2013-07-13-21-04-57 - University of Alberta

Rethinking Coercion as a Cogni0ve Phenomenon:

Processing, Frequency, and Seman0c Compa0bility

Suzanne Kemmer

Rice University

Soyeon Yoon Seoul Na0onal University/

Rice University

ICLC-‐12, Edmonton, June 2013

Coercion •  Resolu0on of seman0c incompa0bility between a construc0on and a lexical item occurring in it (Michaelis 2005) •  Give me some pillow! •  I sneezed the napkin off the table. •  I’m liking it.

•  A contextual reinterpreta0on that occurs when seman0c specifica0ons clash (Pustejovsky) –

•  An adjustment of specifica0ons “repairs the mismatch” (de Swart)

•  “Special meaning effects” (de Swart)

2

Issues with coercion as typically conceived

•  Changes in binary feature specifica0ons (Michaelis 2005; de Swart 2000) are inconsistent with a frame-‐based, gradient seman0cs

•  It is not clear how the coercion ‘mechanism’ relates to online processing or other aspects of language use:

-‐-‐ Theories (Construc0on Grammar; formal grammars) are generally either silent or inexplicit about how processing relates (but see Traxler et al. 2002, Piñango et al. 1999, Piñango et al. 2006). -‐-‐or, they explicitly divorce language structure from processing (e.g. Sign-‐based Construc0on Grammar).

3

Usage-‐based Model:

Correla0on of Four Dimensions General predic,on (Kemmer 2008 following from Langacker 1987, 1990, 2000 inter alia.)

There should be some correla0on between:

•  Seman0c compa,bility of a host construc0on with lexical item

•  Frequency of use (distribu0onal/behavioral correlate of cogni0ve entrenchment)

•  Processing 0me

•  Acceptability judgments

How much? How does it play out? An empirical ques0on.

But an empirically-‐demonstrated overall correla0on will support the usage-‐based model as described by Langacker: dynamic, gradient, integrated

4

Why? Seman0c compa0bility;

Frequency;

Processing;

Acceptability

-‐-‐Why and how should these relate?

5

Interrela0on of 4 dimensions

In a dynamic usage-‐based language system:

•  Greatest seman,c compa,bility: Maximal conven0onality, minimal seman0c extension; schemas and exemplars fit together in their specifica0ons, no clash

•  Frequency: Construc0ons are schema0za0ons over many exemplars; they derive exactly from repe00on of exemplars that (therefore) best fit them. Highly frequent exemplars are analogical anractors for novel exemplars of less frequency and less compa0bility – including coercions

6

Interrela0on, cont. •  Processing: Generally, cogni0ve mismatches should be

harder to process. Specifically, in a cogni0ve compe00on model, ambivalence/difficulty of categoriza0on should take more 0me. Also-‐-‐a well-‐known property of cogni0on: the more frequent the experience, the easier (and therefore faster) it is to process.

•  Acceptability judgments: Speakers like most what they have most heard before: schemas with their usual exemplars in prototypical rela0ons. Minimal mismatches. (Boas 2011 shows rela0on of coercion, seman0c compa0bility, and variable acceptability)

All subject to incremental change over 0me; and construc0on and its conven0onal and produc0ve uses developing as the individual’s language system matures.

7

Inves0gated for one construc0on in Yoon (2012)

English Ditransi0ve Construc0on [V NP1 NP2]

•  Sally gave John the book. •  Construc0onal meaning: transfer of possession

from an Agent to a Recipient

The criteria of seman0c compa0bility •  the number of par0cipants in the prototypical

event scene of the verb •  the possibility that the Pa0ent is transferrable

as a result of the ac0on prototypically designated by the verb (e.g. kill)

8

Seman0c compa0bility of verbs in Ditransi0ve Construc0on (DC)

9

Seman,c Compa,bility (1 most, 5 least)

Seman,c Type of Verb

Eg.

SemCom1 SemCom2 SemCom3 SemCom4

SemCom5

Inherent transfer Poten0al transfer Prevented transfer Impossible transfer

Events internal to the Agent

give, send cook, find refuse, deny cut, break

think, stay

Categories 1-‐3 based on Pinker (1989) and Goldberg (1995)

Verbs said to not occur with the DC (Goldberg 1995: 128)

•  Verbs of fulfilling (X gives something to Y that Y deserves, needs, or is worthy of)

present, donate, provide •  Verbs of con0nuous causa0on of accompanied mo0on in some

manner pull, carry, push

•  Verbs of manner of speaking shout, murmur, whisper

•  Verbs of proposi0on and proposi0onal attude say, claim

•  Verbs of choosing choose, pick

10

More verbs to be examined

Method •  Collexeme Analysis (Stefanowitsch and Gries 2003)

Corpus •  BNC, spoken subcorpus -‐ ca. 1,450,000 words

•  # of DC exemplars: 1,374 •  # of verbs used in the DC: 49

1. Frequency of verbs in DC

11

Result •  Verbs more compa0ble with the DC tend to be more frequently associated with the DC (higher collostruc0onal rank).

•  Verbs less compa0ble tend to occur less frequently in the DC or do not occur at all.

Table shows the rela0on:

12

13

Compa,bility Average Collostruc,on rank

# of verbs found

Examples

SemCom 1 17 25 give, send, tell SemCom2 SemCom3

33 29

20 2

buy, make, find refuse, deny

SemCom4 SemCom5

-‐ 34

0 2

-‐ run, wish

Compa0bility and Collostruc0onal Rank

Verbs and collostruc0onal rank

Next chart shows rela0on of specific seman0c classes of verbs (and their individual verbs) and collostruc0onal rank

14

15

2. Processing effort and acceptability judgments (DC)

•  Experiment Design •  S0muli

•  35 verbs selected from seman0c compa0bility categories and result of corpus analysis

•  35 sentences where each verb was used as a main verb in the DC •  (1) Eddie told Kim the news last month. (tell from

SemCom1) •  (2) Billy found Jane the ring six days ago. (find from

SemCom2)

16

Design, cont. Task

•  27 par0cipants read the sentences in a self-‐paced reading task.

•  The 0me taken to read the second NP (underlined in (1) and (2)) was recorded.

Acceptability judgments •  Aver reading each sentence, the par0cipants

judged its naturalness on 7-‐point-‐scale.

17

Verb seman,c class (from 1, most compa0ble, to 5, least compa0ble)

Verb subclass Selected verbs

1.  Verbs of inherent transfer

Inherently signifying giving give Communica0on tell Instrument of communica0on fax Future having owe, promise,

leave, allow

Sending send Deic0c bring

2.  Verbs of possible transfer Ballis0c mo0on throw, drop Crea0on create, cook Obtaining find, buy, rent

(hire in BE)

3. Verbs of refused transfer Refusal refuse, deny

4. Verbs of impossible/impaired transfer

Damaging break, cut

5. Verbs of events internal to the Agent

Emo0on/cogni0on/desire think, want, wish

intransi0ve stay, sneeze Verbs occurring only in the corpus (placed in 2nd most compa0ble)

Loca0on put, set General causa0on cause

Verbs that were expected not to occur in the DC (the least compa0ble)

present, donate, provide, push, whisper, say, choose

18

Result

Significant trend •  If seman0cally less compa0ble, processed slower

•  Judged as less acceptable

19

Seman0c compa0bility with processing 0me; with acceptability judgments

20

y = 1.067x + 0.675 R² = 0.8457

0

1

2

3

4

5

6

7

SemCom1 SemCom2 SemCom3 SemCom4 SemCom5

y = 19.336x + 780.1 R² = 0.65817 740

760

780

800

820

840

860

880

900


Figure 1. Average processing 0me of each seman0c compa0bility category (Linear Trend: t(26) = 30.29, p < .001)

Average processing 0me of each seman0c compa0bility category (Linear Trend: t(26) = 3.02, p < .01)

Average naturalness score of each seman0c compa0bility category (Linear Trend: t(26) = 30.29, p < .001)

Excluding outliers (misclassified?) put, set, and cause

A more linear trend

21

y = 1.16x + 0.21 R² = 0.91509 0

1

2

3

4

5

6

7


y = 21.968x + 766.94 R² = 0.65421 740

760

780

800

820

840

860

880

900


Average naturalness score of each seman0c compa0bility category (excluding put, set, and cause)

Average processing 0me of each seman0c compa0bility category (excluding put, set, and cause)

give, fax, allow, bring

Correla0on of Four Dimensions (DC)

All four aspects were significantly correlated with each other.

22

(*p < .01, **p < .001) SemCom ColloRank NatScore ProcessingT

SemCom ColloRank .42** NatScore .54** .41** ProcessingT .09* .12 ** .13**

Gradient Nature of Coercion If seman0c compa0bility is gradable, will coercion be the same for all different degrees of seman0c compa0bility? à No, coercion is also gradable:

Kelly sent Ryan the card. Billy found Jane the ring. Larry refused Kim the lunch.

Jean broke David the bread. Ricky stayed Sue the space.

Seman0c compa0bility correlates with: •  how oven the resolved co-‐occurrences are used •  how difficult the resolu0on is to process •  how natural the speakers feel the co-‐occurrences are

23

Coercion, Usage, Processing Coercion is closely related with usage,

specifically, processing. •  In comprehension, speaker requires different

amounts of actual processing effort, depending on the amount of seman0c incompa0bility.

•  Coercion can be thought of not as a theore0cal mechanism in the “grammar”, separate from processing (and usage in general) but as part of an actual psychological process during language use: resolving seman0c incompa0bility online in usage events

24

Direc0onality of coercion Seman0cs of the target lexical item and the construc0on

Some0mes, the meaning of a lexical item overrides the construc0onal meaning. à challenges Override Principle (Michaelis

2005) claiming construc0on always coerces lexical item

Larry refused Kim the lunch. Kevin caused Liz the fire.

25

Linguis0c and extralinguis0c context affect coercion

•  Linguis0c context Some0mes, coercion is easier with par0cular linguis0c contexts – par0cularly V NP colloca0ons (via ac0va0on of general or specific frames)

•  Larry owed Jane the watch. vs. Larry owed Jane $10. •  Kevin caused Liz the fire. vs. Kevin caused Liz trouble.

•  Extra-‐linguis0c context •  Speakers try to resolve the incompa0bility by exploi0ng extra-‐linguis0c context. •  David broke Jean the bread. •  She squinted into the room. (Kemmer 2008)

26

What is coercion, really? What people call “coercion” is a subcase of dynamic seman0c integra0on of construc0onal schemas, lexical schemas, their associated conven0onal frames, and contextual elements

-‐-‐where the incompa0bility is no0ceable (there’s some viola0on of a generaliza0on that works in prototypical cases)

-‐-‐during syntac0c/seman0c composi0on of open-‐slot construc0ons with lexical items

(purely seman0c composi0on/resolu0on as in colorless green ideas, has not been of much interest in modern Linguis0cs).

27

Why inves0gate coercions in par0cular?

Relevance: •  Coercions are rela0vely novel mo0vated usages that partly

conform to an exis0ng construc0onal schema. Thus they are relevant to syntax.

•  We can closely observe the synchronic grammar and its processing at an interes0ng point: where conven0onaliza0on of a construc0on is intermediate, and it works with some classes of lexical items but not others.

•  Diachrony: Emergence and change of construc0ons can be studied. As exemplars of a par0cular type become more entrenched, the construc0on changes its specifica0ons (cf. Israel 1996).

•  Acquisi,on: Can inves0gate learning of a construc0on and expansion to new lexical items/classes of lexical items.

•  Varia,on: Can observe varia0on among and within speakers.

28

Conclusions Coercion is a concept widely invoked to allow for/explain seman0c mismatches and to argue for existence of construc0ons.

We conclude:

1. Since theore0cal ideas rest on it, its nature should be more closely inves0gated.

2. Coercion is a gradient cogni0ve process reflected in variable processing 0me. It is not a unitary or “all or nothing” device or process.

29

Conclusions 3. Coercion is the set of syntac0cally relevant subtypes of the dynamic process of seman0c integra0on of:

•  conven0onal linguis0c specifica0ons

•  frame-‐based knowledge not specific to language

•  contextual elements This general process occurs in language usage in general, not just in syntac0c construc0ons no0ced by linguists.

30

Coercion Phenomena given the name “coercion” are disparate, e.g.:

•  “NP-‐coercion” – specifically mass construed as unit (Give me a beer) – is highly conven0onalized in English

•  A schema with seman0cs ‘conven0onally uni0zed drink’ has entrenched exemplars with par0cular lexical items associated with par0cular frames; is compa0ble with count noun construc0ons (singular indef. ar0cle, pluraliza0on); and can be licensed for non-‐conven0onalized nouns (new drink names, masses not usually uni0zed etc.), in contexts ac0va0ng the frames associated with the schema

•  Under usage-‐based model, entrenched cases like a beer do not actually involve coercion. They are expected to be processed more quickly, show higher frequency, and have greater acceptability than found in cases of real incompa0bility (genuine coercions)

31

Conclusions 4. Coercion can be inves0gated for specific construc0ons, but we need to take into account the degree of entrenchment of relevant construc0onal schemas, specific and general.

Doing so will provide:

•  A more general and accurate descrip0on of coercion phenomena

•  Stronger theore0cal grounding

•  Natural rela0on to acquisi0on, synchronic varia0on, and diachrony

32

References

Boas, Hans. 2011. Coercion and leaking argument structures in Construc0on Grammar. LinguisGcs 49-‐6.

De Swart, Henriëne. 2000. Tense, aspect and coercion in a cross-‐linguis0c perspec0ve. Proceedings of the Berkeley Formal Grammar Conference. Stanford: CSLI Publica0ons.

Goldberg, Adele. 1995. ConstrucGons: A ConstrucGon Grammar approach to argument structure. Chicago: University of Chicago Press.

Israel, Michael. 1996. The Way-‐Construc0ons Grow. In Adele Goldberg, ed., Conceptual Structure, Discourse, and Language. Stanford: CSLI.

Kemmer, Suzanne. 2008. New dimensions of dimensions: Frequency, produc0vity, domains and coercion. Presented at Cogni0ve Linguis0cs: Between Universality and Varia0on. Dubrovnik, Croa0a.

Langacker, Ronald. 1987. FoundaGons of CogniGve Grammar. Vol. I. Stanford: Stanford University Press.

Langacker, Ronald. 1990. A usage-‐based model. Chapter 10 of Concept, Image and Symbol: The CogniGve Basis of Grammar , 261-‐288. Berlin: Mouton de Gruyter.

Langacker, Ronald. 2000. A dynamic usage-‐based model. In Michael Barlow and Suzanne Kemmer, eds. Usage-‐based Models of language, 1-‐63. Stanford: CSLI.

Michaelis, Laura A. 2005. En0ty and event coercion in a symbolic theory of syntax. In Jan-‐Ola Østman and Miriam Fried, Eds. ConstrucGon Grammar(s): CogniGve Grounding and TheoreGcal Extensions. (Construc0onal Approaches to Language 3.) Amsterdam: Benjamins.

Piñango, M.M., A.E. Zurif, and Ray Jackendoff, 1999. Real-‐0me processing implica0ons of aspectual coercion at the syntax-‐seman0cs interface. Journal of PsycholinguisGc Research 28, 395-‐414.

Piñango, M.M.,A. Winnick, R. Ullah, and E. Zurif. 2006. Time-‐course of seman0c composi0on: The case of aspectual coercion. Journal of PsycholinguisGc Research 35, 233-‐244.

Pinker, Steven. 1989. Learnability and CogniGon: The acquisiGon of argument structure. Cambridge, MA: The MIT Press.

Pustejovsky, J. 1995. Linguis0c Constraints on Type Coercion. In P. Saint-‐Dizier and E. Viegas (eds.), ComputaGonal Lexical SemanGcs, 71-‐97. Cambridge; New York; Melbourne: Cambridge University Press.

Stefanowitsch, Anatol, and Stefan Gries. 2003. Collostruc0ons: Inves0ga0ng the interac0on of words and construc0ons. InternaGonal Journal of Corpus LinguisGcs 8, 209-‐243.

Traxler, M. J., M. J. Pickering, and B. McElree. 2002. Coercion in sentence processing: Evidence from eye-‐movements and self-‐paced reading. Journal of Memory and Language 47, 530-‐547.

Yoon, Soyeon. 2012. ConstrucGons, SemanGc CompaGbility and Coercion: An empirical usage-‐based approach. Doctoral disserta0on, Dept. of Linguis0cs, Rice University.

33

Addi0onal Slides Details on regressions.

34

Details: Regressions correla0ng the factors

Regression 1 y NatScore = .79xSemCom + .03xColloRank + .001x ProcessingT – 2.87

(p <.001) (p < .001) (p < .05) •  Seman0cally less compa0ble construc0on and verb •  Less frequently used together •  Processed with more effort •  Judged less acceptable •  More coercion

Regression 2 yProcessingT = 7.79xSemCom + 1.47xColloRank – 67.72

(p = .22) (p < .01)

35

Mul0ple Regression

36

Unstandardized coefficient

Standardized coefficient p

Step 1 Constant -34.70 r2 = .01 ( p < .01) SemCom 15.37 .09 p < .01

Step 2 Constant -67.72 r2 = .02 ( p < .001) SemCom 7.79 .04 p = .218

ColloRank 1.47 .10 p < .01

ProcessingT

ColloRank

SemCom

Documents

2013-07-13-21-04-57 - University of Alberta