Constraint-based projection Judith Tonhauser University of - - PowerPoint PPT Presentation

Constraint-based projection

Judith Tonhauser University of Stuttgart (& The Ohio State University) Based on joint work with Judith Degen, Stanford University Semantics and Linguistic Theory 30 August 17-20, 2020

Goal motivate a constraint-based projection analysis Outline

• 1. Projection
• 2. The constraint-based approach to projection
• 3. Exp 1: Lexical meaning matters, but not as expected
• 4. Exp 2: Listener belief influence projection
• 5. Conclusions and outlook

Projection Taylor: “Did Kim discover that Sandy’s work is plagiarized?” Do you (the listener) infer that Taylor is committed to the content

• f the complement (CC), that Sandy’s work is plagiarized?

“yes” CC projects “no” CC does not project increase in strength of inference / projection

Frege 1892, Strawson 1950, Kiparsky & Kiparsky 1970, Karttunen 1971, Prince 1978, Karttunen & Peters 1979, Atlas & Levinson 1981, and many more

Projection Projective content is ubiquitous in natural language: appositives, deictic and definite expressions, tense, verbs, adverbs…

(e.g., Levinson 1983, Potts 2005, Tonhauser et al 2013, Tonhauser in print)

Taylor: “Did Kim discover that Sandy’s work is plagiarized?” Empirical domain in today’s talk: The content of the complement (CC) of clause-embedding predicates know, be annoyed, announce, believe, pretend,… English: about 1,000 (White & Rawlins 2016)

Many information sources influence projection

(e.g., Stalnaker 1972, Karttunen 1974; Gazdar 1979; Heim 1982, 1983)

• 1. Common ground

Taylor: “Did Kim discover that Sandy’s work is plagiarized?” Context 1: Taylor is a professor. Her TA Kim called a student, Sandy, in for a meeting. Taylor asks another TA: Context 2: Taylor, Cam and Sandy are collaborating students. Sandy was called in for a meeting by Kim, their TA. Taylor asks Cam: Inference to CC is stronger, i.e., CC is more projective, in Context 2 than Context 1.

(e.g., Kiparsky & Kiparsky 1970; Xue & Onea 2011, Tonhauser, Beaver & Degen 2018)

Taylor: “Does Kim that Sandy’s work is plagiarized?”

• 2. Predicate

Inference to CC is stronger, i.e., CC is more projective, with discover than with think. discover think Many information sources influence projection

(e.g., Beaver 2010, Tonhauser 2016, Djärv & Bacovcin 2017)

Taylor: “Did Kim discover that Sandy’s work is plagiarized?”

• 3. Information structure

Prosody 1 (focus: Kim) Inference to CC is stronger, i.e., CC is more projective, with Prosody 1 than Prosody 2. Prosody 2 (focus: Sandy) Many information sources influence projection

• 4. Question Under Discussion / At-issueness

Many information sources influence projection

• nly

discover know stop stupid NRRC annoyed NomApp possNP

• 0.7

0.8 0.9 1.0 0.7 0.8 0.9 1.0

Mean not−at−issueness rating ('asking whether') Mean projectivity rating

Tonhauser, Beaver & Degen 2018 (Journal of Semantics)

(e.g., Simons et al 2010, 2017; Xue & Onea 2011; Cummins & Rohde 2015)

Taylor: “Did Kim discover / Does Kim know that Sandy’s work is plagiarized?” The more the CC is not-at-issue, the more projective it is.

Schlenker 2010; de Marneffe et al 2012; Mahler 2020

• 5. Information about the subject of the attitude or the speaker

(e.g., reliability, credibility, political affiliation) Many information sources influence projection Mahler 2020 manipulated the political affiliation of the speaker Cindy: “Ben doesn’t know that Obama improved the American economy.” Listeners’ inferences that Cindy is committed to the CC are stronger when Cindy is a Democrat than a Republican.

Interim summary: Listeners rely on multiple sources of information in inferring speaker commitment to the CC, i.e., in inferring projection of the CC. inferences about speaker commitment bottom-up information predicate prosody top-down information information structure common ground info about speaker QUD

Goal motivate a constraint-based projection analysis Outline

• 1. Projection
• 2. Constraint-based approach to projection
• 3. Exp 1: Lexical meaning matters, but not as expected
• 4. Exp 2: Listener beliefs influence projection
• 5. Conclusions and outlook

Constraint-based approaches to pragmatics To identify speaker meaning, listeners integrate probabilistic information from multiple sources.

(e.g., Degen & Tanenhaus 2015, 2019)

bottom-up information top-down information speaker meaning expectations about speaker linguistic context lexical meaning prosody syntax

Constraint-based approach to projection To draw inferences about speaker commitment, listeners integrate probabilistic information from multiple sources. Big question: What are the relevant information sources in the empirical domain and how are they integrated? inferences about speaker commitment bottom-up information predicate prosody top-down information information structure common ground info about speaker QUD

Contemporary projection analyses

• Lexicalist (e.g., Heim 1983, van der Sandt 1992)
• Entailment-based (e.g., Abrusán 2011, 2016; Simons, Beaver, Roberts &

Tonhauser 2017)

• Alternatives-based (e.g., Chemla 2009; Abusch 2002, 2010; Romoli 2015)
• Context-dependent triggering (Schlenker ms/2019)

Common theme: Analyses only apply to “presupposed” CCs, i.e., predicates or utterances for which the inference that the speaker is committed to the CC is “sufficiently strong”. factive know discover non-factive think be right inform announce lexicalist, entailment- and alternatives-based context-dependent triggering

inferences about speaker commitment bottom-up information predicate prosody top-down information information structure common ground info about speaker Recasting contemporary projection analyses in the constraint-based framework Is this empirically adequate? QUD

Goal motivate a constraint-based projection analysis Outline

• 1. Projection
• 2. Constraint-based approach to projection
• 3. Exp 1: Lexical meaning matters, but not as expected
• 4. Exp 2: Listener beliefs influence projection
• 5. Conclusions

Experiment 1: Lexical meaning

(Tonhauser & Degen under review; see LingBuzz) How does lexical meaning contribute to projection? Is it empirically adequate for projection analyses to disregard the CCs of particular classes of predicates (e.g., non-factive)?

Experiment 1: Materials 20 clause-embedding predicates

• Factive:
• Non-factive:

Each predicate was combined with one of 20 complement clauses, for 400 predicate/clause combinations. Lexicalist, entailment- and alternatives-based analyses predict that the CC of factive predicates is projective but they make no predictions about the CC of most non-factive predicates. know, be annoyed, discover, reveal, see (5)

• Veridical non-factive: be right, demonstrate (2)
• Optionally factive: prove, confirm, establish, announce,

confess, admit, ackowledge, hear, inform (9)

• Non-veridical non-factive: pretend, think, say, suggest (4)

(Kiparsky & Kiparsky 1970)

‘certain that’ diagnostic for projection

(e.g., Tonhauser 2016, Djärv & Bacovcin 2017, Tonhauser, Beaver & Degen 2018

Each participant rated the projectivity of the CC for each of the 20 clause-embedding predicates and 6 non-projecting controls. projection question response utterance

6 non-projecting main clause controls Sandy: “Is Zack coming to the meeting tomorrow?” Is Sandy certain that Zack is coming to the meeting tomorrow?

• 0.0

0.2 0.4 0.6 0.8 1.0

M C p r e t e n d b e _ r i g h t t h i n k s u g g e s t s a y p r

• v

e c

• n

f i r m e s t a b l i s h d e m

• n

s t r a t e a n n

• u

n c e c

• n

f e s s a d m i t r e v e a l a c k n

• w

l e d g e h e a r d i s c

• v

e r i n f

• r

m s e e k n

• w

b e _ a n n

• y

e d Predicate Mean certainty rating

Factive predicates are not categorically different from non-factive predicates.

266 self-declared native speakers of American English

• 0.0

0.2 0.4 0.6 0.8 1.0

M C p r e t e n d b e _ r i g h t t h i n k s u g g e s t s a y p r

• v

e c

• n

f i r m e s t a b l i s h d e m

• n

s t r a t e a n n

• u

n c e c

• n

f e s s a d m i t r e v e a l a c k n

• w

l e d g e h e a r d i s c

• v

e r i n f

• r

m s e e k n

• w

b e _ a n n

• y

e d Predicate Mean certainty rating

The CC of all predicates is at least mildly projective; there is no non-arbitrary cutoff for “presupposed CCs”

266 self-declared native speakers of American English

(Bayesian ME Beta regression predicting certainty ratings from predicate (treatment coding, MC as reference level); random by-participant and -item intercepts)

Discussion Predicate meaning influences projection — as long recognized! It is not empirically adequate to

• privilege particular classes of predicates, like ‘factives’, or

CCs, like ‘entailed CCs’ or

• assume a projection threshold to identify “presuppositions”.

Four pieces of converging evidence: (Tonhauser & Degen under review)

• 1. Experiment with categorical response options (‘yes’, ‘no’)

datasets that differ in materials and projection diagnostic

• 2. CommitmentBank (de Marneffe et al 2018, SuB)
• 3. VerbVeridicality (Ross & Pavlick 2019, EMNLP)
• 4. MegaVeridicality (White & Rawlins 2018, NELS)

Converging evidence: MegaVeridicality dataset

• ●
• ●
• ●
• ●
• acknowledge

admit announce be_annoyed confess confirm demonstrate discover establish hear inform know pretend prove reveal say see suggest think

−0.5 0.0 1.0

Predicate Mean projectivity rating Predicate type a a a a

non−veridical non−factive veridical non−factive

• ptionally

factive factive

517 predicates Somebody didn’t [PRED] that a particular thing happened. Did that thing happen?

(White & Rawlins 2018, NELS)

No empirical evidence that some classes of predicates are extremely privileged: projection diagnostic sentence

Interim summary

• No. The CCs of other predicates are projective, too,

sometimes even more so! How does lexical meaning contribute to projection? Is it empirically adequate for projection analyses to disregard the CCs of particular classes of predicates (e.g., non-factive)? The lexical meaning of clause-embedding predicates influences the projection of the CC; lexical meaning is a relatively stable predictor across multiple experiments and datasets. Thus: An empirically adequate projection must consider the influence on projection by the lexical meaning of all predicates.

Constraint-based approach to projection Big question: What are the relevant information sources in the empirical domain and how are they integrated? inferences about speaker commitment bottom-up information predicate prosody top-down information information structure common ground info about speaker QUD To draw inferences about speaker commitment, listeners integrate probabilistic information from multiple sources.

Goal motivate a constraint-based projection analysis Outline

• 1. Projection
• 2. Constraint-based approach to projection
• 3. Exp 1: Lexical meaning matters, but not as expected
• 4. Exp 2: Listener beliefs influence projection
• 5. Conclusions and outlook

Experiment 2: Listener belief

(Degen & Tonhauser in prep) How does listener belief influence projection?

Listener beliefs influence interpretation

• Pronoun resolution
• Ambiguity resolution
• Scalar implicatures
• Genericity
• Projection

Listeners bring their beliefs about the world, including their beliefs about the speaker’s epistemic state, (≈ world knowledge) to bear on utterance interpretation:

e.g., Winograd 1972; Altmann & Kamide 1999; Chambers et al 2002, 2004; Hagoort et al 2004; Bicknell & Rohde 2009; Degen et al 2015; Kravtchenko & Demberg 2015; Tessler & Goodman 2019; Mahler 2020

Sally: “Does Kim know that…?”

• 0.0

0.2 0.4 0.6 0.8 1.0 play alcatraz cheat soccer cupcakes bmw hat aunt

Content Mean projectivity rating Tonhauser, Beaver & Degen 2018 (Journal of Semantics) Exp 1a

Some lexical content is more projective than other Jane has a sick aunt Jack has been playing outside with the kids

Hypothesis from Tonhauser, Beaver & Degen 2018: 500 Content is more projective the higher its prior probability, i.e., the stronger listeners’ prior belief in the content.

Experiment 2: Materials 20 clause-embedding predicates (same as Exp 1)

• Factive:
• Non-factive:

know, be annoyed, discover, reveal, see (5)

• Veridical non-factive: be right, demonstrate (2)
• Optionally factive: prove, confirm, establish, announce,

confess, admit, ackowledge, hear, inform (9)

• Non-veridical non-factive: pretend, think, say, suggest (4)

(Kiparsky & Kiparsky 1970)

Each predicate was combined with one of 20 complement clauses, for 400 predicate/clause combinations, as in Exp 1. Additional manipulation: Prior probability of the CC

Experiment 2: Materials and procedure 400 polar questions Sally: “Did Kim discover that Julian dances salsa?” Manipulation of prior probability of the CC

• 1. Higher prior probability fact: Julian is Cuban
• 2. Lower prior probability fact: Julian is German

286 participants (AMT) Block 1: Prior probability of the CC, given the fact Block 2: Projection of the CC, given the fact and the predicate (block order randomized) 800 combinations of a polar question and a fact

Block 1: Prior probability of the CC, given the fact Every participant rated the prior probability of 20 CCs: 10 with higher and 10 with lower probability facts

lower probability fact complement

0.0 0.2 0.4 0.6 0.8 1.0

7 16 8 6 18 5 4 12 3 14 11 2 10 17 15 19 13 9 20 1 Content Mean prior probability rating Fact higher probability lower probability

Julian is German Julian is Cuban Julian dances salsa Mary is pregnant Mary is taking a prenatal yoga class Mary is a middle school student Prior probability of the 20 CCs is influenced by their facts

286 self-reported speakers of American English

Each participant rated the projectivity of 20 CCs given a fact and a unique predicate and 6 main clause controls

lower probability fact

Block 2: Projection of the CC, given fact and predicate projection question response fact + utterance

complement

• 0.0

0.2 0.4 0.6 0.8 1.0

M C p r e t e n d s u g g e s t b e _ r i g h t c

• n

f i r m s a y t h i n k p r

• v

e e s t a b l i s h d e m

• n

s t r a t e a n n

• u

n c e c

• n

f e s s a d m i t r e v e a l a c k n

• w

l e d g e d i s c

• v

e r s e e h e a r k n

• w

i n f

• r

m b e _ a n n

• y

e d Predicate Mean certainty rating Fact

• higher probability

lower probability main clause

Higher-probability CCs are more projective than lower-probability CCs

LMEM predicting certainty rating from prior probability rating; random effects for participant, predicate, CC; by-participant slope for prior probability (β = .27, SE = .02, t = 12.8, p < .0001)

Listener prior belief predicts projection

discover 286 participants’ projection and prior ratings

• ●
• 0.5

1 0.5 1

Prior probability rating Certainty rating Fact

• higher probability

lower probability

r = .27

Experiment 2: Summary of findings

• 1. The CC of all 20 predicates, including non-factive ones, is

at least mildly projective (as in Exp 1).

• 2. The higher a listeners’ prior belief, the stronger their

inference that the speaker is committed to the CC, i.e., the more projective is the CC (see Tonhauser, Beaver & Degen’s

2018 hypothesis).

Constraint-based approach to projection Big question: What are the relevant information sources in the empirical domain and how are they integrated? inferences about speaker commitment bottom-up information prosody top-down information listener prior belief information structure common ground info about speaker QUD predicate To draw inferences about speaker commitment, listeners integrate probabilistic information from multiple sources.

Discussion

• Lexicalist (e.g., Heim 1983, van der Sandt 1992)
• Entailment-based (e.g., Abrusán 2011, 2016; Simons, Beaver, Roberts &

Tonhauser 2017)

• Alternatives-based (e.g., Chemla 2009; Abusch 2002, 2010; Romoli 2015)
• Context-dependent triggering (Schlenker ms/2019)

Do contemporary projection analyses lead us to expect this influence of listeners’ prior beliefs on projection?

Context-dependent triggering (Schlenker ms/2019) Fact: Julian is German. Taylor: “Did Kim discover that Julian dances salsa?” Context: Julian is German. Sentence: Kim discovered that Julian dances salsa. Does it definitely follow that Julian dances salsa? the CC is contextually entailed Simplified characterization: The CC is presupposed (i.e., a commitment of the speaker) if 1. 2.

• ●
• 0.5

1 0.5 1

Prior probability rating Certainty rating Fact

• higher probability

lower probability

discover

286 participants’ ratings

(Schlenker ms/2019)

Context-dependent triggering Fact: Julian is German. Taylor: “Did Kim discover that Julian dances salsa?” the CC is contextually entailed Simplified characterization: The CC is presupposed (i.e., a commitment of the speaker) if 1. 2. the probability that a generic agent believes the CC given that they believe the content of the utterance in that context is at least as high as threshold a.

Discussion

• Lexicalist (e.g., Heim 1983, van der Sandt 1992)
• Entailment-based (e.g., Abrusán 2011, 2016; Simons, Beaver, Roberts &

Tonhauser 2017)

• Alternatives-based (e.g., Chemla 2009; Abusch 2002, 2010; Romoli 2015)
• Context-dependent triggering (Schlenker ms/2019)

Do contemporary projection analyses lead us to expect this influence of listeners’ prior beliefs on projection?

• Yes, insofar as listeners’ prior beliefs influence their

posterior beliefs.

• But how strong does the inference about speaker

commitment need to be to count as a presupposition?

Goal motivate a constraint-based projection analysis Outline

• 1. Projection
• 2. Constraint-based approach to projection
• 3. Exp 1: Lexical meaning matters, but not as expected
• 4. Exp 2: Listener beliefs influence projection
• 5. Conclusions and outlook

Constraint-based approach to projection Big question: What are the relevant information sources in the empirical domain and how are they integrated? inferences about speaker commitment bottom-up information prosody top-down information listener prior belief information structure common ground info about speaker QUD predicate Exp 1 Exp 2

(Djärv & Bacovcin 2017, Mahler et al 2019)

• 1. Principled

incorporation of information sources

• 2. Integration of

information sources

• 3. Research on

languages other than English To draw inferences about speaker commitment, listeners integrate probabilistic information from multiple sources.

SUPPLEMENTARY SLIDES

• 4. Question Under Discussion / At-issueness

Many information sources influence projection

(e.g., Simons et al 2010, 2017; Cummins & Rohde 2015)

discover 210 participants’ projection and at-issueness ratings r = .43

• ●
• ●
• ●
• 0.5

1 0.5 1

Not−at−issueness rating Projectivity rating

Tonhauser, Beaver & Degen 2018 (Journal of Semantics)

Comparing projectivity ratings

• 0.00

0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

Mean certainty rating (with prior) Mean certainty rating (no prior)

Spearman’s rank correlation n = 20 rS = .983

Converging evidence: CommitmentBank

−2 −1 1 2

• c

c u r ( 1 ) s u p p

• s

e ( 5 ) s e e m ( 2 ) e x p e c t ( 4 ) p r e t e n d ( 4 ) s u g g e s t ( 1 7 ) t h i n k ( 3 7 8 ) b e l i e v e ( 4 6 ) f e e l ( 2 9 ) c

• n

v i n c e ( 5 ) h

• p

e ( 8 ) d e m a n d ( 2 ) i n s i s t ( 3 ) m e a n ( 5 ) i m a g i n e ( 1 5 ) a s s u m e ( 5 ) f i g u r e ( 1 ) s e e ( 1 2 ) h y p

• t

h e s i z e ( 1 ) s a y ( 6 7 ) p r

• v

e ( 2 ) s i g n a l ( 2 ) f e a r ( 3 ) h e a r ( 9 ) d e c i d e ( 1 1 ) t a k e ( 1 ) f i n d ( 1 7 ) a d m i t ( 5 ) k n

• w

( 1 2 2 ) f

• r

e s e e ( 1 ) s u s p e c t ( 1 8 ) b e t ( 1 ) a n n

• u

n c e ( 3 ) g u e s s ( 1 6 ) r e a l i z e ( 2 9 ) s h

• w

( 2 ) r e m e m b e r ( 5 ) t e l l ( 1 8 ) b

• t

h e r ( 1 ) r e c

• g

n i z e ( 1 ) a c c e p t ( 2 ) u n d e r s t a n d ( 7 ) n

• t

i c e ( 3 2 ) l e a r n ( 6 ) f

• r

g e t ( 1 3 ) Predicate (with number of discourses) Mean certainty rating Predicate type

non−veridical non−factive

• ptionally

factive factive

Converging evidence: VerbVeridicality

• acknowledge

admit announce confirm demonstrate discover hear know prove reveal saw say see suggest think

−1 1 2

Predicate Mean projectivity rating Predicate type a a a a

non−veridical non−factive veridical non−factive

• ptionally

factive factive