. New York: Methuen.Garnsey, S. M., Pearlmutter, N. J., Myers, E., and Lotocky, M. A. (1997) The contributions

of verb bias and plausibility to the comprehension of temporarily ambiguous sen-tences. Journal of Memory and Language 37:58–93.

Hodges, J. R., Patterson, K., and Tyler, L. K. (1994). Loss of semantic memory:Implications for the modularity of mind. Cognitive Neuropsychology 11:505–542.

Holmes, V. M., Stowe, L., and Cupples, L. (1989) Lexical expectations in parsing comple-ment-verb sentences. Journal of Memory and Language 28:668–689.

Joshi, A., Vijay-Shanker, K. and Weir, D. (1991) The convergence of mildly context-sensitive formalisms. In The Processing of Linguistic Structure, ed. P. Sells, S. Shieber,and T. Wasow. Cambridge, MA: MIT Press, 31–91.

Juliano, C. and Tanenhaus, M. K. (1994) A constraint-based account of the subject/objectambiguity. Journal of Psycholinguistic Research 23:459–471.

Levy, Y. (1996). Modularity of language reconsidered. Brain and Language 55:240–263.MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994) The lexical nature of

syntactic ambiguity resolution. Psychological Review 101:676–803.Marcus, M. P. (1995). New trends in natural language processing: Statistical natural lan-

guage processing. Proceedings of National Academy of Sciences 92:10052–10059.Marcus, M. P., Santorini, B., and Marcinkiewicz, M. A. (1993) Building a large annotated

corpus of English: The Penn Treebank. Computational Linguistics 19:313–330.

The Organization and Use of the Lexicon 343

Marslen-Wilson, W. D. and Tyler, L. K. (1987) Against modularity. In Modularity inKnowledge Representations and Natural Language Understanding, ed. K. Garfield.Cambridge, MA: MIT Press.

McClelland, J. L. (1987) The case for interactionism in language processing. In Attentionand Performance XII, ed. M. Coltheart. Hillsdale, NJ: Lawrence Erlbaum Associates.

Mitchell, D. C., Cuetos, F., Corley, M., and Brysbaert (1995) Exposure-based models ofhuman parsing. Journal of Psycholinguistic Research 24:469–488.

Pearlmutter N. J., and MacDonald, M. C. (1995) Individual differences and probabilisticconstraints in syntactic ambiguity resolution. Journal of Memory and Language34:521–542

Pollard, C. and Sag, I. (1994) Head-driven Phrase Structure Grammar. Chicago, IL:University of Chicago Press.

Rayner, K., Carlson, M., and Frazier, L. (1983) The interaction of syntax and semanticsduring sentence processing. Journal of Verbal Learning and Verbal Behavior22:358–374.

Rayner, K. and Frazier, L. (1989) Selectional mechanisms in reading lexically ambiguouswords. Journal of Experimental Psychology: Learning, Memory and Cognition 5:779–890.

Rayner, K., Pacht, J. M., and Duffy, S. A. (1994) Effects of prior encounter and discoursebias on the processing of lexically ambiguous words: Evidence from eye fixations.Journal of Memory and Language 33:527–544.

Rayner, K., Sereno, S., Lesch, M., and Pollatsek, A. (1995) Phonological cues are automat-ically activated during reading: Evidence from eye movement priming paradigm.Psychological Science 6(1):26–32.

Schmauder, A. R. (1991) Argument structure frames: A lexical complexity metric? Journalof Experimental Psychology: Learning, Memory and Cognition 17:49–65.

Schmauder, A. R., Kennison, S., and Clifton, C., Jr. (1991) On the conditions necessary forobserving argument structure complexity effects. Journal of ExperimentalPsychology: Learning, Memory, and Cognition 17:1188–1192.

Schwartz, M. F., Marin, O. S. M., and Saffran, E. M. (1979) Dissociations of language func-tion in dementia: A case study. Brain and Language 7:277–306.

Sereno, S. C. (1995) Resolution of lexical ambiguity: Evidence from an eye movementpriming paradigm. Journal of Experimental Psychology: Learning, Memory andCognition 21(3):582–595.

Sereno, S. C., Pacht, J. M., and Rayner, K. (1992) The effect of meaning frequency on pro-cessing lexically ambiguous words: Evidence from eye fixations. PsychologicalScience 3(5):296–299.

Sereno, S. C. and Rayner, K. (1992) Fast priming during eye fixations in reading. Journal ofExperimental Psychology: Human Perception and Performance 18:173–184.

Shapiro, L. P., Zurif, E., and Grimshaw, J. (1987) Sentence processing and the mental rep-resentation of verbs. Cognition 27:219–246.

Shapiro, L. P., Zurif, E., and Grimshaw, J. (1989) Verb representation and sentence pro-cessing: contextual impenetrability. Journal of Psycholinguistic Research 18:223–243.

Steedman, M. (1996) Surface Structure and Interpretation. Cambridge, MA: MIT Press.Taraban, R. and McClelland, J. (1988) Constituent attachment and thematic role assign-

ment in sentence processing: Influences of content-based expectations. Journal ofMemory and Language 27:1–36.

Trueswell, J. C. (1996) The Role of Lexical Frequency in Syntactic Ambiguity Resolution.Journal of Memory and Language 35:566–585.

Trueswell, J. C., Kim, A., and Shapiro, S. (1997) Using verb information during listeningand reading: Semantic fit and argument frequency effects for the alternating da-tive. In preparation.

344 John C. Trueswell

Trueswell, J. C. and Kim, A. (1998) How to prune a garden-path by nipping it in the bud:Fast priming of verb argument structure. Journal of Memory and Language39:102–123.

Trueswell, J.C., and Tanenhaus, M.K. (1994) Toward a lexicalist framework for constraint-based syntactic ambiguity resolution. In Perspectives on Sentence Processing, ed. C.Clifton, K. Rayner, and L. Frazier. Hillsdale, NJ: Erlbaum.

Trueswell, J. C., Tanenhaus, M. K., and Garnsey, S. M. (1994) Semantic influences on pars-ing: Use of thematic role information in syntactic ambiguity resolution. Journal ofMemory and Language 33:285–318.

Trueswell, J. C., Tanenhaus, M. K., and Kello, C. (1993) Verb-specific constraints in sen-tence processing: Separating effects of lexical preference from garden-paths.Journal of Experimental Psychology: Learning, Memory and Cognition 19:528–553.

Tyler, L. K. and Marslen-Wilson, W. D. (1977) The on-line effects of semantic context onsyntactic processing. Journal of Verbal Learning and Verbal Behavior 16:683–692.

West, R. F. and Stanovich, K. E. (1986) Robust effects of syntactic structure on visual wordprocessing. Memory and Cognition 14:104–112.

The Organization and Use of the Lexicon 345

Cynthia FisherUniversity of Illinois

Susan Goldin-MeadowUniversity of Chicago

Kathy Hirsh-PasekTemple University

John JonidesDepartment of PsychologyUniversity of Michigan

Philip J. KellmanUniversity of California, LosAngeles

Michael KellyDepartment of PsychologyUniversity of Pennsylvania

Donald S. LammNorton Publishers

Barbara LandauUniversity of Delaware

Jack NachmiasUniversity of Pennsylvania

Letitia R. NaiglesUniversity of Connecticut

Elissa L. NewportDepartment of Brain andCognitive SciencesUniversity of Rochester

W. Gerrod ParrottGeorgetown University

Daniel ReisbergReed College

Robert A. RescorlaUniversity of Pennsylvania

Paul RozinUniversity of Pennsylvania

John SabiniDepartment of Psychology University of Pennsylvania

Elizabeth F. ShipleyUniversity of Pennsylvania

Thomas F. ShipleyDepartment of PsychologyTemple University

John C. TrueswellDepartment of PsychologyUniversity of Pensylvania

Contributors

Abrahamsen, Adele, 209Adams, R. M., 316Adaptive staircase procedure, 178Adjectives, in verse, 313Alphanumeric stimuli, 10Alternating dative ambiguity, 334–335American Sign Language (ASL)children’s acquisition of, 113topicalized structures in, 116–117

Anderson, R. L., 235Andrade, J., 141Aphasia, 10Appearance. See also Physical objectsidentity determined by, 79, 80and object naming, 221

Argument linking, in syntax acquisition, 283Argument structurefast lexical priming of, 336–340of verbs, 333and word recognition, 331

Armstrong, Sharon, 16Articulation, concurrentand imagery, 148and memory-span performance, 149

Asch, Solomon, 4Aslin, Richard, 107Auditory imagerylimitations in, 146and parallel phenomena, 154and perception, 143, 144, 145perceptual understanding in, 151production of, 146–147and subvocalized support, 141and task-irrelevant noises, 141

Auditory images, creation of, 150Awh, E., 89

Baddeley, Alan, 89, 141Baker, E., 195

Battig, W. F., 317Bauer, L., 319Behavioral properties, in identification of

physical objects, 77–80Besag, J., 143, 155n.2Blackmore, S. J., 295Blind childrenlanguage learning in, 14, 15, 210objects named by, 214–216spatial knowledge in, 15visual verbs and, 16

Bloomfield, Leonard, 6Boundariesin 3-D world, 170interpolation, 158, 186motion signals for, 301

Boundary assignment, 165Boundary localization, and relatability, 178Bound morpheme comprehension, study

of, 197–206Bound morphemes, 193data analysis of, 205and sentence processing, 202

Bower, Gordon, 239Bowerman, M., 277Bowes, Thomas, 233Brain activationcomputer display of, 102dose response curves for, 95–97and memory load, 98in neuroimaging experiments, 92–94

Brelstaff, G., 295Brent, M. R., 283Bright, Timothy, 233Brown, R., 218Bruner, J., 128Burton, Richard, 234Bush, Robert, 5, 27, 53Bushnell, Emily, 7

Index

Caregiver speech, 11. See also MothereseCarew, Richard, 233Carey, S., 71, 72Categoriesentrenched properties and, 73–76membership in, 70psychological essentialism of, 70–73

Category developmentin children, 73establishing identities in, 291

Causal agent, in object naming, 220Certainty distinction, and MSV acquisi-

tion, 250–251, 257–258, 266Chalkley, M. A., 107, 194Changemotion-based model of, 301perception of, 293–394and perception of occlusion, 299during saccades, 295

Charron, Pierre, 233Cheese Seminar, 8, 191. See also SeminarsChiasson, L., 255Children. See also Blind children; Deaf

childrenacquisition of mental state verbs in, 245bound morpheme comprehension in,197–200, 202

categorization of objects by, 69entrenched categories acquired by, 75–76notions about insides of, 81–82preschool experience of, 263–266sensitivity to morphological cues of, 196statistical learning in, 109–110

Children, hearing, gesture systems of, 123Chinese culturedevelopment of gestural systems in, 129mother-child interaction in, 128

Chomsky, Carol, 4Chomsky, Noam, 4, 6, 275–276Chronometric analysis, 87Clark, E. V., 317Clark, H. H., 108, 317Clark, Kenneth, 1Classification. See also Categoriessimilarity as basis for, 215and word learning, 216

Cognitive developmentand MSV acquisition, 254–255similarities in, 215

Cognitive psychological approach, tomental verbs, 249–250

Cognitive science, 87

and creative language use, 312foundation of, 213

Coloration, and adjectives, 219Colwill, Ruth, 43–44Combinatory categorial grammars

(CCGs), 341“Common Fate, Factor of,” 181–182Complete objects, perception of, 182Compound nouns, structure of, 6Comprehension studies, 195, 205. See also

Cognitive scienceComputational analysis, of visualizing

objects, 163Computational linguistics, 4, 341Computer models, in study of cognition,

87Computersadvances in, 12UNIVAC2, 3

Conjunctive phrases, contractions of, 319Conjuncts, 320Continuation principle, 163, 166Continuityfirst-order, 164Gestalt notions of, 165

Continuity errors, 295–296Contours3-D illusory, 170and good continuation, 163illusory vs. occluded, 160–161occluded, 170relatability, 174

Counting-out poems, 321–322, 324Count noun context, for object naming,

219Crabbe, G., 157Crowder, R., 142Cunningham, Douglas, 297

Darbishire, H., 316Das, A., 186Davidge, J., 255Deaf childrenacquisition of language by, 14Chinese, 129gesture systems of, 123–124, 126, 133language acquisition of, 113language use of, 126and parent-child interaction, 127, 128

Deception, emotional, 236Decomposing, of complex cognitive

processes, 87

350 Index

Dell, Gary, 7Deprivation paradigmin language acquisition, 245pioneered by Gleitmans, 269

Depth relatability experiments, stimuli in,170–171

Depth relationship, between objects, 160Detachment gain, 155Diachronic survey, 318Diggle, P. J., 143, 155n.2Discontinuity, first-order, 164Discovery, image-based, 152Distributional analysis, of language acqui-

sition, 17–18Distributional information, for language,

107Donders, R. C., 91, 95–96DO/S ambiguity, 336Dot localization paradigm, 178–179Down’s syndrome, language learning in,

14Drama, H. Gleitman’s interest in, 231–232.

See also Plays; TheaterDynamic occlusiondisplays, 299, 300, 301paradigm, 183–187

Ecological analyses of perception, 181Edge, occluding, 165–166Edge classification, 165, 168Edge continuation, and interpolation, 186Edge-insensitive process, 181Edge interactions, 187Edge relatabilityand object perception, 181and similarity, 179spatiotemporal, 182–183

Edge-sensitive process, 182Education, undergraduate, 54Elyot, Sir Thomas, 233Embarrassment, emotion of, 232Emotionin Elizabethan psychology, 234–241everyday meaning of, 240–241folk psychologies of, 232modern academic theories of, 241modern psychology of, 240–241

Enacted images, 147English language, lexical innovation in,

317. See also LanguageEntrenchment, concept of, 73Equifinality, in language learning, 129–130

Ergative languages, 286, 287Errors, patterns of, 87Essences, psychological, 70–71Essentialism, psychological, 70–73Ethics, emotion and, 241–242Events, in Elizabethan psychology, 237

Factivity, determination of, 248–249Factivity dimensionchildren’s understanding of, 253in grade-schoolers, 252and MSV acquisition, 254in three-year-olds, 251

Factor of direction, 162False beliefs, first-order vs. second-order,

269n.3Familiar morpheme hypothesis, 204Fast lexical priming, 337, 338Feldman, Heidi, 7, 13, 14, 123, 210Field, D., 168Fisher, Cynthia, 220, 275–290Form class, 193grammatical morphemes in, 194and spatial relationships, 213

Fowler, C. A., 313Frank, R., 255Free report procedure, 300Function, 221in object naming, 221–225smooth, 162

Furrow, D., 253, 255

Gabor patches, 168Galanter, Eugene, 3–4, 5Gallistel, Randy, 15Garden-path effects, 338in language learning, 327–328magnitude of, 339

Garnsey, S. M., 329Geer, Sandra, 7Gelman, R., 81–82Gelman, S. A., 82, 195Generalization, in object naming, 226Gerken, L. A., 196, 197, 201Gestalt principlesgood continuation, 188good form, 174–177law of proximity, 169in object perception, 158object segregation, 157and unit formation, 161

Gestalt psychology

Index 351

good continuation principle in, 161–165,166

identity hypothesis and, 159–161neural models of, 186–187update of, 158

Gesture creation, and language develop-ment, 129

Gesture-speech mismatch, and readiness-to-learn, 131

Gesture-speech system, integrated,130–131

Gesture systemsof deaf children, 123effect on learning of, 132–134environmental conditions for, 126–129grammatical categories for, 125and mother-child interaction, 128sentence-level structure in, 124–125word-level structure of, 125

Gibson, J. J., 293, 294, 296Gilbert, C. D., 186Gleitman, Claire, 6, 28Gleitman, Ellen, 6Gleitman, Henryacademic style of, 29–30educational philosophy of, 311–312formal education of, 1honors and awards of, 18–19interdisciplinary approach of, 341–342interest in drama of, 28, 35, 231as mentor, 211A Midsummer Night’s Dream directed by,311

photos of, 35, 36, 37published with students, 13, 14, 24, 41at Swarthmore, 4–5, 23on teaching, 49teaching of, 191, 210

Gleitman, Lila R., 36, 38cited, 192, 194, 195, 197, 202, 206, 209, 210,218, 220, 247

collaboration with P. Rozin, 28coteaching with H. Gleitman, 54–55educational philosophy of, 312honors and awards of, 19interdisciplinary approach of, 341–342as mentor, 211Ph.D. work of, 6–7Psychology reviewed by, 64published work of, 11, 12, 13teaching of, 191at Univ. of Penn, 13

Gleitman, Phillip, 34Gleitman family, 33Gleitology, 57–65Global instruction condition, 178Global symmetry, 188Goddard, David, 5Goffman, Erving, 53, 232Goldberg, A., 276Goldfish, memory studies in, 9, 27Goldin-Meadow, Susan, 7, 13, 14, 121–137,

210Goldsmith, John, 7Golinkoff, Roberta, 192, 197, 284Good continuationand object perception, 188principle, 161–163, 163–165, 166and relatability, 166–167updated notion of, 174

Good formprinciple of, 174–177putative examples of, 175

Goodman, Nelson, 69, 73, 82Grammaracquisition of, 288discovery of, 192in gesture systems, 125–126learning, 204–205lexicalized formalisms in, 341and lexicon, 341universal architectural principles of,116–117

universal (UG), 276Grammatical categoriesdistributional view of, 194extensions of, 317–318

“Great Verb Game,” 17Grief, in Shakespeare’s plays, 237–238Groupingbasic problem in, 158–159and notions of smoothness, 163understanding, 187

Hall, G., 220Hall, W. S., 254, 255, 264Harris, Zellig, 3, 17, 112Hayes, A., 168Hayes, J. R., 108Head-driven phrase-structure grammars

(HPSGs), 341Hearing children, gestures of, 133Hess, R. F., 168Hilgard, E. R., 39

352 Index

Hirsh-Pasek, Katherine, 191–208, 284Hiz, Henry, 6Hoenigswald, Henry, 3, 6Hoff-Ginsberg, E., 268Holophrastic listeners, and grammatical

morphemes, 195Homophone judgment, 153Housum, J., 313Hull, 24Humor, role of expectation and surprise

in, 14Hurvich, Leo, 5Huttenlocher, J., 264

Iambic pentameter, 312–313, 315Identification. See also Object namingeffects of insides upon, 80–82entrenched properties and, 76–80

Identification judgments, of children,82–83

Identityand perception, 292–294psychological, 291

Identity hypothesisand global-local controversy, 176–177in object completion, 159–161

Imagery. See also Auditory imageryauditory vs. visual, 152binary definition of, 323functioning of, 151and memory, 322

Imagesinterpretation of, 155n.4nature of, 152pathways for creation of, 142–148

Imagined sounds, multidimensional spacefor, 142

Inductive inferencesentrenchment concept in, 73and psychological essentialism, 72role of entrenchment in, 74

Infantsenactment tasks for, 195function morphemes identified by, 196language learning in, 6morpheme sensitivity of, 201–202statistical learning in, 109–110understanding in, 195

Information, functional, and shape bias,221–225

Information, verbal, working memory for,89

Informativeness, and poetic meter, 313Insides, and determination of identity,

80–82Instrumental learning, experiment in,

44–45Instrument timbres, study of, 142–148Intention, and object naming, 221Intermodal preferential looking paradigm

(IPLP), testing, 198–200Interpolation, in perception, 182Intransitives, subjects of, 286–287Item recognition, study of, 9–10Iverson, P., 143

Jackendoff, R., 281James, William, 36, 37Jameson, Dottie, 5Johansson, G., 292Johnson, E. C., 72Jones, S. S., 71Jonides, John, 7, 9, 13, 34, 87–103Joshi, Aravind, 4Jusczyk, Peter, 7

Kalish, C. W., 72Kanizsa’s Demonstrations, 175Kanizsa, G., 160Kaplan, G., 294Katz, N., 195Katz, S., 89Kauffman, Bruria, 4Keil, F. C., 71, 81Kellman, Philip J., 14, 157–190, 301Kelly, Michael, 311–326Kinetic screen effect, 294Koeppe, R. A., 89Koffka, K., 181Köhler, Wolfgang, 4, 23, 61, 307Krumhansl, C., 143

Lamm, Donald S., 5, 57–65Landau, Barbara, 14, 71, 209–230, 288Languageergative, 286, 287and perceptual bias, 220–221resilience of, 123–124and space, 210–211, 227

Language acquisition, 275in children with Down’s syndrome, 129comprehension data in, 195, 205conceptual structures in, 281deprivation paradigm in, 245

Index 353

Language acquisition (cont.)distributional analysis of, 17–18grammatical morphemes’ role in, 192and inconsistent linguistic input, 113–115natural experiments in, 112nature-nurture question in, 11–12output grammar, 113and partial information, 288problem of, 105–106role of syntax in, 218–221transitional probabilities in, 108–110understanding, 13

Language and Experience (Landau andGleitman), 211, 214

Language comprehensionincremental nature of, 327lexical items in, 341

Language developmentlanguage comprehension in, 195sensitivity to morphological cues in, 196

Language learning, 122equifinality in, 129–130matching problem in, 291object naming in, 225–227research in, 6resilient properties of, 124–126of second language, 14spatial experience in, 211

Learningdelayed response, 41gesture’s role in, 130, 132segmentation in, 111and teaching, 154

Leeuwenberg, E., 174Lennard, Samson, 233Levin, D. T., 296, 307Lexicalized tree-adjoining grammars

(LTAGs), 341Lexical learning, 211–212Lexical preference, and syntactic ambigui-

ties, 331–336Lexical priming technique, 337Lexical stress, and English spelling, 312,

314, 315, 317. See also StressLichtenberg, Lila, 2–3. See also Gleitman,

Lila R.Linguistic inputbuilding structure with, 116–118inconsistent, 113–115natural experiments of, 112–118reshaping and restructuring of, 117–118

Linguistics

computational, 341mental verbs in, 248–249primitives and, 276

Liquid-crystal-diode (LCD) shutterglasses, 171

Literature, folk, 240Location, and object naming, 220Locke, John, 211Lorenz, Konrad, 32Luce, Duncan, 5, 6Luminance masking, 295

MacMillan, Deborah, 7MacNamara, J., 195Malt, B. C., 72Maratsos, M., 107, 194Marin, Oscar, 10Markman, E. M., 74Material, and object naming, 220Maternal speechmental verbs in, 247–248and MSV acquisition, 255–256

May, Robert, 7McDill, M., 301McGill, Bill, 25McIntosh, B. J., 196, 197, 201Meaningsfor blind child, 212of gestures, 125of MSVs, 250–251and presyntactic cues, 279–280sentence structure and, 276–277

Memorygoldfish, 8–9investigations of, 5visual, 296–297

Memory, working, 88–99architecture of verbal, 89–99defined, 88differing subsystems of, 99–102parametric studies of, 92, 95spatial vs. verbal, 99–100, 102studies in, 148

Memory control, in neuroimaging experi-ments, 89–90

Mental context, detachment from, 154Mental image, understandings attached

to, 150Mental representationsprototypes for, 16shape bias in, 218

Mental state verb acquisition

354 Index

experiments with, 258–263and preschool experience, 263–266and radical translation problem, 247–248theory of, 254–256

Mental state verbs (MSVs)adult usage of, 264characteristics of, 246–247children’s acquisition of, 245–246degree of certainty of, 250–251, 257–258,266

developmental understanding of,251–252

first uses of, 251measuring comprehension of, 260, 261,262

polysemy of, 248–251and theories of mind (TOM), 252–254

Michotte, A., 157, 293, 296, 306Miller, George A., 58, 59, 60Milton, John, 315, 316Mind. See Theory of mindMintz, Toby, 107Montague, W. E., 317Moore, C., 253, 255, 257, 258Morgan, J., 194Morphemes, grammatical, 193. See also

Bound morphemesacquisition of, 113in form class assignments, 194in language acquisition, 192–193sensitivity to, 195in syntactic development, 194

Morphologybound, 204–205in language acquisition, 114probability used, 118

Motherese, 11–12, 13, 105, 122, 196, 246Mother’s speechMSVs in, 269n.4studies of, 105

Motion, perceiving structure from, 292Motion signalsboundaries defined by, 302–306to perceive moving boundary, 301–302sequential pattern of, 303

Murphy, Gardner, 1Music, segmentation in, 111

Nachmias, Jacob, 4, 5, 23–25, 40Nagy, W. E., 254Naigles, Letitia R., 245–274, 284Nakayama, K., 165

Naming latencies, 335Naming patterns, study of, 222. See also

Object namingNature-nurture questions, 112“N-back” task, in working memory exper-

iments, 89, 93, 96, 97Necker cube, 151“Neg-raising”defined, 269n.1syntactic phenomenon of, 248–249

Neisser, Ulrich, 4, 23, 24, 296Nelson, K., 295Neural models, 186–187Neuroimaging techniquesapplication of, 102–103experiments with, 89–99in reaction time studies, 91–92in study of cognition, 87–88

Newport, Elissa, 7, 10–11, 13, 105–119Norton, 59Nounsin child-directed speech, 192–193compared with verbs, 317–318contexts of, 219first-syllable stress, 318presyntactic primitives as, 283–285prosodic properties of, 193

Object completiondepth information in, 170–174dynamic, 183–187and edge interactions, 187global notions of, 179hypothesis, 172identity hypothesis in, 159–161similarity in, 179

Object constancy, occlusion and, 297–301Object function, studies of, 226. See also

FunctionObject names, for blind children, 214–216Object namingchildren’s vs. adult’s, 225–227form vs. function in, 223–224generalization in, 226mature, 225–227role of syntax in, 218–221and shape bias, 217–218shape in, 216–217

Object perception. See also Perceptionmodels of, 188multiple tasks in, 165relatability in, 166–168

Index 355

Object permanence, in occlusion displays,307

Object recognition system, object namingin, 227

Objectscategorization of, 69in Elizabethan psychology, 237(nonnaming) judgments about, 222persistence of, 293in word acquisition, 213

Object shape, 216–217. See also Shape;Shape bias

Object unity, in 3–D world, 170Occlusiondynamic, 299and object constancy, 297sequential, 303, 304, 305

Opacity, of moving surfaces, 305Optics, ecological, 163Oral traditions, and study of memory,

323–324O’Regan, J. K., 295Orthography, English, variability in, 314,

317Outcomes, in instrumental learning,

44–45, 47

Paivio, A., 322Palmer, Evan, 183Paquin, M., 255Parental input, and mental verb under-

standing, 255, 263. See also Maternalspeech

Parent-child interaction, and MSV distinc-tion, 267–268

Parentheticals, MSVs in discourse struc-ture of, 249

Parrott, W. Gerrod, 14, 231–244Parser, development of, 4Parsingalternatives to encapsulated, 328–331lexicalist approaches to, 337

Partial sentence representation (PSR)explained, 281and utterance structure, 285

Participle frequency, effects of, 333Particular morpheme hypothesis, of mor-

pheme sensitivity, 203–204Paulesu, E., 90Pavlovian conditioning, exploration of, 46Peabody Picture Vocabulary Test (PPVT)

Revised, 256

Peer language use, 268Pennsylvania, University of. See also

Psychology Dept.Gleitmans’ move to, 5H. Gleitman as chairman at, 53H. Gleitman’s arrival at, 60job-talk ritual at, 121–122

Perceptionand auditory imagery, 143and conscious experience, 307and continuity errors, 295–296depth relationship, 160ecological analyses of, 181in Elizabethan psychology, 237of form, 151in Gleitmans’ seminars, 157identity and, 292and internal representation, 293, 294–295and luminance masking, 295of object constancy, 297

Perception, visual, and object naming,216–217

Perceptual biases, language and, 221Perceptual constancies, 292Perceptual development, and Gestalt prin-

ciples, 158Perceptual organizationdevelopment of, 157understanding, 187

Perceptual processing, continuity in, 165Perceptual unit formation, 291Perceptual verbs, 250Persistence, perception of, 293, 294Petersik, J. T., 301PET scanner, working memory experi-

ments with, 101Petter, G., 159Petter’s effect, 160, 176Philosophy, educational, 311Phrasal stress. See StressPhrase and Paraphrase (Gleitman and

Gleitman), 7Physical objects. See also Object naming;

Object perceptionimportance of, 71psychological categories of, 69transformation of properties of, 71

Picture pointing tasks, limitations of, 198Pinker, S., 193, 194Pitt, M., 142Place, in word acquisition, 213. See also

Space

356 Index

Plays. See also Drama; Theaterdirected by H. Gleitman, 29produced by H. Gleitman, 25psychology in, 232–234Shakespeare’s, 232–240

Poetic meteriambic pentameter, 312–313, 315information in, 313

Poetry, and study of memory, 323–324. Seealso Verse

Polat, U., 186Prägnanz principle, 187Predicatesargument, 283spatial relationships as, 213

Preschoolers, vs. home-schooled children,265

Preschool experience, and mental verbunderstanding, 263–266

Priming studies, 177PrimitivesChomsky on, 275–276nouns as presyntactic, 283–285

Problem solving, and working memory,88

Prosodic bracketing, 194Prosody, in language acquisition, 193Prototype representations, 16Proximity, Gestalt law of, 169Pseudowords, disyllabic, 314Psycholinguistics, 192, 206, 341and sentence processing, 341and word blends, 321

PsychologyAmerican, 13Elizabethan, 234–240in English Renaissance, 232–234folk, 232, 240–242identity problem in, 291Renaissance, 240–242teaching of, 27

Psychology course, introductory, 39–40recording of, 60–61syllabus for, 58, 59, 60

Psychology dept., Univ. of Penn, 5–6, 61rules of, 54teaching at, 154

Psychology (Gleitman), 311beginning of, 5completion of, 19, 65editions of, 65first chapter of, 62

historical component of, 63origins of, 58, 59Renaissance equivalent of, 232reviews of, 63–64success of, 64–65

Psychophysiology, Elizabethans’, 235Puppets, in experiments with children,

258–259

Quine, W. V. O., 214, 215, 216, 246

Radical translation, problem of, 246, 247Radio Free Europe, 25Rakowitz, S., 220Rayner, K., 337Reaction times, study of, 91, 94Readiness to learn, gesture and speech as

index of, 131–132Readingambiguity resolution during, 336studies of, 9

Reading acquisition, early study of, 27–28Reasoning, and working memory, 88Recall cues, effectiveness of, 322Recognition from partial information

(RPI), 176, 177Reduced relative clauses, and participle

frequency, 333Rehearsal, memory, and concurrent artic-

ulation, 149Rehearsal control, in neuroimaging exper-

iments, 89–90, 91Reisberg, Daniel, 139–156Relatabilityin cases of minimal gaps, 169–170construction used to define, 1673-D, 170–174in edge interpolation, 186experimental evidence about, 168–169and localization of boundaries, 178in object perception, 166–168spatiotemporal, 182–183, 184

Relatable displays, accuracy of, 185Relative clause ambiguity, 333, 336Rensink, R. A., 295, 307Repetitions, imagined vs. perceived, 149Rescorla, Robert A., 39–47Reynolds, H., 294Rhyme judgment, 153Rhyme patterns, in child verse, 321–324Rice, M., 256Rips, Lance, 7

Index 357

Ritter, E., 277Rock, I., 301Rosch, Eleanor, 16Rosen, S. T., 277Ross, D. S., 115Rozin, Paul, 8, 9, 10, 27–38, 42

Sabini, John, 49–56Saccades, changes during, 295Saffran, Jenny, 107Sagar, Naomi, 4Sagi, D., 186Same kind, in object naming, 225Scherer, Martin, 1Schmidt, Hillary, 14Scholnick, E., 254Schumacher, E. H., 89Schwanenflugel, P., 250, 252Schweisguth, Melissa, 192, 197Second language learning, 14Segmentationbasic problem in, 158–159and notions of smoothness, 163understanding, 187word, 107–110

Selective interference experiments, 100Self-control, as virtue, 239, 240Self-splitting figures (SSOs), 159–160, 161Semantic constraints, lexical frequency

and, 331Semantic informationlexically specific, 329and verb recognition, 330

Semantics, and language acquisition, 193Seminarscheese, 8, 191Gleitmans’ research, 7–8, 8, 24, 40, 55, 87,103, 118, 139, 157, 188, 191, 221, 242

L. Gleitman’s graduate, 54, 56weekly evening, 122–123, 209–210

Sentence interpretation, and sentencestructure, 281–283

Sentence processing, psycholinguistictheories of, 341

Sentences, children’s gesture, 124–125Sentence structureand conceptual structure, 282–283and meaning, 276–277presyntactic, 288

Sereno, S. C., 337“Sesame Street,” 256

Shafer, V. L., 196Shakespeare, plays ofemotion in, 234–240psychology in, 232–234

Shapeaccurate identification of, 303artifacts used in studies of, 221and object naming, 216–217, 219

Shape bias, 75in adults vs. children, 223–224and functional information, 221–225and object naming, 217–218

Shatz, Marilyn, 7–8, 251Shepard, Roger, 140Shimojo, S., 165Shipley, Elizabeth, 6, 8, 11, 12, 13, 69–85,

192, 194, 195, 197, 202, 206, 226Shipley, Thomas F., 291–309Shipley, Tim, 157, 179, 183Shows. See also Theaterdirected by H. Gleitman, 29produced by H. Gleitman, 25

Shucard, D., 196Shucard, J., 196Siblings, effect on MSV acquisition, 268Siegal, Muffy, 7Sigman, E., 301Silverman, G., 165Similarityand edge relatability, 179and object naming, 214–215, 217

Simons, D. J., 81, 296, 307Singer, D., 256Singer, J., 256Singleton, J. L., 116Siskind, J., 282Sloan Group, 211Smith, Carlotta, 6, 11, 12, 13, 192, 194, 195,

197, 202, 206Smith, E. E. (Ed), 88, 89Smith, L. B., 71Smoothness, notions of, 162–163Smythe, P. C., 322Solomon, Dick, 5, 41, 42Sounds. See also Auditory imagerymental images of, 144mental representations of, 153ratings of perceptions of, 143, 145

Space, vs. language, 210–211, 227Spatial representation, and object naming,

227

358 Index

Speechcovert, 155n.1prosodic principles of, 313

Spelke, Elizabeth, 14, 15, 157Spelling, and stress, 314Stabilityappearance of, 294change as information for, 293–294illusory, 296, 297, 300, 307motion-based model of, 301and motion signals, 305–306perception of, 292, 296

Statistical information, 107Statistical learning, 106–107, 117in language acquisition, 111vs. nonstatistical learning, 111and word segmentation, 107–108

Sternberg, S., 92Storytelling, and study of memory,

323–324Stress. See also Versefirst- vs. second-syllable, 318spelling and, 314

Stress patterns, noun vs. verb, 318Students, psych 1, 51–52. See also

Psychology course, introductorySubjects, and transitive vs. intransitive

sentences, 285–287“Subordinate bias” effect, 332Subrahmanyam, K., 220Subtraction strategy, in cognition studies,

89Subvocalization, 140and auditory imagery, 152detachment provided by, 153planning mechanisms for, 155n.3

Surface completion process, 180Surface filling, 186Surface interpolation, examples of, 158Surface quality, spreading of, 179–181Surfaces, illusory vs. occluded, 160–161Surface texture, and adjectives, 219Swarthmore College, 4–5, 7, 23Syllable number, and memory, 322–323Symmetry testing, 179Syntactic ambiguitiesand lexically specific argument, 331–336and lexical preference, 331–336and lexical priming, 339

Syntactic bootstrappingdefined, 277–279

original proposal of, 288pioneered by Gleitmans, 276presyntactic mechanism for, 284–285sentence interpretation in, 283in verb learning, 280

Syntactic context, and object naming, 219Syntactic diversity, 268Syntactic evidence, child’s understanding

of, 279–280Syntactic information, lexically specific, 329Syntax acquisition, 276argument linking in, 283early education of, 246and presyntactic division, 287utterances in, 282verb learning in, 280

Tanenhaus, M. K., 329Tangent discontinuity (TD), 163continuity and, 165–166and good form, 174

Taylor, M., 195Teacher-preschooler interactions, and

MSV distinction, 267–268Teaching. See also Seminarsof graduate students, 154H. Gleitman on, 49of L. Gleitman, 54–55, 56of psychology by H. Gleitman, 27

Teaching assistants, 42Teitelbaum, Philip, 5Telegraphic listeners, and grammatical

morphemes, 195Television inputand MSV acquisition, 270n.6and MSV understanding, 266and PPVT scores, 256

Textbooks. See also Psychologypublishing of, 51writing of, 51

Theater, H. Gleitman’s love of, 28, 36. Seealso Plays

Theory of mind (TOM)developing, 252–254and preschool experience, 264, 268

Thines, G., 157Thinkingand mental verbs, 247as perceptual experience, 249

Thinking out loud. See also Auditory im-agery

Index 359

Thinking out loud (cont.)reasons for, 154research on, 139–140

Thoughtseffect of emotion on, 239externalized forms of, 140

Toddlers. See also Childrenbound morpheme comprehension in,197–200, 202

comprehension studies in, 196and sentence structure, 280understanding in, 195

Tolman, Edward Chace, 2, 47Transformation and Discourse Analysis

Project (TDAP), 3Transitional probabilities, 108, 111Transitive sentencesvs. intransitive sentences, 287object arguments of, 286–287

Transparency phenomena, 158, 161Troscianko, T., 295Trueswell, John C., 327–345Tunneling, demonstration of, 306

Unit formation, Gestalt principles and, 161Unity, perceived, and relatability, 172UNIVAC2 computer, 3Universal Grammar (UG), 276Utterances, grouping of words into,

281–282

Vanderhelm, P., 174Vanlier, R. J., 174Verbal auxiliaries, position of, 13Verbal memory, in neuroimaging experi-

ments, 89–90Verb game, 17Verb meaningslearning of, 16presyntactic structural cues to, 279–280understanding of, 18

Verbs. See also Mental state verbs (MSVs)argument structures of, 333compared with nouns, 317–318interpretation of, 278learning, 278, 288perceptual, 250prosodic properties of, 193relational meaning of, 277second-syllable stress, 318semantic structures of, 281and sentence structure, 277

Verbs, action, in syntactic context, 220Verrekia, L., 315Versenoun-verb stress difference in, 318–319rhyme patterns in child, 321–324rhythmic structure of, 312–314

Virtue, emotion and, 241Vision, models of, 163Visual imagery, and Gestalt principles,

150–151Visual memory, 296–297Visual recognition, study of, 10Visual sequences, segmentation in, 110Vives, Juan Luis, 233Vocabularyof blind children, 211, 214and shape bias, 227and television input, 256–263

Vocabulary learning, and sentence struc-ture, 282–283

Wallach, Hans, 4, 23, 292Watson, John, 139Wellman, H. M., 82Wertheimer, M., 161, 181Wheeler, K., 294Whole units, words stored as, 202–203Williams, Dave, 27Word acquisition, for blind child, 212–213Word blends, 319–321elements in, 320predictions of, 320–321

Wordgleits, 28Word learning, and category membership,

216Word order patterns, 319–320Word recognitionargument structures in, 336–340syntactic aspects of, 340

Word segmentationand statistical learning, 107–108studies, 108–110

World events, in language acquisition, 277Wright, Thomas, 234Writing, detachment from, 155

Yantis, S., 297Yin, Carol, 180Yuille, J. C., 322

360 Index