1 Introduction phenomenon Introduction phenomenon Methods & - - PDF document

▶
1
SMART_READER_LITE
LIVE PREVIEW

1 Introduction phenomenon Introduction phenomenon Methods & - - PDF document

Oct 08, 2022 471 likes •567 views

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis Sense-contingent lexical bias & Structure of the talk QITL-2 its role for initial parsing University of Osnabrck Introduction

slide-1
SLIDE 1

1

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 1

Sense-contingent lexical bias & its role for initial parsing decisions

Daniel Wiechmann Friedrich-Schiller-Universität Jena

QITL-2

University of Osnabrück Institute of Cognitive Science

June 2006

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 2

Structure of the talk

 Introduction

 Phenomenon: Local NP/S-ambiguity  Assumptions and hypothesis

 Sense-contingent lexical guidance

 Methods and Results

 Corpus-based estimation of lexical bias

 form-based vs. sense-contingent preferences

 Comparison with experimental data (Hare et al. 2003)

 Conclusion and afterthoughts

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 3

local syntactic ambiguity

 

Since Jay always jogs a mile … doesn’t seem that long. (clause boundary)

  

The horse raced past the barn … fell. (MV/RR)

    The student knew the solution … was wrong. (NP/S) Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

t

syntactic structure is ambiguous at t

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 4

local NP/S-ambiguity

  • The student knew the solution

was incorrect. (construction DOS)

  • The student knew the

solution to the problem. (construction DONP)

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 5

local NP/S-ambiguity

Some variants are usually preferred over possible alternatives. The question is: Why?

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 6

Some accounts of sentence processing

Structural models

  • Kimball’s ‘7 Principles’
  • ‘Garden Path Theory’´

(Frazier, Fodor) Grammar-based models

  • Theta-Attachment

(Pritchett, Abney)

  • Argument Attachment

(Crocker)

  • Dependency

(Pickering) Memory-/Capacity-based (Gibson, Just und Carpenter) Experience-based

  • Lexical Guidance

(Ford)

  • Tuning

(Mitchell)

  • Referential Theory

(Altman, Steedman)

  • Completely probabilistic

(Jurafsky, Crocker und Brants)

  • Attractor-based (dynamical systems)

(Elman, Tanenhaus, … )

  • Embodied sentence comprehension

(Zwaan, Madden, Bergen and Chang; … ) Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

slide-2
SLIDE 2

2

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 7

Some accounts of sentence processing

parsing

informationally restricted

informationall y unrestricted

serial parallel serial ?

parallel

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 8

constraints on constructing an interpretation

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

  • Phrase-formation constraints
  • Contextual constraints
  • Plausibility
  • Referential contexts
  • Computational resource constraints
  • Locality (memory cost constraints)
  • Phrase-level contingent (and other) frequency constraints
  • Prosodic constraints
  • Lexically-based constraints
  • Grammatical category of perceived element
  • Subcategorization preferences
  • ...

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 9

constraints on constructing an interpretation

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

  • Phrase-formation constraints
  • Contextual constraints
  • Plausibility
  • Referential contexts
  • Computational resource constraints
  • Locality (memory cost constraints)
  • Phrase-level contingent (and other) frequency constraints
  • Prosodic constraints
  • Lexically-based constraints
  • Grammatical category of perceived element
  • Subcategorization preferences
  • ...

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 10

lexically specific information: SUBCAT preferences

transitive verbs

  • nly nominal

like,...

  • nly sentential

think,... feel, find, claim, bet, ...

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 11

lexical guidance hypothesis

Parsing preferences are guided by structural expectations resulting from sense-contingent lexically specific preferences.

Introduction phenomenon Methods & results background Conclusion & afterthoughts hypothesis

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 12

Means to estimate lexical preferences derived from corpora

experimentally derived sentence completion sentence production string frequencies, ...

estimating lexical preferences

measures of association

Introduction preliminaries Methods & results data Conclusion & afterthoughts method

slide-3
SLIDE 3

3

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 13

Estimations of lexical bias on the level of verb-form are problematic

Verbs are (conventionally) used to express a number of semantic relations and each of these have their own preferences

 

Peter VP [V admitted1 NP [ his ex-girlfriend] PP [to the club]].

  

Peter VP [V admitted2 S [NP [his ex-girlfriend] was nicer than his current

  • ne]].

    Peter VP [V admitted2 NP [his error]]

 LBVerb(form) ≠ LBVerb(sense)

(cf. Roland et al. 2000: experimental vs. corpus-based norms)

lexical preferences: form vs. sense

Introduction preliminaries Methods & results data Conclusion & afterthoughts methods

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 14

procedure

  Corpus-based estimation of LB both on the level of verb form and sense 2. LBs are compared with reading-time latencies in self-paced moving window experiment (Hare et al. 2003)

Introduction preliminaries Methods & results data Conclusion & afterthoughts methods

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 15

corpus: 17 mio word sample (BNC)

[compiled as to mirror structural properties of ICE-GB corpus]

Step 1: Extraction of all [V NP]-constructions for 20 verbs

[pronominal NPs have been ignored (case)]

Number of tokens per type: N > 3000 : random sample: 10% 300 < N < 3000 : random sample: n=300 N < 300 : all tokens ___________ tokentotal: 4960

data

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 16

Step 2 → All instances of the relevant construction were coded with respect to

  • the grammatical role played by the postverbal NP
  • the verb sense instantiated by the verb in the matrix clause using a lexical

database (WordNet 2.0) Step 3 → Lexical bias was calculated both on the level of form and sense using Distinctive Collexeme Analysis (DCA)

data

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 17

Introduction data Methods & results method Conclusion & afterthoughts results

distinctive collexeme analysis (DCA)

  • DCA is a variant of ‘Collostructional Analysis’

(Stefanowitsch and Gries 2003)

  • Assumes a construction grammar (CxG) framework
  • CxG assumes that linguistic knowledge at all levels can be characterized as constructions,

i.e. pairings of form and meaning

  • Assesses the degree of association between two constructions of arbitrary degrees of

specificity

  • DCA measures the relationship of a lexical construction towards more abstract

constructions it can occur in

  • Lexical preferences are expressed in terms of association scores
  • → outputs a gradual measure

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 18

1. Collection of data above 2. Statistical analysis of the observed distribution; → Application of Fisher‘s exact test (FET) as a measure of association

  • Fisher 1922, Pedersen 1996, Gries und Stefanowitsch 2004 for a discussion
  • Evert 2004 for a comprehensive discussion of a wide array of association measures

NP-complementation NP-complementation observed

  • bserved

(expected) (expected)

  • bserved
  • bserved

(expected) (expected) R 1 R 1 S-complementation S-complementation observed

  • bserved

(expected) (expected)

  • bserved
  • bserved

(expected) (expected) R 2 R 2 column column totals totals C1 C1 C2 C2 N find find

  • ther
  • ther verbs

verbs row row totals totals contingency contingency table table DCA DCA Introduction data Methods & results method Conclusion & afterthoughts results

distinctive collexeme analysis (DCA)

slide-4
SLIDE 4

4

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 19

Introduction data Methods & results method Conclusion & afterthoughts results

lexical bias: form-based

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 20

lexical bias: form vs. sense

comparable

  • qual. different
  • quan. different

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 21

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 22

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 23

Introduction data Methods & results method Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 24

There are substantial differences between form-based and sense-contingent preferences.

form-based vs. sense-contingent preferences

Introduction data Methods & results method Conclusion & afterthoughts results

slide-5
SLIDE 5

5

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 25

If lexical bias is in fact an indicative factor for the subconscious expectations that guide early parsing, this should be reflected in experimental findings (e.g. reading time experiments) Correlational analysis: association scores - reading time delta (Hare et al. 2003)

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

predictive power of lexical preferences

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 26

Materials:

  • 20 polysemous verbs (2 senses each → source: WordNet 2.0)
  • sense 1  preference DONP
  • sense 2  preference DOS

Procedure Subject read 2 sentences: 1 prime- & 1 target sentence

  • Subjects read prime (→ activate sense 1 or 2)
  • Subject read target (always sentential complementation)

 Reading time at each word is measured  Ambiguity effect is assessed

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

Hare et al. 2003: Self-paced reading

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 27

condition 1: p evokes a scenario compatible with REALIZE-sense of find (prefers S)

(p) The intro psychology students hated having to read the assigned text because it was boring. (t) They found (that) the book was written poorly and difficult to understand.

condition 2: p evokes a scenario compatible with LOCATE-sense of find (prefers NP)

(p) Allison and her friends had been searching for John Grisham’s new novel for a week, but yesterday they finally were successful. (t) They found (that) the book was written poorly and were annoyed that they had spent so much time trying to get it. Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

context sentence (p) - target sentence (t)

example: find

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 28

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

pos1 pos2 disambiguation region (DR)

example: find

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 29

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

Area of interest is the second word of the disambiguation region (DRPOS2)

example: find

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 30

300 320 340 360 380 400 420 440

that the book was written poorly and Reading time in ms

Lemma S, no that Lemma S - that p primes nominal comp sense - no that p primes DO NP sense, that

DRPOS2

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

example: find

slide-6
SLIDE 6

6

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 31

correlation association scoresFET – RT deltaDR Pos2

Prediction: Lexical Guidance The greater the strength of association in direction DOS, the smaller the ambiguity effect (i.e. the lower the logarithmically scaled pFET-value, the smaller the reading time delta at DRPOS2)

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 32

Results: correlationKendall is not significant (tau= 0.0613 ; z = 0.5487 ; p = 0.583)  form-based lexical bias cannot predict reading times

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

correlation: form-based association scoresFET – RT deltaDR Pos2

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 33

Results: correlationKendall is significant (tau= -0.1981; z = -1.8012; p = 0.047)  sense-contingent lexical bias can predict reading times

Introduction Hare et al. 2003 Methods & results correlation AS-RT Conclusion & afterthoughts results

correlation: sense-based association scoresFET – RT deltaDR Pos2

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 34

conclusion

The present study presented corpus-based evidence for a dominant role of sense-contingent lexical preferences (expressed by a measure of association from statistical hypothesis testing) for the resolution of local syntactic NP/S- ambiguity...

...and by implication, for a sense-contingent lexically driven comprehension system

Introduction conclusion Methods & results afterthoughts Conclusion references

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 35

afterthoughts: simulation-based language understanding model

Phonological schemas Utterance Constructions Conceptual schemas Communicative context

SIMULATION

Inferences ANALYSIS Semantic specification

form meaning

Introduction conclusion Methods & results afterthoughts Conclusion references

Embodied Construction Grammar parser

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 36

Parsing in ECG

  • Parsing = analysis process, which takes an input utterance in

context and determines the set of constructions most likely to be responsible for it

  • cf. Bryant 2003 for a computational model
  • Cued construction potentially supply top-down constraints on their

constituents

  • Constructions and their constraints should be regarded not as deterministic

but as fitting a given utterance and context to some quantifiable degree

  • Constructions and their constraints could be associated with

connection weights

  • → association strengths

Introduction conclusion Methods & results afterthoughts Conclusion references

afterthoughts: Parsing in Embodied Construction Grammar

slide-7
SLIDE 7

7

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 37

Thank you.

Introduction conclusion Methods & results afterthoughts Conclusion references

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 38

References

Fisher, R.A. 1922. "On the interpretation of χ2 from contingency tables, and the calculation of P". Journal of the Royal Statistical Society 85(1):87-94

Fodor, J.D. 1978. Parsing strategies and constraints on

  • transformations. Linguistic Inquiry, 9, 427-473.

Frazier, L. 1987. Sentence Processing: A tutorial review. In Coltheart,

  • M. (Ed.) Attention and Performance XII. Hillsdale, NJ: Erlbaum.

Gompel, R.P.G. van, and Pickering, M.J. 2001. Lexical guidance in sentence processing: A note on Adams, Clifton and Mitchell (1998). Psychonomic Bulletin and Review, 8, 851-857.

Gries, St.Th., and Stefanowitsch. A. 2004. Extending collostructional analysis: A corpus-based perspectives on 'alternations'. International Journal of Corpus Linguistics, 9, 97-129.

Hare, M. L., McRae, K., and Elman, J.L. 2003. Sense and structure: Meaning as a determinant of verb subcategorization preferences. Journal of Memory and Language, 48(2), 281-303.

Jennings, F., Randall, B., and Tyler, L.K. 1997. Graded effects of verb subcategorization preferences on parsing: Support for constraint- satisfaction models. Language and Cognitive Processes, 12(4), 485- 504.

Jurafsky, D. 1996. A Probabilistic Model of Lexical and Syntactic Access and Disambiguation. Cognitive Science, 20, 137-194.

Jurafsky, D. 2002. Probabilistic modeling in psycholinguistics: Linguistic comprehension and production. In R. Bod, J. Hay, and S. Jannedy (Eds.), Probabilistic Linguistics. Cambridge, MA: MIT Press, p.39-97.

Kawamoto, A.H. 1993. Nonlinear dynamics in the resolution of lexical ambiguity: A parallel distributed processing account. Journal of Memory and language, 32, 474-516.  MacDonald, M.C. 1999. Distributional information in language comprehension, production, and acquisition: Three puzzles and a moral. In B. MacWhinney (Ed.), The Emergence of

  • Language. Mahwah, NJ: Lawrence Erlbaum.

MacDonald, M.C., Pearlmutter, N.J., and Seidenberg, M.S.

  • 1994. Lexical nature of syntactic ambiguity resolution.

Psychological Review, 101, 676-703.

  • D. Pecher & R.A. Zwaan (Eds.), The grounding of cognition:

The role of perception and action in memory, language, and

  • thinking. Cambridge, UK: Cambridge University Press.

Roland, D., and Jurafsky, D. 2002. Verb sense and verb subcategorization probabilities. In S. Stevenson, and P. Merlo (Eds.), The Lexical Basis of Sentence Processing: Formal, Computational, and Experimental Issues. Amsterdam: John Benjamins.

Stallings, L.M., MacDonald, M.C., and O’Seaghdha, P.G.

  • 1998. Phrasal ordering constraints in sentence production:

Phrase length and verb disposition in heavy-NP shift. Journal

  • f Memory and Language, 39, 392-417.

Trueswell, J.C., Tanenhaus, M.K., and Kello, C. 1993. Verb- specific constraints in sentence processing: Separating effects

  • f lexical preference from garden-paths. Journal of

experimental Psychology: Learning, Memory and Cognition, 19, 528-553.

Fishing for Exactness (Pedersen) - Appears in the Proceedings of the South - Central SAS Users Group Conference (SCSUG-96), Oct 27-29, 1996, Austin, TX (Also available from CMP-LG E-Print Archive as #9608010 ) Introduction conclusion Methods & results afterthoughts Conclusion references FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 39 Introduction preliminaries Methods & results phenomenon Conclusion & afterthoughts assumptions

Background: mental representation

Concept of mental lexicon as static storage = good approximation

(cf. sense enumeration model; Pustejowski 1996)

→ mechanisms: access, retrieval, integration etc. But: There is no such storage in simple recurrent networks New metaphor: Words are ‘operators’ and not ‘operands’ → words are stimuli that have causal effects on mental states → “words do not have meaning, they are cues to meaning”

(cf., e.g., Tabor and Tanenhaus 2001)

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 40

Introduction preliminaries Methods & results phenomenon Conclusion & afterthoughts assumptions

Background: association measures

Why Fisher’s exact test (and not some other measure)? Pedersen 1996 → good for skewed and sparse data; better than asymptotic tests of significance (e.g. t-test, Pearson chi-squre, likelihood ratio chi square)

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 41

Evert 2004: Four major approaches to measuring association

  • Significance of association group:
  • Likelihood measures (compute the probability of observed contingency table)
  • Asymptotic statistical hypothesis tests (compute significanc or p-value)
  • Exact statistical hypothesis tests (compute a test statistic)
  • Degree of association group (estimates coefficients of association strength)
  • Point estimates (maximum likelihood estimates)
  • Conservative estimates (are based on confidence intervals obtained from hypothesis tests)
  • Information theory group

(based on concepts of entropy, cross-entropy, mutual information → quantify the non-homogeneity of the observed contigency table)

  • Number of heuristic formulae

(combine sample values that are considered good estimators)

Introduction preliminaries Methods & results phenomenon Conclusion & afterthoughts assumptions

Background: association measures

FSU Jena Anglistische Sprachwissenschaft: Sprache und Kognition 42

Background: But where do the preferences come from?

  • All mental representations are experiential, i.e., related to perception and action.
  • referent representations (RR)
  • linguistic representations (LR)
  • RR are traces laid down in memory because of perceptions of and interactions with the environment.
  • RR are multi-modal
  • RR are schematic  because of attentional limitations
  • LR are laid down, as linguistic information is being received or produced.
  • All (RR & LR) constructions are interconnected
  • LR are also connected to RR
  • How are these interconnections established?
  • The main mechanism is co-occurrence (e.g., Hebb, 1949)
  • Certain entities/events in the environment tend to co-occur.
  • Because of these spatio-temporal co-occurrences, combinations of entities and events become part
  • f the same experiential trace.
  • Because they co-occur with the entities/events, linguistic become associated with RR