Haskins Laboratories Status Report on Speech Research1992, SR-111! 112, 147-160

The Reading Process is Different for DifferentOrthographies: The Orthographic Depth Hypothesis*

Leonard Katzt and Ram Frost:j:

It has been said that most languages get theorthography they deserve and there is a kernel oftruth in that statement. There is generally anunderlying rationale of efficiency in matching alanguage's characteristic phonology andmorphology to a written form. Although the finalproduct may turn out to be more suitable for somelanguages than for others, there are certain basicprinciples of the fit that can be observed. Theattempt to make an efficient match between thewritten form, on the one hand, and morphologyand phonology, on the other, typically determineswhether the orthography chosen is a syllabary, asyllabary-cum-logography, or an alphabet.Further, within the group of alphabeticorthographies itself, there are varying degrees ofdependence on the strict alphabetic principle: therange of correspondence between grapheme andphoneme varies both in consistency andcompleteness. The degree of this dependence is tosome extent a function of a language'scharacteristic phonology and morphology, just aswas the choice of .the kind of orthography itself.We discuss here what this varying dependence onthe alphabetic principle may mean for the mentalprocesses involved in reading and writing.

Diversity in writing systemsAlthough writing systems are, in general terms,

systems for communication, what they actuallycommunicate is the spoken language-as opposedto communicating nonverbal ideas and meanings.

Work on this paper was supported by National Institute ofChild Health and Human Development Grant HD-01994 toHaskins Laboratories. The authors wish to thank 1. G.Mattingly for comments on an earlier draft.

147

DeFrancis (1989) reinforced this point with hisanalysis of so-called pictographic languages,writing systems whose elements are picturesand symbols that do not stand for words.DeFrancis argued that true pictographic systemsare not, in principle, effective and showed thatexisting examples of pictographic systems hadbeen designed only as novelties or playfulcommunication systems. In practice, they werenever used to communicate without substantialancillary aid from spoken language. The exampleof pictographic writing emphasizes the poverty ofwritten communication that is not based onlanguage. Therefore, because writing systems aresystems for representing the spoken language, itis reasonable to suggest that an understanding ofthe psychological processing involved in using awriting system must include an understanding ofthe processing of the spoken language. Althoughthe former will not necessarily parallel the latter,it will be constrained by it. A major constraintarises from the spoken language's morphemeswhich are the smallest units that carry meaning.It is these morphemes of speech that will be thefocus of communication, both spoken and written.Word stems are all morphemes and so are theirderivational and inflectional affixes; these units ofthe spoken language must be easily recoverablefrom the written language.

A large number and variety of writing systemshave flourished, evolved and developed, and inmany cases, died, over the centuries. Each of theknown systems can be categorized as eitherlogographic-phonetic, syllabic, or alphabetic(DeFrancis, 1989). These distinctions are made onthe basis of how a script (a set of symbols) relatesto the structure of its language. This relationshipbetween a script and its language is what isdescribed by the term orthography (Scheerer,

148 Katz and Frost

1986). The kind of script system and itsorthography are typically not wholly the result ofaccident. It is not accidental that the Chineselanguages, for example, have a logographic­phonetic system. In the Chinese orthography, thetypical character has two parts to it: a logographicand a phonetic component, the former providing avisually distinctive cue to the semantics and thelatter giving the reader a partial guide to thepronunciation. Together, the two componentsmake a combination that specifies completely aunique spoken morpheme. Words may be mono- orpolymorphemic.

Chinese morphemes are mainly monosyllabicand, because the variety of possible syllables islimited, there is a high degree of homophony inthe language. Such a language is best served byan orthography that distinguishes between thedifferent meanings of morphemes that soundalike. Instead, if the orthography had representedthe spoken form alone (e.g., only the phoneticcomponent in the printed word), the reader wouldnot be able to determine the intended meaning ofeach homophone except, possibly, from the wordor sentential context (many polymorphemic wordsare unique compounds of homophonousmorphemes)-but not without the additionalcognitive cost needed to resolve the ambiguity.The homophony problem is more of a problem fora reader than for a listener because the listenerhas more nonverbal contextual informationavailable to assist in determining word meanings.

But a pure logography would not suffice either,for Chinese. If the orthography had no phoneticcomponent, a reader would have to remember,without a phonetic cue, a pronunciation for each ofseveral thousand logographs. This would haveeffectively limited the number of printedcharacters that a reader could remember andname to an unacceptably small number. Instead,in modern Mandarin, it is necessary to rememberonly a small number of phonetic componentstogether with a smaller number of semantic signs.DeFrancis gives the number of these as 895 and214, respectively (DeFrancis, 1989). A phoneticand a semantic component are paired to produce acharacter; the effective set consists of 4300characters, the approximate number considerednecessary for full literacy.

In contrast to Chinese, spoken Japanese ispolysyllabic and is composed of regular syllable­like components, called moras. Because thenumber of syllables is small (fewer than 113), it isfeasible to represent them by means of a

syllabary. The Japanese orthography called kanawas such a system, adapted from Chinesecharacters. However, because there is a good dealof homophony in Japanese, the use of a syllabaryalone would not have been without problems anda logography also came into use. That logographyis still in use today and is routinely mixed withthe use of the kana, the syllabaries being usedprimarily for morphological affixes andgrammatical function words, foreign loan-words,and words not already covered by/the Chinese.

Indo-European languages have less homophonyand more polysyllabic morphemes than Chineseand Japanese. In addition, the structure of theIndo-European syllable itself is generally morecomplex than Chinese or Japanese, containing alarger number of phonologically permissibleclusters. English is said to have at least 8000syllables in its phonology, compared to fewer than1300 for Chinese (DeFrancis 1989). Eight­thousand is far too large a number for an effectivesyllabary. For English, an alphabet, representingphonemes, is more efficient for learning to readand write. Similarly, the Semitic languages(which include Arabic and Hebrew) would be lesssuitably represented by a syllabary than by analphabet. Like Indo-European languages, they toohave complex syllable structures. Historically, aconsonantal alphabet that developed for WestSemitic was the alphabet from which we trace theevolution of modern alphabets.

One of the characteristics of Semitic languagesthat may have led to the invention of the alphabetis the Semitic triconsonantal root: Semitic wordsthat are related by derivation or inflection have acommon core (usually three consonants). Althoughthe vowels in each of the different relatives maybe quite different and although there may be ad­ditional consonants in a prefix or suffix, there re­mains an invariant series of three consonants, theroot. This core has a strong linguistic salience-itrepresents a morphological communality of mean­ing among its family members (see Frost &Bentin, this volume). One can speculate that sev­eral early attempts were made at developing awriting system but only an alphabetic systemcould have captured this communality ofmorphology efficiently. Syllabic representationswould not be optimal: in Hebrew, morphemeboundaries fail to coincide with syllableboundaries (a condition that is true of Indo­European languages as well). Therefore, syllabicrepresentations would not be appropriate forrepresenting morphological units.

The Reading Process is Different [or Different Orthographies: The Orthographic Depth Hypothesis 149

The causes of diversity in alphabeticorthographies

Even among the various alphabetic writingsystems themselves there are major differences inthe degree to which they mirror the phonemicstructure of their respective spoken languages.Again, the reason for the differences is largelyaccounted for by the particular phonological andmorphological characteristics of each language.For example, standard written Hebrew is anorthography in which all diacritics (or points) areomitted. These diacritics represent nearly all ofthe vowels and are also used to disambiguatesome of the consonants. Nevertheless, writingwithout diacritics is usually sufficient to indicatethe exact intended (i.e., spoken) word if it issupported by a phrasal or sentential context.Thus, although the printed root may beinsufficient to allow an unequivocal identificationwhen presented in isolation, when it is presentedin a normal context-even a printed one-thecombined sources of information are enough forword identification.

In strong contrast to the Hebrew orthography isthe Serbo-Croatian. Serbo-Croatian is a majorlanguage of the Balkan Peninsula. Its presentalphabet was introduced in the early nineteenthcentury following the principle, "Spell a word likeit sounds and speak it the way it is spelled." Eachletter represents only one phoneme and eachphoneme is represented by only one letter.Moreover, no phoneme in the spoken word is everexcluded in the spelled word. The relationbetween letters and phonemes is isomorphic andexhaustive. To this day, the Serbo-Croatianspelling system follows the phonemic structure ofspoken words. So regular is the relation betweenspeech and writing that the writings of peoplewith different regional pronunciations will showdifferent spellings, mirroring the minordifferences in their spoken language. This simplewriting system works well for Serbo-Croatianbecause morphemic variations in the languagedue to inflection and derivation do not oftenproduce alterations in the phonemic structure ofword stems; word stems largely remain intact.

A differen.t state of affairs exists in English.English is somewhere between Hebrew and Serbo­Croatian in the directness with which its phonol­ogy is represented in its spelling; there is a largeamount of regular phonologic change amongwords in the same derivational family. Examplesof this are the contrasts between the derivationalrelatives HEAL and HEALTH, between STEAL

and STEALTH, etc. Chomsky and Halle (1968)argue that English spelling represents a morpho­phonemic invariance common to these word pairs,an abstract phonological communality that is be­low their surface difference in pronunciation. Inreading English aloud, the reader must either re­member the pronunciation of such a word as awhole or remember the appropriate context-de­pendent rule for pronunciation. An alternativewriting system might have spelled English in thesame way that Serbo-Croatian is spelled: with anisomorphic relation between letter and phoneme.However, that method would rob printed Englishof the advantageous common spelling for wordswith common morphology. The words HEAL andHEALTH, for example, might then be spelledHEEL and HELTH, disguising their commonmeaning. The printed form HEEL would also suf­fer from a double meaning as a consequence of itsbeing homophonic with the word meaning, "part ofa foot." English spelling represents a compromisebetween the attempt to maintain a consistent let­ter-phoneme relation and the attempt to representmorphological communality among words even atthe cost of inconsistency in the letter-phoneme re­lation.

Thus, alphabetic writing systems reflect thespoken forms of their respective languages withdifferent degrees of consistency and completenessbetween letter and phoneme. Some of thedifferences in writing systems have purelypolitical, cultural, or economic causes. But manydifferences have been motivated by two factorsthat are purely linguistic. The first has to do withhow complex the spoken language is in therelation between phonology and morphology-onlya phonologically complex language can have adeep alphabetic orthography. For example, Serbo­Croatian is not phonologically complex. Allmorphologically related words have a commonphonologically invariant core. Two words that aremorphologically related will share a common wordstem that will necessarily sound the same in bothwords. Both instances of that common stem will,of course, be spelled the same. Thus, whenevaluated by the characteristic of phonologicalcomplexity, a language that is not complex can bewritten (and generally will be written) in ashallow orthography, an orthography that tracksthe phonology. Secondly, if a language is one thatis phonologically complex then the orthographyhas the option of representing eithermorphological invariance (a deep orthography) orfollowing grapheme-phoneme invariance (ashallow orthography). As we suggested above,

150 Katz and Frost

English qualifies as quite complex, phonologically.In principle, it could have been written either as ashallow or a deep orthography. The advantage toEnglish in choosing a deep orthography is in theconsistent spelling of morphemic invariances.However, that choice having been made, there arethen different pronunciations of the same spellingon occasion (e.g., HE A L-HEALTH) and,inadvertently, identical pronunciations fordifferent spellings (e.g., PEEL-DEAL).

A different situation exists in Hebrew. Hebrew'sphonology is complex; morphemes may undergoconsiderable sound change under eitherinflectional or derivational change. On the otherhand, because of the pervasiveness of thetriconsonantal root in Hebrew, a great deal ofmorphological constancy exists. Therefore, therewas an historical choice, so to speak, for theevolution of the Hebrew orthography: It couldhave opted for either morphemic or phonemicinvariance but, unlike Serbo-Croatian, it could nothave contained both in a single orthographybecause of its phonological complexity. Hebrewinitially evolved as an orthography in which themorphology was preserved at the expense ofphonological completeness. Vowels were omittedthereby emphasizing the morphologically basedconsonantal invariance in a given family of wordroots. Vowel points were added to the script at alater stage in the orthography's development onlybecause the language was no longer being spokenas a primary language and it was feared that itspronunciation would become corrupted unlessvowels were included in the script. Nowadays, theorthography used by adults is the unpointed one,which is graphemically incomplete and somewhatinconsistent to the reader because it omits nearlyall of the vowels and makes some of theconsonants ambiguous.

In summary, all alphabetic orthographies can beclassified according to the transparency of theirletter-to-phoneme correspondence, a factor thathas been referred to as orthographic depth(Liberman, Liberman, Mattingly, & Shankweiler,1980). An orthography in which the letters areisomorphic to phonemes in the spoken word(completely and consistently), is orthographicallyshallow. An orthography in which the letter­phoneme relation is substantially equivocal is saidto be deep (e.g., some letters have more than onesound and some phonemes can be written in morethan one way or are not represented in theorthography). Shallow orthographies arecharacteristic of languages in which morphemicrelatives have consistent pronunciations.

Differences among alphabetic orthographiesin processing printed words: The orthographicdepth hypothesis

Our discussion to this point has made thestandard argument that there are differencesamong alphabetic orthographies in orthographicdepth and that these differences are a result ofdifferences in their languages' phonology andmorphology. In this section, we propose that thedifferences in orthographic depth lead toprocessing differences for naming and lexicaldecision. This proposal is referred to as theorthographic depth hypothesis (ODH). It statesthat shallow orthographies are more easily able tosupport a word recognition process that involvesthe language's phonology. In contrast, deeporthographies encourage a reader to processprinted words by referring to their morphology viathe printed word's visual-orthographic structure.

We would like to make two points, eachindependent of the other. The first states that,because shallow orthographies are optimized forassembling phonology from a word's componentletters, phonology is more easily available to thereader prelexically than is the case for a deeporthography. The second states that the easier itis to obtain prelexical phonology, the more likely itwill be used for both pronunciation and lexicalaccess. Both statements together suggest that theuse of assembled phonology should be moreprevalent when reading a shallow orthographythan when reading a deep orthography. Becauseshallow orthographies have relatively simple,consistent, and complete connections betweenletter and phoneme, it is easier for readers torecover more of a printed word's phonologyprelexically by assembling it from letter-phonemecorrespondences.

Suggested by the above is our assumption thatthere will always be at least some dependence onphonological coding for the process of reading inany orthography. That is, the processing of (atleast) some words will include assembledphonology (at least in part). This assumption canbe easily motivated for alphabetic orthographies.The assembling of phonology has a certainprecedence in a reader's experience; instruction inreading typically means instruction in decoding,i.e., learning how to use letter-phonemecorrespondences. It is well established thatbeginning readers find it easier to learn to read inshallow orthographies, where thosecorrespondences are most consistent (see, forexample, Cossu, Shankweiler, Liberman, Katz, &

The Reading Process is Different for Different Orthographies: The Orthographic Depth Hypothesis 151

Tola, 1988). Even in learning to read Hebrew,instruction is typically given in the shallowpointed orthography instead of the deep unpointedone (the transition to the unpointed formbeginning in the third grade). In any orthography,after learning to read by using assembledphonology routines, skilled readers may continueits use to the extent that the cost of doing so islow. This will be particularly true when theorthography is shallow. However, given theexperimental evidence, some of which we discusslater, it seems certain that assembled phonology isnot used exclusively. More likely, a mix of bothassembled phonology and visual-orthographiccodings are nearly always involved, even inshallow orthographies.

Two versions of the orthographic depthhypothesis

Two versions of the orthographic depthhypothesis (ODH) exist in the current literature.What can be called the strong ODH states thatphonological representations derived fromassembled phonology alone are sufficient fornaming and lexical decision in shalloworthographies. Thus, according to the strong ODH,rapid naming in shallow orthographies is a resultof only this prelexical analytic process and doesnot involve pronunciation obtained from memory,i.e., the lexicon. However, we submit that thestrong form of the ODH is patently untenablewhen applied to the orthographies that havetypically been used in research on wordperception. It is insufficient to account forpronunciation even in a shallow orthography likeSerbo-Croatian. This is so because Serbo-Croatiandoes not represent syllable stress and, eventhough stress is often predictable, it is not alwayspredictable. Because the final syllable is neverstressed, stress is completely predictable for two­syllable words but for words of more than two, it isnot. However, one- and two-syllable words makeup a large part of normal running text so much ormost of the words a reader encounters can bepronounced by means of a prelexical subwordanalysis. But, of course, many words will begreater than two syllables in length and these canbe pronounced correctly only by reference tolexically stored information. In addition, there aresome exceptions to the rule that a letter mustrepresent only one phoneme; some final consonantvoicing changes occur in speech that are notmirrored in the conventional spelling (thesechanges are predictable, however). Thus, Serbo­Croatian, although it should be considered an

essentially shallow orthography, is not the perfectparadigm of a shallow orthography. We should notexpect a strong ODH to make sense for such anorthography.

We support the weak ODH. In this version, thephonology needed for the pronunciation of printedwords comes not only from prelexical letter­phonology correspondences but also from storedlexical phonology, that is to say, from memory.The latter is the result of a visual-orthographicaddressing of lexicon, i.e., a search process thatmatches the spelling of a whole word or morphemewith its stored phonology. The degree to which aprelexical process is active in naming is a functionof an orthography's depth; prelexical analyticprocesses will be more functional (less costly) inshallow orthographies. However, whether or notthese prelexical processes actually dominateorthographic processing for any particularorthography is a question of the demands the twokinds of processes make on the reader's processingresources, a question we discuss further below. Weproposed (and supported) this weak form of theODH in Katz and Feldman (1983) and Frost, Katz,and Bentin (1987); further details are given laterin this chapter.

With regard to word recognition (as in lexicaldecision), some of our colleagues have argued thatSerbo-Croatian necessarily involves prelexical(i.e., assembled) phonology (Feldman & Turvey,1983; Lukatela & Turvey, 1990a). Others havemade a similar claim for the obligatoryinvolvement of prelexical phonology in English(Van Orden, Pennington & Stone, 1990; Perfetti,Bell, & Delaney, 1988). However, theseresearchers have not argued for the exclusiveinvolvement of assembled phonology. Logically,assembled prelexical phonological information,without syllable stress information, is sufficient toidentify the great majority of words in the Englishlexicon. However, irregularly spelled words,foreign borrowings, etc., would pose problems foran exclusively phonological mechanism, and,therefore, such a view seems less plausible.Finally, note that we are confining this discussionto the problems of naming and lexical decision; itis an entirely different question to ask whetherphonological representations are necessary forpostlexical processes like syntactic parsing andtext comprehension.

Evidence on the questions of phonologicalrecoding and the weak ODH

We discuss next the evidence for the hypothesesthat the lexicon is addressed by assembled

152 Katz and Frost

phonology, presumably in combination withvisual-orthographic addressing, and that thespecific mix of the two types of codes depends onorthographic depth. We show why single-languagestudies, in general, are not suitable for testing theweak ODH and mention the few exceptions.Experiments that directly compare orthographieswith each other provide the most direct evidence.We will argue that these cross-languagecomparisons are absolutely critical to aninvestigation of orthographic depth effects.

It is important to realize that, in the controversyover whether visual-orthographic recoding orphonological recoding is used in word perception,there is little direct evidence of visual­orthographic effects. Instead, the burden of proofis placed on assembled phonology; if no effect ofphonology is found, then visual-orthographiccoding is said to win, by default. The assumptionis not unreasonable because a visual-orthographicrepresentation is obviously available in principleand seems to be the only alternative to assembledphonology. However, it should be kept in Illindthat, because of this, the experimental evidencehinges on the sensitivity of the experimental taskand its dependent measures to phonology. If theyfail to indicate the presence of phonology, it maynot be because phonology is not operative.

In fact, several experiments have demonstratedthat phonological recoding effects can be foundeven in deep orthographies. In Hebrew, Frost hasshown that, if available, the full phonology givenby the pointed orthography is preferred fornaming; this seems to suggest that even if wordrecognition does not normally proceed viaassembled phonology in Hebrew, the recognitionprocess is prepared to default to its use (Frost, inpress). In English, Perfetti and his associates andVan Orden and his associates have presented·strong evidence for phonological recoding inlexical access (Perfetti, Bell, & Delaney, 1988,Perfetti, Zhang, & Berent, 1992; Van Orden,Pennington, & Stone 1990; Van Orden, Stone,Garlington, Markson, Pinnt, Simonfy, &Brichetto, 1992).

However, it is one thing to find the activepresence of phonological recoding, another todetermine the conditions under whichphonological recoding occurs, and yet another todetermine the degree to which naming or lexicalaccess is dependent on it. Is assembled phonologyobligatory or is it rarely used; if neither of theseextremes, is it the more preferred or the lesspreferred code? In this vein, Seidenberg(Seidenberg, 1985; Seidenberg, 1992) has argued

that word frequency is the primary factor thatdetermines whether or not assembled phonology isused to access the lexicon. His argument is that inany orthography, whether deep or shallow,frequently seen words will become familiar visual­orthographic patterns and, therefore, rapid visualaccess will occur before the (presumably) slowerphonological code can be assembled from theprint. Low frequency words, being less familiar,produce visual-orthographic representations thatare less functional; lexical activation builds upmore slowly. This gives time for phonologicalrecoding to contact the lexicon first. However, wecannot presently answer questions that areconcerned with the relative iIllportance of wordfrequency to orthographic depth or concerned withthe relative dominance of the two kinds ofrepresentation, visual-orthographic andphonological. But we can meaningfully addressthe question of whether or not the relative amountof assembled phonological coding decreases withincreasing orthographic depth: the orthographicdepth hypothesis.

Comparisons across orthographiesCross-language experimentation, in which dif­

ferent languages are directly compared, are thecritical methodology for studying the orthographicdepth hypothesis. Single-language experimentsare adequate for testing only the strong form ofthe ODH, in which shallow orthographies are saidto never use lexically stored information for nam­ing-but, as we showed, this is a claim that can berejected on logical grounds. Single-language ex­periments are not without interest, however; theycan be useful in indicating how easy it is to findeffects of phonological coding. This may suggest­but only suggest weakly-what the dominant rep­resentation is for an orthography. If it is easy tofind effects of phonological coding in Serbo­Croatian and difficult to find those effects inHebrew, using more or less similar experimentaltechniques, we may suspect that phonological cod- .ing is the dominant (preferred) code in Serbo­Croatian but not in Hebrew. However, such exper­iments can not rule out the additional use of thealternate type of representation for either orthog­raphy. In this vein, we know that it is difficult tofind effects of phonological coding in English usingstandard lexical decision paradigms but, neverthe­less, phonological effects can be found usingPerfetti's backward masking paradigm, which isapparently more sensitive (Perfetti, 1988). Finally,however, an accurate answer to the question ofwhich type of representation is dominant for a

The Reading Process is Different for Different Orthogri1phies: The Orthographic Depth Hlfpothesis 153

particular orthography can not be given by thecurrent experimental paradigms; the results fromthese experiments may only reflect the adequacyof the paradigms in capturing the true wordrecognition process.

The weak form of the ODH proposes that (1)both orthographic information and prelexicaliy as­sembled phonological information are used forlexical access and (2) the degree to which one kindof information predominates is a function of thestructural relationship between orthography andthe lexical entry. The Serbo-Croatian orthography,with its simple and consistent letter-phoneme re­lationships, makes it easy for the reader to learnand maintain the use of assembled phonology.This assembled phonology must address,presumably, the same abstract phonologyaddressed by a listener's spoken language lexicon(although it may be only a subset of the fullspoken phonology, because of the absence of stressinformation, at the least). In contrast, the Hebreworthography, because it lacks most of the vowelsand has many ambiguous consonants, is incapableof providing enough assembled phonology that willconsistently identify a unique word in thephonological lexicon (only the consonants can beassembled); therefore, there are fewer benefits ingenerating ph,mological information byassembling it from grapheme-phoneme correspon­dences. These are lessons that the developingreader can learn tacitly, lessons that may lead,eventually, to different dominant modes of printedword processing for Serbo-Croatian and Hebrewreaders. For languages that are in between thesetwo extremes, the relative balance of assembledphonology to orthographic representation shouldreflect the relative efficacy of the two kinds of

. information in that orthography; some letters orletter-sequences may be simple and consistent andthe assembled phonology derived from these maybe used along with orthographic information. It isnot possible to make a more precise statementwithout an understanding of the details of thelexical recognition process and the processingresources that are required. For example, for avisual-orthographically coded word to berecognized, a mental representation of that wordmust have been created previously and stored inthe reader's memory. We do not know how tocompare the resources needed to create a neworthographic representation with the resourcesneeded to generate assembled phonology; which ismore demanding? Neither can we automaticallyassume that it is easier to access lexicon via avisual-orthographic representation.

Additional complications arise when we try to bemore specific about the phonological nature of theinformation in the lexicon itself; what, exactly, isthe information that is represented in lexicon thatis, presumably, addressed by assembled phonol­ogy? Alphabets mainly represent phonemes butare words in the spoken lexicon to be representedas phoneme sequences? If so, why do syllabic or­thographies work at all, since the printed units ofsyllabaries map onto syllables, not phonemes? IIifact, there are several different theoretical de­scriptions that have been proposed by speech re­searchers for the lexical representation of a word.Perhaps, the spoken lexicon contains multiplephonological descriptions of a single word, e.g.,phonetic, phonemic, syllabic, gestural, etc. Thephonology produced by reading may be a subset ofthe information constituting the spoken lexicon orit may even be different in kind (although related).We do not propose to discuss this problem in de­tail here, but only wish to point out that there is acompanion to the question of how phonologicalinformation is used in reading, namely, the ques­tion of the nature of the phonological informationthat is used in spoken word recognition.

Evidence supporting the orthographic depthhypothesis

Some early evidence that there is a relationshipbetween orthographic depth and lexical accesswas obtained by Katz and Feldman (1981). Theycompared lexical decision times in Serbo-Croatianand English for printed stimuli that were dividedwith a slash character. The stimuli were dividedeither at a syllable boundary (e.g., WAITER) orone character to either the left or right of theboundary (e.g., WIATER or WATIER). If wordrecognition involves recoding of the stimulus to aphonological form, and if that phonology includesthe syllable as a unit, then division at a syllableboundary-which preserves the syllable units­should be less disruptive than division off theboundary. Pseudowords were similarly divided.Lexical decisions to words and pseudowords thatwere irregularly divided were slower than lexicaldecisions to their regularly divided counterpartsand the disruptive effect of irregular division wasstronger for Serbo-Croatian. The data, then, wereconsistent with a model of word recognition thatassumes the operation of at least somephonological recoding of print prior to lexicalaccess and, further, assumes that phonologicalrecoding is a more consistent determiner of accessin shallower orthographies (e.g., Serbo-Croatian)than deeper ones (e.g., English).

154 Katz and Frost

In a second study, Katz and Feldman (1983)made a direct test of a second prediction of theorthographic depth hypothesis: that pronouncinga word (naming) depends more on assembledphonology in Serbo-Croatian than in English. Alexical decision experiment and a namingexperiment were run in both languages; stimuluswords had the same (or similar) meanings in bothlanguages. The subjects were native speakers whowere tested in their native countries (Serbia or theUnited States), in order to avoid subjects whowere fluent in both languages. This was donebecause the experience of a bilingual speakermight have affected his or her strategy forreading.

Each test stimulus (the target), whether it was aword or a nonword, was always preceded by thebrief (600 ms) presentation of a real word. On halfof those trials when the target was a word, thispredecessor was semantically related to the target(e.g., MUSIC-JAZZ); on the other half, it wasunrelated. Words also preceded the nonwordtargets. It is well established, that preceding astimulus with a semantically related word willfacilitate and speed a lexical decision response tothe target. Thus, it was expected that reactiontime to the target JAZZ would be faster for thosesubjects who saw it preceded by the word MUSICthan for those subjects who saw it preceded by theword GLASS, which has no st:r::ong semanticrelation to the target. The critical fact for thiskind of experimental technique is that thefacilitating effect of MUSIC can only occur byactivating the semantic link between it and JAZZand this linkage necessarily must be within thelexicon. Thus, to the extent that there isfacilitation in the subject's recognition of JAZZ, itindicates that recognition is being assisted byactivity within the lexicon. Such lexical activitymay facilitate whether the lexicon is addressed byorthography or by phonology.

For naming, however, the prediction is different.Naming can be accomplished largely withoutaccessing the lexicon by means of subword letter­to-phonology recoding. Of course, in neitherEnglish or Serbo-Croatian can the process beentirely without reference to lexical memory,because the stress of polysyllabic words is notspecified in the orthography. Nevertheless, theprocess of naming can, in principle, be carried outsubstantiallY without reference to the lexicon.Thus, if naming in Serbo-Croatian is moredependent on phonological recoding (and lessdependent on lexical look-up) than English,naming in Serbo-Croatian ought not to be affected

by the semantic priming manipulation, which isnecessarily lexical in its locus of operation.Results supported this suggestion: target wordsthat were preceded by semantically related words(e.g., MUSIC-JAZZ) were pronounced faster thantarget words that were preceded by unrelatedwords (e.g., GLASS-JAZZ) in the case of Englishbut not in the case of Serbo-Croatian. In contrast,there were equivalent strong effects of semanticpriming for lexical decisions, in both languages.

A three-way comparison of Hebrew, Serbo­Croatian, and English increased the range oforthographic depth examined in a single study(Frost, Katz, & Bentin 1987). Necessary to thesuccess of any cross-language experiment is a setof stimulus words that have equivalent criticalcharacteristics in all the languages under consid­eration: for example, equivalent frequencies ofoccurrence. Subjectively estimated frequencies ofoccurrence were obtained and equivalent sets ofstimulus words were used in all three testlanguages. The important comparison, over thethree orthographies, was the relation betweennaming and lexical decision reaction times forwords versus nonwords. The importance of thiscomparison was based on the rationale that inshallow orthographies the lexicon plays only aminor role in the naming process compared to itsrole in the lexical decision process. The oppositeassumption, i.e., that even in shalloworthographies, a skilled reader always employsthe orthographic route to accessing the lexicon,predicts that readers in all three orthographiesshould perform similarly. On the other hand, ifthe orthographic depth hypothesis is correct, thegreatest difference between naming and lexicaldecision reaction times should be in Serbo­Croatian, which has the shallowest orthographywhile Hebrew should show the greatest similarity.Results were in line with the orthographic depthhypothesis; naming times were considerably fasterthan lexical decision times in Serbo-Croatian but,in Hebrew, lexical decision and naming lookedquite similar. In Hebrew, it took as long to name aword as to recognize it: a suggestion that namingwas accomplished postlexically. In addition, inSerbo-Croatian, the faster responding for namingversus lexical decision was even greater forpseudowords than for words. In thesecomparisons, English was intermediate. Thus, theresults support the hypothesis that the shallowerthe orthography, the greater the amount ofphonological recoding that is carried out innaming. Subsequent experiments in this study,which maximized the potential for lexical

The RelUiing Process is Different for Different Orthographies: The Orthographic Depth Hypothesis 155

processing by semantically priming target wordsand by varying the relative number of pseu­dowords further supported this interpretation.

In all the experiments we have discussed to thispoint, the experimental paradigms used have beennaming and lexical decision tasks. These taskshave a disadvantage as methodologies because thephonological variation that is used to affect thesubject's response (e.g., the consistency of thegrapheme - phoneme relation) is obtained throughmanipulating the orthography (e.g., differentalphabets) and not by manipulating the putativephonology directly. The experimenter neverobserves any phonologic recoding; its presence isonly inferred. Thus, one can not be certain thatthe differences that are observed are true effectsof phonological recoding or, instead, are only theresult of orthographic effects which happen to becorrelated with phonology. Frost and Katz (1989)addressed this issue by introducing a paradigm inwhich subjects had to compare a spoken word anda printed word. This paradigm requires subjects toperceive and use phonology in their taskprocessing. Subjects were required to simul­taneously read and listen to two words presentedby computer and judge whether or not they werethe same (i.e., represented the same lexical item).In order to make the comparison, the subject hadto mentally place both spoken and printed stimuliinto a common representation. This could havebeen done, in principle, in several ways, althoughonly two possibilities seemed reasonable. Thespoken word could have been imagined as aspelled (printed) word or subjects could havegenerated the phonology of the printed word. Theevidence indicated that subjects chose the latter.This was not surprising: subjects have had farmore practice reading than spelling. Afterconverting the printed stimulus to a phonologicalrepresentation, both phonological representationscould then have been compared in order todetermine if they matched. Over a list of 144 ormore trials, subjects made the judgment "Same"or "Different" about each pair of printed andspoken words on each trial. There were threeconditions: clear speech and clear print, degradedspeech (noise added) and clear print, and clearspeech and degraded print (visual noise added).Serbo-Croatian and English native speakers weretested on comparable materials. The effects ofdegrading were marked; when either the print orthe spoken word was degraded, performancedeclined sharply. However, the difference inlatency between the slower responses to print or

speech that had been degraded compared to clearprint or speech was four times greater in theorthographically deep English than the shallowerSerbo-Croatian; degradation had a much strongerdeleterious effect in English.

An interactive activation network model can beextended easily to account for these results. Themodel contains parallel orthographic andphonologic systems that are each multilevel withlateral connections between the two systems atevery level. In particular, the sets of graphemicand phonemic nodes are connected together in amanner that reflects the correspondences in aparticular orthography: mainly isomorphicconnections in a shallow orthography and morecomplex connections in a deep orthography. Thesimple isomorphic connections in a shalloworthography should enable subjects to use theprinted graphemes to activate their corresponding(unambiguous) phonological nodes, supplementingweaker activation generated more directly bydegraded speech. This higher, aggregated,activation should reach threshold fast compared toa network representing a deep orthography withits weaker grapheme-phoneme connections.

Evidence against the orthographic depthhypothesis

We mentioned above the study by Seidenberg(1985) who studied word naming in Chinese andEnglish. In English, he found no differencebetween regularly spelled words and exceptionwords as long as their word frequency was high,suggesting that phonologic representations playno role in naming frequent words. Differenceswere found, however, for low frequency words,exception words having the longer latencies. InCantonese, an analogous pattern was found:There was no significant latency differencebetween phonograms (compound characters thatcontain both a phonetic component and alogographic signific) and non-phonograms(characters that contain no phonetic)-as long asthey were high frequency. For low frequencyitems, phonograms were named faster. However,what seemed to drive naming latency moststrongly in both languages was word frequency.The results suggest that the effect of frequency, aneffect that was similar in both orthographies, maybe of overriding importance in determining whichkind of lexical access code is successful;differences between orthographies that can affectthe coding of phonology may be irrelevant whenfrequent words are compared.

156 Katz and Frost

Besner and Hildebrandt (1987) capitalized onthe fact that Japanese is written in three scriptsystems. One of the scripts is a logography that isderived from the Chinese but one in which thecharacters are pronounced differently. Two ofthese scripts are essentially syllabic orthogra­phies, katakana and hiragana. Historically, theirgraphemes evolved from separate sources but theyboth address the same phonology, similar in thisregard to the dual Cyrillic and Roman alphabetsof Serbo-Croatian. Unlike Cyrillic and Roman,however, the Japanese scripts are rarely used towrite the same words; instead, they "specialize,»being used for mutually exclusive vocabularies. InJapanese, then, the pronunciations of those wordsthat are logographic must be recalled lexically.However, those words that are normally printedin katakana and those words that are normallyprinted in hiragana can, in principle, bepronounced via grapheme-syllable corre­spondences. In a simple but direct experimentthat compared only the katakana and hiraganascripts, subjects named words that were printedeither in their normal script or in the other script.When printed in the normal script, naming timeswere 47 to 65 milliseconds faster than whenprinted in their atypical script. Besner andHildebrandt interpreted the results to mean thatsubjects were not using grapheme-syllablecorrespondences in order to pronounce thenormally printed stimuli because changing thevisual-orthographic form had been detrimental; ifsubjects had been assembling the phonology fornaming, they would not have been slowed by thechange in grapheme-syllable script system. Thus,there is reason to suspect that Japanese readersalways adopt a visual-orthographic mode fornaming no matter the depth of the script systemthey are reading: the deep logography or theshallow syllabaries. Besner and Smith (1992) offerfurther details.

Additional evidence against the orthographicdepth hypothesis was reported by Baluch andBesner (1991). They studied naming in Persian,an orthography which offers a comparisonbetween words that omit the vowels, like all wordsin Hebrew (opaque words) and other words thatare spelled with a relatively full phonologicalspecification, like Serbo-Croatian (transparentwords). The difference lies in the representation ofthe vowels; opaque words have one or more ofcertain specific vowels that can be written asdiacritics but, instead, are typically omitted fromthe spelling (as in Hebrew) while transparentwords contain only those vowels that are never

omitted. The authors found that semantic primingequally facilitated both transparent words andopaque words; the weak orthographic depthhypothesis would predict less facilitation for thetransparent words, which can be pronouncedlargely via assembled phonology, needing lexicalinformation only for syllable stress. Differencesbetween opaque and transparent words didappear when pseudowords were included in thelist of words to be pronounced; then, only theopaque words were facilitated. The inclusion ofpseudowords presumably biased the subjecttoward the use of addressed phonology as thedefault because the pseudowords would have hadno addressed phonology and grapheme-phonemecorrespondence rules were therefore an effectivealternative. Apparently, when the recognitionprocess is biased in this way, transparent wordsare no longer processed via visual-orthographiccoding and they are no longer facilitated bysemantic priming. Their results suggest that innormal reading, where there are no pseudowords,subjects may always use the direct lexical route,without the use of assembled phonology. Baluchand Besner note that in all the studies in whichsubjects used phonological recoding, pseudowordshad been included in the stimulus list of words,perhaps biasing subjects toward an atypicalprocessing strategy (e.g., (Katz et al., 1983; Frostet al., 1987; etc.).

A similar point was made by Tabossi and Laghi(1992). In a clever series of experiments, theyshowed that semantic priming effects in naming,which are indicative of lexical processing,disappeared or were attenuated . whenpseudowords were introduced. The implication is,then, that visual-orthographic coding was thepreferred strategy for their subjects. The authorsinterpret their results to suggest that assembledphonology is produced only under artificialconditions such as when pseudowords are present.Because their experiments used a shalloworthography (Italian), the authors suggest that allorthographies, shallow and deep, use the samemechanism for processing print, i.e., the visual­orthographic route.

However, the alternative explanation, the ODR,is not directly addressed by their study (with oneexception, discussed shortly). As we suggestabove, no standard experiment on a single scriptsystem will be able to test the claim of the ODRthat the amount of lexical involvement is greaterin shallow than in deep orthographies. The ODRis a statement about relationships amongorthographies; it does not categorically disallow

The Readin'{ Process is Different for Different Orthographies: The Orthographic Depth Hypothesis 157

the use of either assembled phonology or visual­orthographic processing for any orthography(except for the strong form of the hypothesis,which is clearly unacceptable on rationalgrounds). Thus, it is still not known if somephonological processing is occurring in Italian butis not being observed because of specialcharacteristics of the experiment itself (e.g., thetask, stimuli, etc.). As in all situations in whichthe experimenter is pressing the null hypothesis("Can we show that no phonological processing isoccurring?"), the better the set of alternativemodels for comparison, the more convincing theoutcome. In the case of the ODH, the moreconvincing argument against it would be to showthat manipulations that affect semanticfacilitation cause identical effects (do not causedifferential effects) between Italian and someorthography that is deeper than Italian. Tabossiand Laghi (1992) do, in fact, make this test; whenthey do, they find evidence that is consistent withthe ODH. Semantically primed words were namedfaster than controls in English but not in Italian,suggesting that naming involved the lexicon morein the deeper English than in Italian. However,both word lists contained pseudowords whichwould tend to increase the amount of phonologicalprocessing for both.

A similar demonstration was made in the Frost,Katz, and Bentin (1987) study. In Experiment 3,the authors explicitly examined the samehypothesis that put forward by Baluch and Besner(1991). In this experiment the ratio of words topseudowords in the stimulus list was manipulatedand its effect on naming was measured in Hebrew,English, and Serbo-Croatian. The results showedmarked differences in the pseudoword ratio effectin the three different orthographies. Whereas inSerbo-Croatian the inclusion of pseudowords hadalmost no effect on naming latencies (consistentwith the notion that assembled phonology is thepreferred strategy for that orthography), muchlarger effects were found in English and Hebrew.The point raised by Baluch and Besner is indeedimportant; pseudowords can, in fact, affectnaming strategies. However, this issue has nodirect relevance to the ODH. The ODH suggeststhat the relative effect of pseudoword inclusionshould be different in deep and in shalloworthographies. The results of both Frost et al.(1987), and Tabossi and Laghi (1992) arecompatible with this notion.

Sebastian-Galles (1991) presented evidence thatwas said to be inconsistent with theories that

propose different mechanisms for shallow anddeep orthographies. Spanish subjects pronouncedpseudowords that had been derived from realwords; each pseudoword was orthographicallysimilar to its counterpart, except for one or twoletters. In some pseudowords, the correctpronunciation of a critical letter (c or g) changedfrom its pronunciation in the real word accordingto grapheme-phoneme correspondence rules,because of a change in the following vowel. Inother pseudowords, there was no such change inthe vowel (and, therefore, no change in thepronunciation from the real word model). Subjectspronounced about 26% of the change pseudowordscontrary to the correspondence rules while onlyabout 10% of the no-change pseudowords werepronounced in that way. Sebastian-Gallesinterpreted this result to mean that subjects wereusing a lexical strategy for pronouncingpseudowords. But this interpretation is warrantedonly if the theory being tested allows no lexicalinvolvement at all in naming: The author wasattacking the strong form of the ODH. A closerlook at the evidence suggests that the data are, infact, consistent with the weak ODH. Seventy-fourpercent of the change pseudowords werepronounced in accordance with spelling-to-soundcorrespondence rules and only 26% werepronounced "lexically." Thus, a mix of the twoprocesses may have been at work. In a secondexperiment, comparing lexical decision andnaming times, Sebastian-Galles found a moderatecorrelation (.455) for latencies between the twotasks. When the latency on each task wascorrelated with word frequency (presumed to bean index of lexical involvement), the correlationwas of greater magnitude for lexical decision(-.497) than for naming (-.298). This result isconsistent with a continuity of lexicalinvolvement, naming being under weaker lexicalcontrol than lexical decision. A final experiment inthis series showed semantic priming for namingunder conditions where it is usually not found inshallow orthographies, viz., when pseudowordsare included in the list of stimuli. Such results dosuggest more lexical involvement in naming for ashallow orthography than other research hassuggested. Nevertheless, in the author'sconclusion, Sebastian-Galles interprets the sum ofthe results to mean that "... lexical access fromprint in Spanish involves the use of orthographicinformation during at least some of the processingtime," a statement that is consistent with theweakODH.

158 Katz and Frost

Concluding remarks

Our working hypothesis has been that all al­phabetic orthographies make some use of assem­bled phonology for word recognition. The proposalthat a mixture of prelexical and visual-ortho­graphic information is used for word recognition isconsistent with the weak form of the ODH. Theapproach we suggest is in line with those modelsthat contain dual phonological and orthographicrepresentations, such as the dual route models(see Paap, Noel, & Johansen, 1992) and networkmodels (see Seidenberg, 1992). The question ofjust how prevalent the use of phonology is,relative to the use of visual-orthographic coding,for any given orthography is an open one and isnot addressed by the ODH itself. It could even bethe case that the predominant lexical access codefor frequent words in the shallow Serbo-Croatianis actually visual-orthographic or, on the otherhand, that the predominant code in Hebrew isbased on the partial phonological information thatcan be assembled from the unpointed letters(although either possibility seems unlikely fromthe present evidence). The ODH does not specifYwhat degree of orthographic depth determinespredominance for, say, visual-orthographic coding.Of course, orthographic depth will not be the onlydeterminer of dominance; the reader's experienceshould playa major role as well.

Which of the codes is dominant might bedetermined by a mechanism like the following.Assume that the processing system is capable ofusing either code: a dual-code model. Suppose thata phonological representation is the default codefor any given word but processing of a word via itsphonological code can be replaced by processingvia its visual-orthographic representation whenthe word has been experienced by the reader asufficient number of times: a word frequencycriterion. The premise that phonology is thedefault code is based on the fact that it is typicallythe code of instruction and the beginning readerreceives much practice in its use. (One piece ofevidence supporting the notion that the defaultcode is phonological is that even adult readers ofHebrew prefer to use phonological information if itis available; Frost, in press). The criterion wordfrequency that is required in order to replaceprocessing by assembled phonology withprocessing by visual-orthographic representationsshould be a function of the costs involved-in part,the cost for assembling the phonologicalrepresentation and using it to access the lexicon.A higher replacement criterion will obtain in a

shallow orthography where assembled phonologyis easy to generate than in a deeper orthography.

What other factors affect the cost? Besides theease of assembling phonology, a second factor isthe ease with which phonology can be used in thelexical search process itself. Likewise, the ease ofgenerating a visual-orthographic representationthat is suitable for lexical search, the ease withwhich that information can address the lexicon,and, of course, the cost involved in establishingthe visual-orthographically coded lexical represen­tation in the first place, all need to be evaluated inestablishing the criterion. The tradeoff betweenthe generation of information that can be used forlexical access and the access process itself is im­portant. Visual-orthographic codes may be costlyfor both the beginning and skilled readers to gen­erate, particularly if a word is morphologicallycomplex (it may require decomposition, in thatcase). However, the search process based on a vi­sual-orthographic representation may be rapid forthe skilled reader once he or she has a well-estab­lished visual-orthographic representation in lexi­cal memory. Phonological codes may be difficult togenerate but, once obtained, may be a fairly natu­ral (i.e., low cost) way of addressing lexicon; afterall, our primary lexicon, the speech lexicon, isbased on phonology. Although the phonologicallexicon may have taken years for a child to de­velop, it is thereafter available free to the readingprocess. However, it is obvious that we know verylittle about this tradeoff in terms of processingcosts. The answer to the question about which ofthe two representations, phonological or visual-or­thographic, is dominant depends on knowing morethan we presently do about the perceptual andcognitive resources involved in word recognition.This question is discussed in some detail bySeidenberg, (1992). Nevertheless, we repeat thateven in the absence of a fuller understanding ofhow the word recognition process draws on theseresources, it is still a plausible (and testable)hypothesis that word recognition in shallow or­thographies will depend more on phonologicalrepresentations simply because such informationis available at less cost in those orthographies.Much of the evidence we have presented here isconsistent with that hypothesis.

There has been substantial progress over thepast decade in understanding the mechanismsbehind naming and recognizing printed words.Importantly, research has involved an increasingvariety of languages and writing systems, forcingtheory to be general enough to encompass thiswider scope. The vitality of this research is great

The Reading Process is Different for Different Orthographies: The Orthographic Depth Hypothesis

and even shows signs of increasing in its pace. Wehope that the main import of this article will be toclarify some issues in this area on the differentialeffects of writing systems on word perception. Ifthere is presently significant (although perhapsnot universal) agreement among researchers thatvisual-orthographic and assembled phonologicalrepresentations may both play roles in wordperception, then the next phase of researchactivity must include ways of assessing theconditions under which they are active, therelative contributions of each, and themechanisms of their action. This requires a switchfrom research designs that address qualitativequestions (e.g., "Is the lexicon accessedphonologically or not?") to designs that addressthe relative balance of phonological and visual­orthographic coding.

The best candida~e for a heuristic framework forthis research may be network modeling whichoffers a natural way of simulating therelationships between orthography and phonology,orthography and morphology, and phonology andmorphology (at one level) and between these codedrepresentations and the lexicon (at another).Likewise, differences between and withinorthographies concerning the consistency,regularity, and frequency of these relationshipscan be implemented as initial constraints on suchnetworks. Seidenberg (1992) suggests thatnetwork architectures offer ways of modeling howthe various general cognitive resources involvedare adapted to the processing of printed words:how the system assembles itself under theconstraints of language, orthography, andmemory. Implicit in this characterization is theadditional possibility of modeling the historicalevolution of a writing system. While the resourcesof memory, perceptual discrimination, and thelike, may be constant, languages and theirorthographies have not been immutable and theirhistories of change are well known in many cases.The orthographic depth hypothesis itself is astatement about a part of this larger issue of thefit between writing systems and humancapabilities. The hypothesis embodies theassumption of covariant learning (Van Orden etal., 1991). That is, the structure and operation ofthe network should reflect the contingenciesamong phonology, morphology, and orthographythat exist for printed words and, therefore, thecontingencies that will be experienced by a reader.Each orthography, shallow or deep, defines itsown pattern of contingencies. Additional progressin this area should come from requiring our ideas

about the differences among orthographies to bemade precise enough to be modeled.

REFERENCESBaluch, B., & Besner, D. (1991). Visual word recognition: Evidence

for strategic control of lexical and nonlexical routines in oralreading. fournal of Experimental Psychology: Learning, Memory,and Cognition, 17,644-652.

Besner, D., & Hilderbrandt, N. (1987). Orthographic andphonological codes in the oral reading of Japanese kana. founzalof Experimental Psychology: Leanzing, Memory, and Cognition, 13,

335-343.Besner, D., & Smith, M. C. (1992). Basic processes in reading: Is

the orthographic depth hypothesis sinking? In L. Katz & R.Frost (Eds.), Orthography, phonology, morphology, and meaning(pp. 45-66). Amsterdam: Elsevier Science Publishers.

Chomsky, N., & Halle, M. (1968). The sound pattern of English. NewYork: Harper & Row.

Cossu, G., Shankweiler, D., Liberman, 1. Y., Katz, L., & Tola, G.(1988). Awareness of phonological segments and readingability in Italian children. Applied Psycholinguistics, 9, 1-16.

DeFrancis, J. (1989). Visible Speech: The diverse oneness of writingsystems. Honolulu: University of Hawaii Press.

Feldman, L. B., & Turvey, M. T. (1983). Word recognition inSerbo-Croation is phonologically analytic. Journal ofExperimental Psychology: Human Perception and Performance, 9,288-298.

Frost, R. (in press). Prelexical and postlexical strategies in reading:Evidence from a deep and a shallow orthography. Journal ofExperimental Psychology: Learning, Memory and Cognition.

Frost, R., & Katz, L. (1989). Orthographic depth and theinteraction of visual and auditory processing in wordrecognition. Memory & Cognition, 17, 302-310.

Frost, R., Katz, L., & Bentin, S. (1987). Strategies for visual wordrecognition and orthographical depth: A multilingualcomparison. Journal of Experimental Psychology: HumanPerception and Perfomwnce, 13, 104-115.

Katz, L., & Feldman, L. B. (1981). Linguistic coding in wordrecognition: Comparisons between a deep and a shalloworthography. In A. M. Lesgold &. C. A. Perfetti (Eds.),Interactive processes in reading. Hillsdale, NJ: Erlbaum.

Katz, L., & Feldman, L. B. (1983). Relation between pronunciationand recognition of printed words in deep and shalloworthographies. fournal of Experimental Psychology: Learning,Memory, and Cognition, 9, 157-166.

Liberman, 1. Y., Liberman, A. M., Mattingly, 1. G., & Shankweiler,D. L (1980). Orthography and the beginning reader. In J. F.Kavanagh &. R. L. Vanezky (Eds.), Orthography, reading anddyslexia (pp. 137-153). Baltimore: University Park Press.

Lukatela, G., & Turvey, M. T. (1990a). Phonemic similarity effectsand prelexicalphonology. European Journal of CognitivePsychology, 18, 128-152.

Paap, K. R., Noel, R. W., & Johansen, L. S. (1992). Dual-routeModels of Print to Sound: Red Herrings and Real Horses. In L.Katz & R. Frost (Eds.), Orthography, phonology, morphology, andmeaning (pp. 293-318). Amsterdam: Elsevier Science Publishers.

Perfetti, C A., Bell, L. C, & Delaney, S. M. (1988). Automatic(prelexical) phonetic activation in silent word reading:Evidence from backward masking. Memory and Language, 27,59-70.

Perfetti, C A., Zhang, S., & Berent, 1. (1992). Reading in Englishand Chinese: Evidence for a "universal" phonologicalprinciple. In L. Katz & R. Frost (Eds.), Orthography, phonology,morphology, and meaning (pp. 277-248). Amsterdam: ElsevierScience Publishers.

160 Katz and Frost

Scheerer, E. (1986). Orthography and lexical access. In G. Augst(Ed.), New trends in graphemics and orthography (pp. 262-286).Berlin: De Gruyter.

Sebastian-Galles, N. (1991). Reading by analogy in a shalloworthography. Journal of Experimental Psychology: HumanPerception and Performance, 17(2), 471-477.

Seidenberg, M. S. (1985). The time-course of phonological codeactivation in two writing systems. Cognition, 19, 1-30.

Seidenberg, M. S. (1992). Beyond orthographic depth in reading:Equitable division of labor. In 1. Katz & R Frost (Eds.),Orthography, phonology, morphology, and meaning (pp. 85-118).Amsterdam: Elsevier Science Publishers.

Seidenberg, M., Waters, G., Barnes, M. A., & Tannenhaus, M. K.(1984). When does irregular spelling and pronunciationinfluence word recognition? Journal of Verbal Learning andVerbal Behavior, 23, 383-404.

Tabossi, P., & Laghi, 1. (1992). Semantic priming in thepronunciation of words in two writing systems: Italian andEnglish. Memory & Cognition, 20, 315-328.

Van Orden, G. C, Pennington, B. F., & Stone, G. O. (1990). Wordidentification in reading and the promise of subsymbolicpsycholinguistics. Psychological Review 97, 488-522.

Van Orden, G. C, Stone, G. 0., Garlington, K. L., Markson, 1. R,Pinnt, G, S., Simonfy, eM., & Brichetto, T. (1992)."Assembled" phonology and reading: A case study on howtheoretical perspective shapes empirical investigation. In 1.Katz & R Frost (Eds.), Orthography, phonology, morphology,and meaning (pp. 249-292). Amsterdam: Elsevier SciencePublishers.

FOOTNOTES'In L. Katz & R Frost (Eds.), Orthography, phonology, morphology,

and meaning (pp. 67-84). Amsterdam: Elsevier Science Publishers(1992).

t Also Department of Psychology, University of Connecticut,Storrs.

tDepartment of Psychology, Hebrew University.