a computational model of s-selection · 14th may, 2016 johns hopkins university 1department of...
TRANSCRIPT
![Page 1: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/1.jpg)
A computational model of S-selection
Aaron Steven White 1 2 Kyle Rawlins 1
Semantics and Linguistic Theory 26University of Texas, Austin14th May, 2016
Johns Hopkins University1Department of Cognitive Science2Center for Language and Speech Processing2Science of Learning Institute
1
![Page 3: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/3.jpg)
Introduction
![Page 4: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/4.jpg)
Introduction
PreliminaryTraditional distributional analyses have had tremendous suc-cess in helping us understand S(emantic)-selection
S-selectionWhat type signatures does a predicate’s denotation have?
ChallengeThese analyses can be difficult to scale to an entire lexicon
4
![Page 5: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/5.jpg)
Introduction
PreliminaryTraditional distributional analyses have had tremendous suc-cess in helping us understand S(emantic)-selection
S-selectionWhat type signatures does a predicate’s denotation have?
ChallengeThese analyses can be difficult to scale to an entire lexicon
4
![Page 6: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/6.jpg)
Introduction
PreliminaryTraditional distributional analyses have had tremendous suc-cess in helping us understand S(emantic)-selection
S-selectionWhat type signatures does a predicate’s denotation have?
ChallengeThese analyses can be difficult to scale to an entire lexicon
4
![Page 7: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/7.jpg)
Introduction
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rules, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
5
![Page 8: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/8.jpg)
Introduction
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rules, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
5
![Page 9: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/9.jpg)
Introduction
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rules, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
5
![Page 10: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/10.jpg)
Introduction
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rules, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions
3. Given syntactic distribution data, use computationaltechniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
5
![Page 11: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/11.jpg)
Introduction
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rules, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy 5
![Page 12: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/12.jpg)
Introduction
Focus
Clause-embedding predicates (∼1000 in English)
Case studyResponsive predicates: take both interrogative and declaratives
(1) John knows {that, whether} it’s raining.
ImportanceDeep literature on S-selection properties of responsivesDo they take questions, propositions, or both? (Karttunen 1977, Groenendijk
& Stokhof 1984, Heim 1994, Ginzburg 1995, Lahiri 2002, George 2011, Rawlins 2013, Spector & Egre 2015, Uegaki 2015)
6
![Page 13: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/13.jpg)
Introduction
Focus
Clause-embedding predicates (∼1000 in English)
Case studyResponsive predicates: take both interrogative and declaratives
(1) John knows {that, whether} it’s raining.
ImportanceDeep literature on S-selection properties of responsivesDo they take questions, propositions, or both? (Karttunen 1977, Groenendijk
& Stokhof 1984, Heim 1994, Ginzburg 1995, Lahiri 2002, George 2011, Rawlins 2013, Spector & Egre 2015, Uegaki 2015)
6
![Page 14: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/14.jpg)
Introduction
Focus
Clause-embedding predicates (∼1000 in English)
Case studyResponsive predicates: take both interrogative and declaratives
(1) John knows {that, whether} it’s raining.
ImportanceDeep literature on S-selection properties of responsivesDo they take questions, propositions, or both? (Karttunen 1977, Groenendijk
& Stokhof 1984, Heim 1994, Ginzburg 1995, Lahiri 2002, George 2011, Rawlins 2013, Spector & Egre 2015, Uegaki 2015)
6
![Page 15: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/15.jpg)
Outline
Introduction
Selection and clausal embedding
The MegaAttitude data set
Model fitting and results
Conclusions and future directions
Appendix
7
![Page 16: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/16.jpg)
Selection and clausal embedding
![Page 17: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/17.jpg)
Multiplicity
Many verbs are syntactically multiplicitous
(2) a. John knows {that, whether} it’s raining.b. John wants {it to rain, rain}.
Syntactic multiplicity does not imply semantic multiplicity
(3) a. John knows [what the answer is]S.b. John knows [the answer]NP.
J(3b)K = J(3a)K suggests it is possible for type(JNPK) = type(JSK)
9
![Page 18: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/18.jpg)
Multiplicity
Many verbs are syntactically multiplicitous
(2) a. John knows {that, whether} it’s raining.b. John wants {it to rain, rain}.
Syntactic multiplicity does not imply semantic multiplicity
(3) a. John knows [what the answer is]S.b. John knows [the answer]NP.
J(3b)K = J(3a)K suggests it is possible for type(JNPK) = type(JSK)
9
![Page 19: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/19.jpg)
Multiplicity
Many verbs are syntactically multiplicitous
(2) a. John knows {that, whether} it’s raining.b. John wants {it to rain, rain}.
Syntactic multiplicity does not imply semantic multiplicity
(3) a. John knows [what the answer is]S.b. John knows [the answer]NP.
J(3b)K = J(3a)K suggests it is possible for type(JNPK) = type(JSK)9
![Page 20: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/20.jpg)
Projection
What do the projection rules look like?How are a verb’s semantic type signatures projected onto itssyntactic type signatures (subcategorization frames)? (Gruber 1965,
Jackendoff 1972, Carter 1976, Grimshaw 1979, 1990, Chomsky 1981, Pesetsky 1982, 1991, Pinker 1984, 1989, Levin 1993)
[ Q]⟨⟨⟨s,t⟩,t⟩, t⟩[ Q] (Grimshaw’s notation)(Montagovian notation)
[ S] [ NP]
Semantic type
Projection
Syntactic type
10
![Page 21: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/21.jpg)
Projection
What do the projection rules look like?How are a verb’s semantic type signatures projected onto itssyntactic type signatures (subcategorization frames)? (Gruber 1965,
Jackendoff 1972, Carter 1976, Grimshaw 1979, 1990, Chomsky 1981, Pesetsky 1982, 1991, Pinker 1984, 1989, Levin 1993)
[ Q]⟨⟨⟨s,t⟩,t⟩, t⟩[ Q] (Grimshaw’s notation)(Montagovian notation)
[ S] [ NP]
Semantic type
Projection
Syntactic type
10
![Page 22: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/22.jpg)
Projection
What do the projection rules look like?How are a verb’s semantic type signatures projected onto itssyntactic type signatures (subcategorization frames)? (Gruber 1965,
Jackendoff 1972, Carter 1976, Grimshaw 1979, 1990, Chomsky 1981, Pesetsky 1982, 1991, Pinker 1984, 1989, Levin 1993)
[ Q]⟨⟨⟨s,t⟩,t⟩, t⟩[ Q] (Grimshaw’s notation)(Montagovian notation)
[ S] [ NP]
Semantic type
Projection
Syntactic type
10
![Page 23: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/23.jpg)
Projection
What do the projection rules look like?How are a verb’s semantic type signatures projected onto itssyntactic type signatures (subcategorization frames)? (Gruber 1965,
Jackendoff 1972, Carter 1976, Grimshaw 1979, 1990, Chomsky 1981, Pesetsky 1982, 1991, Pinker 1984, 1989, Levin 1993)
[ Q]⟨⟨⟨s,t⟩,t⟩, t⟩[ Q] (Grimshaw’s notation)(Montagovian notation)
[ S] [ NP]
Semantic type
Projection
Syntactic type
10
![Page 24: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/24.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
11
![Page 25: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/25.jpg)
Lexical idiosyncrasy
Lexical idiosyncrasyObserved syntactic distributions are not a perfect reflection ofsemantic type + projection rules
Example
Some Q(uestion)-selecting verbs allow concealed questions...
(4) a. Mary asked what time it was.b. Mary asked the time.
...others do not (Grimshaw 1979, Pesetsky 1982, 1991, Nathan 2006, Frana 2010, a.o.)
(5) a. Mary wondered what time it was.b. *Mary wondered the time.
12
![Page 26: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/26.jpg)
Lexical idiosyncrasy
Lexical idiosyncrasyObserved syntactic distributions are not a perfect reflection ofsemantic type + projection rules
Example
Some Q(uestion)-selecting verbs allow concealed questions...
(4) a. Mary asked what time it was.b. Mary asked the time.
...others do not (Grimshaw 1979, Pesetsky 1982, 1991, Nathan 2006, Frana 2010, a.o.)
(5) a. Mary wondered what time it was.b. *Mary wondered the time. 12
![Page 27: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/27.jpg)
Two kinds of lexical idiosyncrasy
Grimshaw (1979)
Verbs are related to semantic type signatures (S-selection) andsyntactic type signatures (C-selection)
Pesetsky (1982, 1991)
Verbs are related to semantic type signatures (S-selection); C-selection is an epiphenomenon of verbs’ abstract case
Shared core
Lexical noise (idiosyncrasy) alters verbs’ idealized syntactic dis-tributions
13
![Page 28: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/28.jpg)
Two kinds of lexical idiosyncrasy
Grimshaw (1979)
Verbs are related to semantic type signatures (S-selection) andsyntactic type signatures (C-selection)
Pesetsky (1982, 1991)
Verbs are related to semantic type signatures (S-selection); C-selection is an epiphenomenon of verbs’ abstract case
Shared core
Lexical noise (idiosyncrasy) alters verbs’ idealized syntactic dis-tributions
13
![Page 29: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/29.jpg)
Two kinds of lexical idiosyncrasy
Grimshaw (1979)
Verbs are related to semantic type signatures (S-selection) andsyntactic type signatures (C-selection)
Pesetsky (1982, 1991)
Verbs are related to semantic type signatures (S-selection); C-selection is an epiphenomenon of verbs’ abstract case
Shared core
Lexical noise (idiosyncrasy) alters verbs’ idealized syntactic dis-tributions
13
![Page 30: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/30.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
14
![Page 31: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/31.jpg)
Specifying the model
QuestionHow do we represent each object in the model?
A minimalistic answerEvery object is a matrix of boolean values
Strategy
1. Give model in terms of sets and functions2. Convert this model into a boolean matrix model
15
![Page 32: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/32.jpg)
Specifying the model
QuestionHow do we represent each object in the model?
A minimalistic answerEvery object is a matrix of boolean values
Strategy
1. Give model in terms of sets and functions2. Convert this model into a boolean matrix model
15
![Page 33: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/33.jpg)
Specifying the model
QuestionHow do we represent each object in the model?
A minimalistic answerEvery object is a matrix of boolean values
Strategy
1. Give model in terms of sets and functions
2. Convert this model into a boolean matrix model
15
![Page 34: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/34.jpg)
Specifying the model
QuestionHow do we represent each object in the model?
A minimalistic answerEvery object is a matrix of boolean values
Strategy
1. Give model in terms of sets and functions2. Convert this model into a boolean matrix model
15
![Page 35: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/35.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
16
![Page 36: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/36.jpg)
A boolean model of S-selection
know → {[ P], [ Q]}think → {[ P]} wonder → {[ Q]}
S =
[ P] [ Q] · · ·
think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·· · ·
...... . . .
17
![Page 37: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/37.jpg)
A boolean model of S-selection
know → {[ P], [ Q]}think → {[ P]} wonder → {[ Q]}
S =
[ P] [ Q] · · ·
think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·· · ·
...... . . .
17
![Page 38: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/38.jpg)
A boolean model of S-selection
know → {[ P], [ Q]}think → {[ P]} wonder → {[ Q]}
S =
[ P] [ Q] · · ·
think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·· · ·
...... . . .
17
![Page 39: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/39.jpg)
A boolean model of S-selection
know → {[ P], [ Q]}think → {[ P]} wonder → {[ Q]}
S =
[ P] [ Q] · · ·
think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·· · ·
...... . . .
17
![Page 40: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/40.jpg)
A boolean model of projection
[ P] → {[ that S], [ NP], ...} [ Q] → {[ whether S], [ NP], ...}
Π =
[ that S] [ whether S] [ NP] · · ·
[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
...... . . .
18
![Page 41: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/41.jpg)
A boolean model of projection
[ P] → {[ that S], [ NP], ...} [ Q] → {[ whether S], [ NP], ...}
Π =
[ that S] [ whether S] [ NP] · · ·
[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
...... . . .
18
![Page 42: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/42.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 43: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/43.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 44: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/44.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 45: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/45.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 46: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/46.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 47: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/47.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
19
![Page 48: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/48.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
20
![Page 49: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/49.jpg)
A boolean model of observed syntactic distribution
∀t ∈ SYNTYPE : D(wonder, t) = D̂(wonder, t) ∧ N(wonder, t)
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 1 1 · · ·know 1 1 1 · · ·wonder 1 1 0 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 0 · · ·
· · ·...
......
. . .
21
![Page 50: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/50.jpg)
A boolean model of observed syntactic distribution
∀t ∈ SYNTYPE : D(wonder, t) = D̂(wonder, t) ∧ N(wonder, t)
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 1 1 · · ·know 1 1 1 · · ·wonder 1 1 0 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 0 · · ·
· · ·...
......
. . .
21
![Page 51: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/51.jpg)
A boolean model of observed syntactic distribution
∀t ∈ SYNTYPE : D(wonder, t) = D̂(wonder, t) ∧ N(wonder, t)
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 1 1 · · ·know 1 1 1 · · ·wonder 1 1 0 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 0 · · ·
· · ·...
......
. . .
21
![Page 52: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/52.jpg)
A boolean model of observed syntactic distribution
∀t ∈ SYNTYPE : D(wonder, t) = D̂(wonder, t) ∧ N(wonder, t)
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 1 1 · · ·know 1 1 1 · · ·wonder 1 1 0 · · ·
· · ·...
......
. . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 0 · · ·
· · ·...
......
. . .
21
![Page 53: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/53.jpg)
Animating abstractions
QuestionWhat is this model useful for?
AnswerIn conjunction withmodern computational techniques, thismodelallow us to scale distributional analysis to an entire lexicon
Basic ideaDistributional analysis corresponds to reversing model arrows
22
![Page 54: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/54.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
23
![Page 55: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/55.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
23
![Page 56: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/56.jpg)
The MegaAttitude data set
![Page 57: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/57.jpg)
MegaAttitude materials
Ordinal (1-7 scale) acceptability ratings
for1000 clause-embedding verbs
×50 syntactic frames
25
![Page 58: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/58.jpg)
MegaAttitude materials
Ordinal (1-7 scale) acceptability ratingsfor
1000 clause-embedding verbs
×50 syntactic frames
25
![Page 59: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/59.jpg)
Verb selection
26
![Page 60: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/60.jpg)
MegaAttitude materials
Ordinal (1-7 scale) acceptability ratingsfor
1000 clause-embedding verbs×
50 syntactic frames
27
![Page 61: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/61.jpg)
Sentence construction
ChallengeAutomate construction of a very large set of frames in a way thatis sufficiently general to many verbs
SolutionConstruct semantically bleached frames using indefinites
(6) Examples of responsivesa. know + NP V {that, whether} S
Someone knew {that, whether} somethinghappened.
b. tell + NP V NP {that, whether} SSomeone told someone {that, whether} somethinghappened.
28
![Page 62: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/62.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 63: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/63.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 64: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/64.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 65: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/65.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 66: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/66.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 67: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/67.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 68: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/68.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 69: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/69.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 70: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/70.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 71: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/71.jpg)
Frame construction
Syntactic type
NP PP S
[ NP] [ PP] [ NP S] [ S][ NP PP] [ PP S]
ACTIVE PASSIVE COMP TENSE
that [+Q] for ∅
whether which NP
[+FIN] [-FIN]
-ed would to ∅ -ing
29
![Page 72: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/72.jpg)
Sentence construction
ChallengeAutomate construction of a very large set of frames in a way thatis sufficiently general to many verbs
SolutionConstruct semantically bleached frames using indefinites
(6) Examples of responsivesa. know + NP V {that, whether} S
Someone knew {that, whether} somethinghappened.
b. tell + NP V NP {that, whether} SSomeone told someone {that, whether} somethinghappened.
30
![Page 73: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/73.jpg)
Sentence construction
ChallengeAutomate construction of a very large set of frames in a way thatis sufficiently general to many verbs
SolutionConstruct semantically bleached frames using indefinites
(6) Examples of responsivesa. know + NP V {that, whether} S
Someone knew {that, whether} somethinghappened.
b. tell + NP V NP {that, whether} SSomeone told someone {that, whether} somethinghappened.
30
![Page 74: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/74.jpg)
Sentence construction
ChallengeAutomate construction of a very large set of frames in a way thatis sufficiently general to many verbs
SolutionConstruct semantically bleached frames using indefinites
(6) Examples of responsivesa. know + NP V {that, whether} S
Someone knew {that, whether} somethinghappened.
b. tell + NP V NP {that, whether} SSomeone told someone {that, whether} somethinghappened. 30
![Page 75: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/75.jpg)
Sentence construction
ChallengeAutomate construction of a very large set of frames in a way thatis sufficiently general to many verbs
SolutionConstruct semantically bleached frames using indefinites
(6) Examples of responsivesa. know + NP V {that, whether} S
Someone knew {that, whether} somethinghappened.
b. tell + NP V NP {that, whether} SSomeone told someone {that, whether} somethinghappened. 30
![Page 76: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/76.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences
• 1,000 lists of 50 items each• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 77: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/77.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 78: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/78.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list
• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 79: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/79.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 80: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/80.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants
• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 81: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/81.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 82: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/82.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item
• No annotator sees the same sentence more than once
31
![Page 83: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/83.jpg)
Data collection
• 1,000 verbs × 50 syntactic frames = 50,000 sentences• 1,000 lists of 50 items each
• Each verb only once per list• Each frame only once per list
• 727 unique Mechanical Turk participants• Annotators allowed to do multiple lists, but never thesame list twice
• 5 judgments per item• No annotator sees the same sentence more than once
31
![Page 84: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/84.jpg)
Task
Turktools (Erlewine & Kotek 2015)
32
![Page 85: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/85.jpg)
Validating the data
Interannotator agreementSpearman rank correlation calculated by list on a pilot 30 verbs
Pilot verb selection
Same verbs used by White (2015), White et al. (2015), selectedbased on Hacquard & Wellwood’s (2012) attitude verb classifi-cation
1. Linguist-to-linguistmedian: 0.70, 95% CI: [0.62, 0.78]
2. Linguist-to-annotatormedian: 0.55, 95% CI: [0.52, 0.58]
3. Annotator-to-annotatormedian: 0.56, 95% CI: [0.53, 0.59]
33
![Page 86: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/86.jpg)
Results
1
2
3
4
5
6
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
34
![Page 87: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/87.jpg)
Results
know
think
want
wonder
1
2
3
4
5
6
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
35
![Page 88: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/88.jpg)
Model fitting and results
![Page 89: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/89.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
37
![Page 90: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/90.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
37
![Page 91: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/91.jpg)
A model of S-selection and projection
SemanticType
SyntacticDistributionIdealizedSyntactic
Distribution
ObservedSyntactic
Distribution
AcceptabilityJudgment
Data
ProjectionRules
LexicalNoiseNoiseModel
37
![Page 92: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/92.jpg)
Fitting the model
GoalFind representations of verbs’ semantic type signatures andprojection rules that best explain the acceptability judgments
Challenges
1. Infeasible to search over 21000T × 250T possibleconfigurations (T = # of type signatures)
2. Finding the best boolean model fails to captureuncertainty inherent in judgment data
38
![Page 93: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/93.jpg)
Fitting the model
GoalFind representations of verbs’ semantic type signatures andprojection rules that best explain the acceptability judgments
Challenges
1. Infeasible to search over 21000T × 250T possibleconfigurations (T = # of type signatures)
2. Finding the best boolean model fails to captureuncertainty inherent in judgment data
38
![Page 94: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/94.jpg)
Fitting the model
SolutionSearch probability distributions over verbs’ semantic type sig-natures and projection rules
Going probabilisticWrap boolean expressions in probability measures
39
![Page 95: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/95.jpg)
Fitting the model
SolutionSearch probability distributions over verbs’ semantic type sig-natures and projection rules
Going probabilisticWrap boolean expressions in probability measures
39
![Page 96: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/96.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
40
![Page 97: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/97.jpg)
A boolean model of idealized syntactic distribution
D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) =∨
t∈{[ P],[ Q],...} S(know, t) ∧Π(t, [ that S])D̂(wonder, [ NP]) =∨
t∈{[ P],[ Q],...} S(wonder, t) ∧Π(t, [ NP])D̂(VERB, SYNTYPE) =∨
t∈SEMTYPES S(VERB, t) ∧Π(t, SYNTYPE)D̂(know, [ that S]) = 1−∏
t∈{[ P],[ Q],...} 1− S(know, t)×Π(t, [ that S])
[ P] [ Q] · · ·think 1 0 · · ·know 1 1 · · ·wonder 0 1 · · ·
· · ·...
.... . .
[ P] [ Q] · · ·think 0.94 0.03 · · ·know 0.97 0.91 · · ·wonder 0.17 0.93 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·[ P] 1 0 1 · · ·[ Q] 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·[ P] 0.99 0.12 · · ·[ Q] 0.07 0.98 · · ·
· · ·...
.... . .
[ that S] [ whether S] [ NP] · · ·think 1 0 1 · · ·know 1 1 1 · · ·wonder 0 1 1 · · ·
· · ·...
......
. . .
[ that S] [ whether S] · · ·think 0.97 0.14 · · ·know 0.95 0.99 · · ·wonder 0.12 0.99 · · ·
· · ·...
.... . .
40
![Page 98: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/98.jpg)
Wrapping with probabilities
P(S[VERB, t] ∧Π[t, SYNTYPE]) = P(S[VERB, t])P(Π[t, SYNTYPE] | S[VERB, t])= P(S[VERB, t])P(Π[t, SYNTYPE])
P
(∨tS[VERB, t] ∧Π[t, SYNTYPE]
)= P
(¬∧t¬(S[VERB, t] ∧Π[t, SYNTYPE])
)
= 1− P
(∧t¬(S[VERB, t] ∧Π[t, SYNTYPE])
)= 1−
∏tP (¬(S[VERB, t] ∧Π[t, SYNTYPE]))
= 1−∏t1− P (S[VERB, t] ∧Π[t, SYNTYPE])
= 1−∏t1− P(S[VERB, t])P(Π[t, SYNTYPE])
41
![Page 99: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/99.jpg)
Fitting the model
AlgorithmProjected gradient descent with adaptive gradient (Duchi et al. 2011)
Remaining challengeDon’t know the number of type signatures T
Standard solutionFit the model with many type signatures and compare using aninformation criterion, e.g., the Akaike Information Criterion (AIC)
42
![Page 100: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/100.jpg)
Fitting the model
AlgorithmProjected gradient descent with adaptive gradient (Duchi et al. 2011)
Remaining challengeDon’t know the number of type signatures T
Standard solutionFit the model with many type signatures and compare using aninformation criterion, e.g., the Akaike Information Criterion (AIC)
42
![Page 101: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/101.jpg)
Fitting the model
AlgorithmProjected gradient descent with adaptive gradient (Duchi et al. 2011)
Remaining challengeDon’t know the number of type signatures T
Standard solutionFit the model with many type signatures and compare using aninformation criterion, e.g., the Akaike Information Criterion (AIC)
42
![Page 102: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/102.jpg)
Akaike Information Criterion
High-level ideaMeasures the information theoretic “distance” to the truemodelfrom the best model with T types signatures (Akaike 1974)
Low-level idea (cf. Gelman et al. 2013)
For each datapoint...
1. ...remove that datapoint from the dataset2. ...fit the model to the remaining data3. ...predict the held-out datapoint
In the limit, model with lowest error on step 3 has lowest AIC
43
![Page 103: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/103.jpg)
Akaike Information Criterion
High-level ideaMeasures the information theoretic “distance” to the truemodelfrom the best model with T types signatures (Akaike 1974)
Low-level idea (cf. Gelman et al. 2013)
For each datapoint...
1. ...remove that datapoint from the dataset2. ...fit the model to the remaining data3. ...predict the held-out datapoint
In the limit, model with lowest error on step 3 has lowest AIC
43
![Page 104: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/104.jpg)
Fitting the model
Result12 is the optimal number of type signatures according to AIC
Reporting findingsRemainder of talk: best model with 12 type signatures
44
![Page 105: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/105.jpg)
Findings
Three findings
1. Cognitive predicates1.1 Two distinct type signatures [ P] and [ Q]
1.2 Coercion of [ P] to [ Q] and [ Q] to [ P]
2. Communicative predicates2.1 Two unified type signatures [ (Ent) P⊕Q] (optional
recipient) and [ Ent P⊕Q] (obligatory recipient)
45
![Page 106: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/106.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
46
![Page 107: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/107.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
46
![Page 108: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/108.jpg)
Findings
Three findings
1. Cognitive predicates1.1 Two distinct type signatures [ P] and [ Q]1.2 Coercion of [ P] to [ Q] and [ Q] to [ P]
2. Communicative predicates2.1 Two unified type signatures [ (Ent) P⊕Q] (optional
recipient) and [ Ent P⊕Q] (obligatory recipient)
47
![Page 109: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/109.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
48
![Page 110: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/110.jpg)
Findings
Three findings
1. Cognitive predicates1.1 Two distinct type signatures [ P] and [ Q]1.2 Coercion of [ P] to [ Q] and [ Q] to [ P]
2. Communicative predicates2.1 Two unified type signatures [ (Ent) P⊕Q] (optional
recipient) and [ Ent P⊕Q] (obligatory recipient)
49
![Page 111: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/111.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
50
![Page 112: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/112.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
50
![Page 113: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/113.jpg)
Hybrid types
QuestionWhat do we mean by P⊕Q?
ExampleStructures with potentially both informative and inquisitive con-tent (Groenendijk & Roelofsen 2009, a.o.)
• S-selectional behavior of responsive predicates on someaccounts (Uegaki 2012; Rawlins 2013)
• Some attitudes whose content is a hybrid Lewisian (1988)subject matter (Rawlins 2013 on think v. think about)
51
![Page 114: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/114.jpg)
Projection
NP Ved to VP[eventive]NP Ved for NP to VP
NP Ved to VP[stative]NP Ved VPing
NP Ved NP to VP[stative]NP Ved NP VP
NP Ved NP to VP[eventive]NP Ved to NP whether S[future]
NP Ved to NP whether SNP Ved to NP that S[future]
NP Ved to NP that SNP Ved to NP that S[-tense]
NP Ved NP to NPNP Ved about whether S
NP Ved about NPNP Ved
NP was Ved whether S[future]NP was Ved whether S
NP was Ved about whether SNP was Ved that S[future]
NP was Ved that SNP was Ved about NPNP Ved NP whichNP S
NP Ved NP that S[-tense]NP was Ved whichNP S
NP was Ved whether to VPNP Ved NP whether S[future]
NP Ved NP whether SNP Ved NP that S[future]
NP Ved NP that SNP was Ved whichNP to VPNP was Ved that S[-tense]
NP was Ved SNP was Ved so
NP was Ved to VP[stative]NP was Ved to VP[eventive]
NP was VedNP Ved whichNP to VPNP Ved whether to VP
NP Ved whether S[future]NP Ved whether SNP Ved whichNP SNP Ved NP VPing
NP Ved NPNP Ved that S[-tense]
NP Ved soS, I V
NP Ved SNP Ved that S[future]
NP Ved that S
52
![Page 115: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/115.jpg)
Projection
NP Ved NP
NP Ved NP VPing
NP Ved SS, I V
NP Ved that S
NP Ved that S[future]
NP was Ved to VP[stative]
NP Ved whether S
NP Ved whether S[future]NP Ved whether to VP
NP Ved whichNP S
NP Ved whichNP to VP
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00[ P]
[Q]
53
![Page 116: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/116.jpg)
Projection
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
sselectP
sselec
tQ
54
![Page 117: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/117.jpg)
acceptacknowledge
admit
affirmagree
announce
assume
attest
believe
decidedetect
expect
figure outfind out
guarantee
hope
swearwish
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
[ P]
[Q]
55
![Page 118: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/118.jpg)
accept
analyzeassumebrainstorm
clarify
contemplate
decidedetect
figure outfind out
miss
outlinequery
question
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
[ P]
[Q]
56
![Page 119: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/119.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
57
![Page 120: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/120.jpg)
Projection
NP Ved NP
NP Ved NP to NP
NP Ved S
S, I V
NP Ved that S
NP Ved that S[future]
NP Ved to NP that S
NP Ved to NP that S[future]
NP Ved to NP that S[-tense]
NP Ved to NP whether S
NP Ved to NP whether S[future]
NP was Ved to VP[stative]
NP Ved whether S
NP Ved whether S[future]
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00[ P]
[(Ent)P
⊕Q]
58
![Page 121: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/121.jpg)
S-selection
acknowledgeadvertise
announcebabble
chat claim
complain
confirmdeny
explain
faxlie
repeat
reveal
sayshare
signalwrite
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
0.9999
[ P]
[(Ent)P
⊕Q]
59
![Page 122: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/122.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
60
![Page 123: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/123.jpg)
Projection
NP Ved NP that S
NP Ved NP that S[future]
NP Ved NP that S[-tense]
NP Ved NP whether S
NP Ved NP whether S[future]
NP was Ved S
NP was Ved about NP
NP was Ved about whether S
NP was Ved
NP was Ved that S
NP was Ved that S[future]
NP was Ved that S[-tense]
NP was Ved whether S
NP was Ved whether S[future]
NP was Ved whether to VPNP was Ved whichNP to VP
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00[ Ent P]
[En
tP⊕Q]
61
![Page 124: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/124.jpg)
S-selection
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
[ Ent P]
[En
tP⊕Q]
62
![Page 125: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/125.jpg)
S-selection
advisealert
ask
bet
fax
notify
remind
tell
0.0001
0.001
0.01
0.10.250.50.750.9
0.99
0.999
[ Ent P]
[En
tP⊕Q]
63
![Page 126: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/126.jpg)
Findings
[ P] [ Q]
[ that S] [ whether S]
[ (Ent) P⊕Q][ Ent P⊕Q]
[ to NP that S] [ to NP whether S][ NP that S] [ NP whether S]
64
![Page 127: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/127.jpg)
Discussion
What we concludeProposition and question types live alongside hybrid types, andthe presence of a hybrid type correlates with communicativity
What we can exclude
Accounts that reduce (or unify) declarative and interrogative se-lection solely to S-selection of a single type + coercion
Methodological pointCoercion can have measurable effects
65
![Page 128: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/128.jpg)
Discussion
What we concludeProposition and question types live alongside hybrid types, andthe presence of a hybrid type correlates with communicativity
What we can exclude
Accounts that reduce (or unify) declarative and interrogative se-lection solely to S-selection of a single type + coercion
Methodological pointCoercion can have measurable effects
65
![Page 129: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/129.jpg)
Discussion
What we concludeProposition and question types live alongside hybrid types, andthe presence of a hybrid type correlates with communicativity
What we can exclude
Accounts that reduce (or unify) declarative and interrogative se-lection solely to S-selection of a single type + coercion
Methodological pointCoercion can have measurable effects
65
![Page 130: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/130.jpg)
Conclusions and future directions
![Page 131: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/131.jpg)
Conclusions
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rule, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
67
![Page 132: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/132.jpg)
Conclusions
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rule, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
67
![Page 133: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/133.jpg)
Conclusions
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rule, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
67
![Page 134: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/134.jpg)
Conclusions
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rule, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions
3. Given syntactic distribution data, use computationaltechniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy
67
![Page 135: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/135.jpg)
Conclusions
Goals
1. Demonstrate a combined experimental-computationalmethod for scaling distributional analysis
2. Show that this method provides insight into generalprinciples governing lexical semantic structure
Basic idea
1. Formalize S(emantic)-selection, projection rule, andlexical idiosyncrasy at Marr’s (1982) computational level
2. Collect data on ∼1000 verbs’ syntactic distributions3. Given syntactic distribution data, use computational
techniques to automate inference of projection rules andverbs’ semantic type, controlling for lexical idiosyncrasy 67
![Page 136: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/136.jpg)
Conclusions
Focus
Clause-embedding predicates (∼1000 in English)
Case studyResponsive predicates and the features that underly their se-lectional behavior.
(7) John knows {that, whether} it’s raining.
By looking at such a large data set, we can discover therelevant s-selectional features, and get an angle on theproblem at the scale of the entire lexicon.
68
![Page 137: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/137.jpg)
Conclusions
Focus
Clause-embedding predicates (∼1000 in English)
Case studyResponsive predicates and the features that underly their se-lectional behavior.
(7) John knows {that, whether} it’s raining.
By looking at such a large data set, we can discover therelevant s-selectional features, and get an angle on theproblem at the scale of the entire lexicon.
68
![Page 138: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/138.jpg)
Future directions
Further investigation of type signaturesSeven other type signatures that are also remarkably coherent
ExampleMany nonfinite-taking verbs
69
![Page 139: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/139.jpg)
Future directions
Atomic v. structured type signaturesCurrently treating type signatures as atomic but type signatureshave rich structure
ExamplePreliminary experiments with models that represent type struc-ture suggest that our glosses for the types are correct
70
![Page 140: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/140.jpg)
Future directions
Homophony v. regular polysemy v. underspecificationPatterns in how semantic type signatures distribute across verbsmay belie regular polysemy rules
ExamplePreliminary experiments with a more elaborated model suggestresponsive predicates display a regular polysemy (cf. George 2011)
71
![Page 141: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/141.jpg)
Thanks
We are grateful to audiences at Johns Hopkins University fordiscussion of this work. We would like to thank Shevaun Lewisand Drew Reisinger in particular for useful comments on thistalk.
This work was funded by NSF DDRIG-1456013 (DoctoralDissertation Research: Learning attitude verb meanings), NSFINSPIRE BCS-1344269 (Gradient symbolic computation), and theJHU Science of Learning Institute.
72
![Page 142: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/142.jpg)
Bibliography I
Akaike, Hirotugu. 1974. A new look at the statistical modelidentification. IEEE Transactions on Automatic Control 19(6).716–723.
Carter, Richard. 1976. Some linking regularities. On Linking: Papers byRichard Carter Cambridge MA: Center for Cognitive Science, MIT(Lexicon Project Working Papers No. 25) .
Chomsky, Noam. 1981. Lectures on Government and Binding: The PisaLectures. Walter de Gruyter.
Duchi, John, Elad Hazan & Yoram Singer. 2011. Adaptive subgradientmethods for online learning and stochastic optimization. TheJournal of Machine Learning Research 12. 2121–2159.
73
![Page 143: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/143.jpg)
Bibliography II
Erlewine, Michael Yoshitaka & Hadas Kotek. 2015. A streamlinedapproach to online linguistic surveys. Natural Language &Linguistic Theory 1–15. doi:10.1007/s11049-015-9305-9.http://link.springer.com/article/10.1007/s11049-015-9305-9.
Frana, Ilaria. 2010. Concealed Questions: in search of answers:University of Massachusetts at Amherst Ph.D. dissertation.
Gelman, Andrew, Jessica Hwang & Aki Vehtari. 2013. Understandingpredictive information criteria for Bayesian models. Statistics andComputing 1–20.
George, Benjamin Ross. 2011. Question embedding and the semanticsof answers: University of California Los Angeles dissertation.
Ginzburg, Jonathan. 1995. Resolving questions, II. Linguistics andPhilosophy 18(6). 567–609.
74
![Page 144: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/144.jpg)
Bibliography III
Grimshaw, Jane. 1979. Complement selection and the lexicon.Linguistic Inquiry 10(2). 279–326.
Grimshaw, Jane. 1990. Argument structure. Cambridge, MA: MIT Press.Groenendijk, Jeroen & Floris Roelofsen. 2009. Inquisitive semantics
and pragmatics. Paper presented at Stanford workshop onLanguage, Communication, and Rational Agency.
Groenendijk, Jeroen & Martin Stokhof. 1984. On the semantics ofquestions and the pragmatics of answers. Varieties of formalsemantics 3. 143–170.
Gruber, Jeffrey Steven. 1965. Studies in lexical relations:Massachusetts Institute of Technology dissertation.
Hacquard, Valentine & Alexis Wellwood. 2012. Embedding epistemicmodals in English: A corpus-based study. Semantics andPragmatics 5(4). 1–29.
75
![Page 145: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/145.jpg)
Bibliography IV
Heim, Irene. 1994. Interrogative semantics and Karttunen’s semanticsfor know. In Proceedings of IATL, vol. 1, 128–144.
Jackendoff, Ray. 1972. Semantic interpretation in generative grammar.Cambridge, MA: MIT Press.
Karttunen, Lauri. 1977. Syntax and semantics of questions. Linguisticsand philosophy 1(1). 3–44.
Lahiri, Utpal. 2002. Questions and answers in embedded contexts.Oxford University Press.
Levin, Beth. 1993. English verb classes and alternations: Apreliminary investigation. University of Chicago Press.
Lewis, David. 1988. Relevant implication. Theoria 54(3). 161–174.
76
![Page 146: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/146.jpg)
Bibliography V
Marr, David. 1982. Vision: a computational investigation into thehuman representation and processing of visual information.Henry Holt and Co. .
Nathan, Lance Edward. 2006. On the interpretation of concealedquestions: Massachusetts Institute of Technology dissertation.
Pesetsky, David. 1982. Paths and categories: MIT dissertation.Pesetsky, David. 1991. Zero syntax: vol. 2: Infinitives.Pinker, Steven. 1984. Language learnability and language
development. Harvard University Press.Pinker, Steven. 1989. Learnability and cognition: The acquisition of
argument structure. Cambridge, MA: MIT Press.Rawlins, Kyle. 2013. About ’about’. In Semantics and Linguistic
Theory, vol. 23, 336–357.
77
![Page 147: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/147.jpg)
Bibliography VI
Spector, Benjamin & Paul Egre. 2015. A uniform semantics forembedded interrogatives: An answer, not necessarily the answer.Synthese 192(6). 1729–1784.
Uegaki, Wataru. 2012. Content nouns and the semantics ofquestion-embedding predicates. In Ana Aguilar-Guevara, AnnaChernilovskaya & Rick Nouwen (eds.), Proceedings of SuB 16, .
Uegaki, Wataru. 2015. Interpreting questions under attitudes: MITdissertation.
White, Aaron Steven. 2015. Information and incrementality insyntactic bootstrapping: University of Maryland dissertation.
White, Aaron Steven, Valentine Hacquard & Jeffrey Lidz. 2015.Projecting attitudes.
78
![Page 148: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/148.jpg)
Appendix
![Page 149: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/149.jpg)
The response model
Two functions
1. Normalize for participants’ judgments so they arecomparable
2. Control for lexicosyntactic noise
80
![Page 150: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/150.jpg)
The response model
Two functions
1. Normalize for participants’ judgments so they arecomparable
2. Control for lexicosyntactic noise
80
![Page 151: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/151.jpg)
The response model
Why normalize judgments?Necessary to control for differences in participants’ use of scale
1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7
81
![Page 152: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/152.jpg)
The response model
Why normalize judgments?Necessary to control for differences in participants’ use of scale
1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7
81
![Page 153: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/153.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
82
![Page 154: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/154.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
82
![Page 155: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/155.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
82
![Page 156: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/156.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
82
![Page 157: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/157.jpg)
The response model
-10
-5
0
5
10
-10 -5 0 5 10
NP V S
NPVwh
ethe
rS
83
![Page 158: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/158.jpg)
The response model
-10
-5
0
5
10
-10 -5 0 5 10
NP V S
NPVwh
ethe
rS
84
![Page 159: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/159.jpg)
The response model
1
2
345
6
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
85
![Page 160: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/160.jpg)
The response model
know
think
want
wonder
1
2
345
6
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
86
![Page 161: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/161.jpg)
The response model
know
think
want
wonder
1
2
345
6
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
87
![Page 162: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/162.jpg)
The response model
know
think
want
wonder
1
23456
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
88
![Page 163: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/163.jpg)
The response model
know
think
want
wonder
1
2
3456
7
1 2 3 4 5 6 7NP V S
NPVwh
ethe
rS
89
![Page 164: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/164.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
90
![Page 165: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/165.jpg)
The response model
0 1
1
−3
2
−2
3
−1
4
0
5
1
6
2
7
3−3 −2 −1 0 1 2 3
R
90
![Page 166: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/166.jpg)
The response model
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00NP V S
NPVwh
ethe
rS
91
![Page 167: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/167.jpg)
Fitting the model
SubgoalFind the optimal number T of type signatures
Goodness of T ↔ model’s ability to... ...fit observed judgments...predict unobserved judgments
• T too small →
bad fitbad prediction
• T too large →
good fitbad prediction
Measure
Akaike Information Criterion (AIC) trades off fit to observeddata and prediction of unobserved data
92
![Page 168: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/168.jpg)
Fitting the model
SubgoalFind the optimal number T of type signatures
Goodness of T ↔ model’s ability to... ...fit observed judgments...predict unobserved judgments
• T too small →
bad fitbad prediction
• T too large →
good fitbad prediction
Measure
Akaike Information Criterion (AIC) trades off fit to observeddata and prediction of unobserved data
92
![Page 169: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/169.jpg)
Fitting the model
SubgoalFind the optimal number T of type signatures
Goodness of T ↔ model’s ability to... ...fit observed judgments...predict unobserved judgments
• T too small →
bad fitbad prediction
• T too large →
good fitbad prediction
Measure
Akaike Information Criterion (AIC) trades off fit to observeddata and prediction of unobserved data
92
![Page 170: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/170.jpg)
Fitting the model
SubgoalFind the optimal number T of type signatures
Goodness of T ↔ model’s ability to... ...fit observed judgments...predict unobserved judgments
• T too small →
bad fitbad prediction
• T too large →
good fitbad prediction
Measure
Akaike Information Criterion (AIC) trades off fit to observeddata and prediction of unobserved data
92
![Page 171: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/171.jpg)
Fitting the model
Number of type signatures
1 2 3 4 5 6 7
Low extremeAll verbs’ syntactic
distributions explainedby single rule
High extreme# types ≥ # frames
every syntactic framehas separate rule
93
![Page 172: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/172.jpg)
Fitting the model
Number of type signatures
1 2 3 4 5 6 7
Low extremeAll verbs’ syntactic
distributions explainedby single rule
High extreme# types ≥ # frames
every syntactic framehas separate rule
93
![Page 173: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/173.jpg)
Fitting the model
Number of type signatures
1 2 3 4 5 6 7
Low extremeAll verbs’ syntactic
distributions explainedby single rule
High extreme# types ≥ # frames
every syntactic framehas separate rule
93
![Page 174: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/174.jpg)
Fitting the model
SubgoalFind the optimal number T of type signatures
Goodness of T ↔ model’s ability to... ...fit observed judgments...predict unobserved judgments
• T too small →
bad fitbad prediction
• T too large →
good fitbad prediction
Measure
Akaike Information Criterion (AIC) trades off fit to observeddata and prediction of unobserved data 94
![Page 175: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/175.jpg)
Model comparison
620000
640000
660000
680000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 151 2 3 4 5 6 7 8 9 10 11 12 13 14 151 2 3 4 5 6 7 8 9 10 11 12 13 14 15Number of semantic type signatures
Akaike
Inform
ationCrite
rion
95
![Page 176: A computational model of S-selection · 14th May, 2016 Johns Hopkins University 1Department of Cognitive Science 2Center for Language and Speech Processing 2Science of Learning Institute](https://reader035.vdocuments.us/reader035/viewer/2022070912/5fb431e60bc498456f6f6f38/html5/thumbnails/176.jpg)
Model comparison
621500
622000
622500
623000
623500
8 9 10 11 12 13 14 158 9 10 11 12 13 14 158 9 10 11 12 13 14 15Number of semantic type signatures
Akaike
Inform
ationCrite
rion
96