Transcript Document

Where to next?
Pronoun interpretation as a
side effect of discourse direction
Hannah Rohde, Andy Kehler, & Jeff Elman
UC San Diego
CUNY Conference on Human Sentence Processing, March 29-31 2007
Transfer of possession (Stevenson et al. 1994)
(1) John handed a book to Bill . He thanked
recommended
______________.
John it
Transfer Verb
Source
(subject)



Goal
(to-phrase)
Ambiguous
Pronoun Prompt
50/50: Goal continuations / Source continuations
No subject preference or grammatical parallelism
Two explanations considered:


Thematic Role Preference
Event Structure Bias
1
Outline


Background: Rohde et al. 2006

Test Thematic-Role and Event-Structure biases

Alternative account: Discourse Coherence
Experiments 1 & 2: test predictions of a coherence-
based model using story continuations

Preliminary results: discourse effects in relative
clause attachment
2
Explaining salience of Goal (Rohde et al. 2006)

Thematic role preference or event structure bias?
(2) JohnSOURCE handed a book to BillGOAL. He ______ .
(3) JohnSOURCE was handing a book to BillGOAL . He ___ .
Equivalent thematic roles but different event structure
Effect of aspect
F(1,48)=50.622
p<0.0001

Goal bias ~ side effect of Event Structure
3
Effects of coherence (Rohde et al. 2006)

Establishing coherence: infer a relationship between the
meanings expressed by two sentences (P&Q below)
(Hobbs 1979, Kehler 2002)

Causal relations (Explanation, Result, Violated Expectation)
(4) Matt passed a sandwich to David. He didn’t want
David to starve.
[Explanation: Q  P]
(5) Matt passed a sandwich to David.
David He said thanks.
[Result: P  Q]
4
Coherence cont.

Similarity relations (Parallel, Elaboration)
(6) Matt passed a sandwich to David. He did so carefully.
[Elaboration: infer P from both S1 and S2]

Contiguity relations (Occasion)
David He ate it up.
(7) Matt passed a sandwich to David.
[Occasion: infer initial state of event described
in S2 to be final state of event described in S1]
5
Discourse coherence effects (Rohde et al.)


Goal bias following perfective context sentences
limited to Occasion & Result (see Arnold 2001)
Interpretation as side effect of coherence distribution
6
Shift coherence  shift interpretation

Test predictions of a coherence-driven model


More Occasion/Result  more Goal resolutions
More Explanation/Elaboration/Violated-Exp  more Source
7
Experiment 1: objects-of-transfer

Proposal: elicit different continuations with different objects

Stimuli: normal and bizarre objects
(8) John handed a book to Bill. He ___________ .
(9) John handed a bloody meat cleaver to Bill. He __ .

Predictions:


If… Abnormal objects  more Explanations
and Explanations  Source bias
More Source continuations for (9) than (8)
8
Methodology





Subjects: 69 monolingual English speakers
Task: write 50 continuations, just like Rohde et al.
Stimuli: 21 transfer-of-possession like Rohde et al.
(+ bizarre objects)
Evaluation: judges assess coherence/interpretation
Analysis:


Effect of within-subject factor of Object Type on
 Coherence (Elab/Expl/Occ/Par/Res/Viol-Exp)
 Pronoun interpretation (Source/Goal)
Mixed-effects logistic regression
 Controls for random effects of Subject and Item
9
Results
Coherence varies by object
p<0.0001
% Coherence
100%
Source
60%
40%
20%
}
Goal
0%
Normal
Coherence
Source
80%
}
Consistent prob(Source|coh)
Exp
Elab
V-E
Res
Occ
Par
Goal
Explanation
Elaboration
Violatedexpectation
Occasion
Result
Parallel
Rohde et al. Exp 1
0.75
0.99
0.82
0.99
0.87
0.81
0.20
0.16
0.45
0.17
0.05
0.71
Abnormal
10
Results

No effect of object type on pronoun interpretation
Subjects: F(1,68)= 0.052
p<0.820
Items: F(1,20)=0.111
p<0.743
11
Experiment 2: ‘What next?’ or ‘Why?’
(10) John handed a book to Bill. He ___________ .

Stimuli & Design: identical to Rohde et al. 2006

Instructions: write continuations answering either
“What happened next?” or “Why?”

Predictions:
 “What next?”  more Occasions  Goal bias
 “Why?”  more Explanations  Source bias
12
Methodology

Subjects: 42 monolingual English speakers
Task: identical to Rohde et al. 2006 (w/instructions)
Stimuli: identical to Rohde et al. 2006
Evaluation: judges assess coherence/interpretation

Analysis:




Effect of between-subject factor of Instruction Type
on coherence distribution & pronoun interpretation
13
Results
Coherence varies w/instruction
(p<0.0001)
Goal
Source
Consistent prob(Source|coh)
Coherence
Explanation
Elaboration
Violatedexpectation
Occasion
Result
Parallel
Rohde et al. Exp 2
0.75
0.99
0.81
1.00
0.87
0.81
0.20
0.16
0.45
0.28
0.11
0.46
14
Results

Effect of Instruction type on pronoun interpretation
F(1,20)=52.672
p<0.0001
15
Predicting pronoun interpretation

Predict % Source Resolutions in Exp 2 using:


Exp2 coherence breakdown
Exp1 conditional probabilities
Coherence p(Source)
Explanation
0.82
Elaboration
0.99
V-E
0.81
Occasion
0.17
Result
0.05
Parallel
0.45
(1) %SR =
 %coh * p(SR | Coh)
coh
 %Exp* p(SR | Exp)  %Elab* p(SR | Elab)  %V - E* p(SR | V - E) 
%Occ* p(SR
 | Occ)  %Res* p(SR | Res)  %Par * p(SR | Par)

16
Capturing subject variation
(1) %SRi =  %cohi * p(SR | Coh)
for subjecti in Exp 2
coh
100%



80%
Observed
%Source
60%
40%
20%
“What next”
0%
1
6
11
16
“Why”
21
26
31
36
Predicted
%Source
using (1)
41
Participant
linear regression R2=0.604
F(1,40)=61.097, p<0.0001
17
Consistency of biases across conditions
Conditional Probability Estimator
R2 value/ANOVA
Exp1: perf, normal objects
R2=0.606, F(1,40)=61.612*
Exp1: imp, normal objects
R2=0.627, F(1,40)=67.371*
Exp1: perf, abnormal objects
R2=0.561, F(1,40)=51.165*
Exp1: imp, abnormal objects
R2=0.586, F(1,40)=51.165*
Exp1: average across verbal
aspects & object types
R2=0.604, F(1,40)=61.097*
* Indicates p<0.0001
18
Summary

Shift coherence  Shift pronoun interpretation
No model relying only on surface-level cues can
account for observed variation, since stimuli were
near-identical (Exp 1) or identical (Exp 2)

Need richer models incorporating discourse-level factors
(see Wolf et al. 2004; Kertz et al. 2006)
19
What else can discourse do for you?

Relative clause attachment ambiguity
high
low
(11) Beth babysits the children of the musician who plays
____
are
at the club
downtown.
musical
prodigies
themselves.

Function of a relative clause
(12) John despises the employee who is always late.
Implicit Causality (NP2 IC) verbs
attribute cause to direct object

Proposal: try to shift RC attachment using verbs
that require Explanations and that attribute
cause to the referent occupying higher NP
20
Predictions & results
100%
100%
80%
80%
60%
low
high
40%
20%
0%
% Coherence
% Attachment
nonIC: Beth babysits the children of the musician who _____
plays
at the club downtown.  low
IC: Beth despises the children of the musician who scream
______
and yell during rehearsals.  high
Mod
Exp
Elab
Occ
Res
Par
60%
40%
20%
0%
nonIC
Verb Type
IC
F(1,51)=31.082
p<0.0001
nonIC
IC
p<0.0001
Verb Type
 Further evidence that discourse influences interpretation
21
References
Arnold, J. E. (2001) The effects of thematic roles on pronoun use and frequency of reference.
Discourse Processes, 31(2): 137-162.
Chambers, G. C. & Smyth, R. (1998) Structural parallelism and discourse coherence: A test of
Centering Theory. Journal of Memory and Language, 39: 593-608.
Crawley, R., Stevenson, R., & Kleinman, D. (1990) The use of heuristic strategies in the interpretation
of pronouns. Journal of Psycholinguistic Research, 4: 245–264.
Kameyama, M. (1996) Indefeasible semantics and defeasible pragmatics. In M. Kanazawa, C. Pinon,
and H. de Swart, editors, Quantifiers, Deduction, and Context. CSLI Stanford, pp. 111-138.
Hobbs, J. R. (1979) Coherence and coreference, Cognitive Science, 3:67-90.
Hobbs, J. R. (1990) Literature and Cognition. CSLI Lecture Notes 21. Stanford, CA.
Kehler, A. (2002) Coherence, reference, and the theory of grammar. CSLI Publications, Stanford, CA.
Kertz, L., Kehler, A., & Elman, J. (2006) Grammatical and Coherence-Based Factors in Pronoun
Interpretation. 28th Annual Conference of the Cognitive Science Society, Vancouver, July 2006.
Moens, M. & Steedman, M. (1988) Temporal ontology and temporal reference. Computational
Linguistics 14(2):15-28.
Smyth, R. H. (1994) Grammatical determinants of ambiguous pronoun resolution. Journal of
Psycholinguistic Research, 23: 197-229.
Stevenson, R., Crawley R., & Kleinman D. (1994) Thematic roles, focusing and the representation of
events. Language and Cognitive Processes, 9:519–548.
Wolf, F., Gibson, E. & Desmet, T. (2004) Coherence and pronoun resolution. Language and Cognitive
Processes, 19(6): 665-675
22
Variation by instruction and aspect
Coherence (instr x aspect
interaction p<0.0001)
Interpretation (instr x aspect
interaction p<0.0001)
…“What happened next?”
(16) John gave a book to Bill. He ___________ .
(17) John was giving a book to Bill. He ___________ .
23
Discourse coherence effects
Imperfective Context Sentences
150
Source
Goal
100
50
Vi
ol
l
lle
ra
Pa
dEx
p
at
e
Re
su
lt
0
O
cc
as
io
n
El
ab
or
at
io
n
Ex
pl
an
at
io
n
Count
200
24