Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Focus Representation and Backward Underspecific...

hfunakura
November 23, 2023
32

Focus Representation and Backward Underspecification in Dependent Type Semantics

hfunakura

November 23, 2023
Tweet

Transcript

  1. 19, Nov, 2023 LENLS 20 Focus Representation and Backward Underspecification

    in Dependent Type Semantics 1 Hayate Funakura Kyoto University
  2. Agenda • Background and motivation • Empirical scope • Proposal

    • Focus context • Backward underspecified term • How the related examples are derived • Unsolved issues 2
  3. Background and Motivation • Main topics: FOCUS and QUESTION-ANSWER CONGRUENCE

    • central in question semantics • (important for logic-guided chatbots) 3
  4. Background and Motivation • Main topics: FOCUS and QUESTION-ANSWER CONGRUENCE

    • central in question semantics • (important for logic-guided chatbots) • Proposals: A method to represent focus information and an analysis of English wh-questions in Dependent Type Semantics 4
  5. Background and Motivation • Dependent Type Semantics [Bekki 2014, etc.]

    • NL semantics based on Martin-Löf type theory • Key item: underspecified terms (UTs) 5
  6. Background and Motivation • Dependent Type Semantics [Bekki 2014, etc.]

    • NL semantics based on Martin-Löf type theory • Key item: underspecified terms (UTs) 6 John knows that Susan danced. ↦ know(j, dance(s), @) in Tanaka+ ( 2 0 1 7 ) Presup. Susan danced. Previous context requires the previous context to entail @ dance(s)
  7. Background and Motivation • In DTS, questions are addressed by

    Watanabe+ 2019, 
 but focus is not covered. • (mention-some) wh-questions are represented as weak- sigma types Who danced? • Informally, a response which entails Someone danced is predicted to be a felicitous answer, regardless of focus ↦ (x : entity) ⊕ dance(x) 7
  8. Empirical scope • Single-wh • Who does John like? •

    John likes [Mary]F. • Multiple-wh • Who likes whom? • [John]F likes [Mary]F. 8 • Phrasal focus • What did John do? • John [met Mary]F. • Answers with anaphoric expressions • John likes [him]F.
  9. Empirical scope • Out of scopes • Answers with quantificational

    expressions / coordination • John likes [everyone]F. • John likes [Mary and Taro]F. • Adjunct wh-questions (when, where, how, etc.) • Polar/alternative questions • Other focus-related phenomena • Comparison with model-theoretical approaches 9
  10. Background and Motivation • Alternative Semantics [Rooth 1992, etc.] is

    a standard approach to focus in formal semantics 10
  11. Background and Motivation • Alternative Semantics [Rooth 1992, etc.] is

    a standard approach to focus in formal semantics • Two options: • Translate Alternative Semantics into DTS language • Explore a unique approach to DTS 11
  12. Background and Motivation • Alternative Semantics [Rooth 1992, etc.] is

    a standard approach to focus in formal semantics • Two options: • Translate Alternative Semantics into DTS language • Explore a unique approach to DTS • The latter seems more challenging 12
  13. Proposal • Focus contexts (FCs) Sequences of type assignments which

    represent 
 focus information • Backward underspecified terms (backward UTs) Gaps in wh-questions which are expected to be filled 
 by an answer 13
  14. Proposal • Focus contexts (FCs) Sequences of type assignments which

    represent 
 focus information • Backward underspecified terms (backward UTs) Gaps in wh-questions which are expected to be filled 
 by an answer 14
  15. Focus contexts • FCs are sequences of type assignments which

    correspond to focused elements • [John]F likes Mary. SR: FC: • John [likes Mary]F. SR: FC: • FCs for each sentence are obtained in parallel with semantic composition (i.e., CCG derivation) like(j, m) j : e like(j, m) like : e → e → 𝚝 𝚢 𝚙 𝚎 , m : e 15
  16. Focus contexts • To derive FCs, I add lexical FC

    to CCG lexical entries • Each lexicon is assigned different FCs depending on whether it appears in a focused constituent or not 16
  17. Focus contexts A CCG derivation for “[John]F likes Mary” 18

    Assigned at the lexical level Merged in line with the derivation
  18. • A case of phrasal focus • Since “met” and

    “Mary” are focused, we obtain a sequence of type assignments for these words Focus contexts 19
  19. Proposal • Focus contexts (FCs) Sequences of type assignments which

    represent 
 focus information • Backward underspecified terms (backward UTs) Gaps in wh-questions which are expected to be filled 
 by an answer 20
  20. Backward underspecified terms • idea: • The meaning of wh-questions

    has an unsolved part • Backward UTs represent that unsolved part • Not an alternative to ordinary underspecified terms, 
 but an additional item 21 Who does John like? } unsolved
  21. Backward underspecified terms 22 Sentence Who does John like? like(j,

    @B) SR FC @B : e • is a backward underspecified term • For a SR which contains a backward UT, 
 the type-checking is postponed until the focus context of a response fills the gap. @B
  22. SR 23 like(j, @B) Response “John likes [Mary]F” SR FC

    like(j, m) Question “Who does John like?” m : e FC @B : e Backward underspecified terms
  23. 24 Response “John likes [Mary]F” 1. search a proof term

    of type e Question “Who does John like?” SR like(j, @B) SR FC like(j, m) m : e FC @B : e Backward underspecified terms
  24. 25 Response “John likes [Mary]F” 1. search a proof term

    of type e Question “Who does John like?” SR like(j, @B) SR FC like(j, m) m : e FC @B : e 2. -elimination @B like(j, m) Backward underspecified terms
  25. 26 Response “John likes [Mary]F” 1. search a proof term

    of type e Question “Who does John like?” SR like(j, @B) SR FC like(j, m) m : e FC @B : e 2. -elimination @B like(j, m) 3. entailment? Backward underspecified terms
  26. How the related examples are derived: single wh-questions 28 Question

    Who does John like? Response [John]F likes [Mary]F. An example of overfocusing
  27. How the related examples are derived: single wh-questions 29 Question

    Who does John like? like(j, @B) SR Response [John]F likes [Mary]F. like(j, m) SR FC An example of overfocusing FC @B : e m : e j : e
  28. How the related examples are derived: single wh-questions 30 Question

    Who does John like? like(j, @B) SR Response [John]F likes [Mary]F. like(j, m) SR FC 1. proof search OK 2. -elimination @B like(j, m) An example of overfocusing FC @B : e m : e j : e
  29. How the related examples are derived: single wh-questions 31 Question

    Who does John like? like(j, @B) SR Response [John]F likes [Mary]F. like(j, m) SR FC 1. proof search OK prediction: incongruent observation: incongruent 2. -elimination @B like(j, m) An example of overfocusing FC @B : e m : e j : e
  30. How the related examples are derived: multiple wh-questions 32 Question

    Who likes whom? Response [John]F likes [Mary]F. An example of a single-pair answer
  31. How the related examples are derived: multiple wh-questions 33 Question

    Who likes whom? like(@B 1 , @B 2 ) SR FC @B 1 : e Response [John]F likes [Mary]F. SR FC @B 2 : e like(j, m) An example of a single-pair answer m : e j : e
  32. How the related examples are derived: multiple wh-questions 34 Question

    Who likes whom? like(@B 1 , @B 2 ) SR FC @B 1 : e Response [John]F likes [Mary]F. SR FC 1. proof search OK prediction: congruent @B 2 : e 3. entailment observation: congruent like(j, m) 2. -elimination @B like(j, m) An example of a single-pair answer (other eliminations are possible, but ommited) m : e j : e
  33. How the related examples are derived: anaphora 35 Question Who

    does John like? An example of answers with anaphoric expressions Response John likes [him]F. SR FC @ : e like(j, @)
  34. How the related examples are derived: single wh-questions 36 Question

    Who does John like? An example of answers with anaphoric expressions Response John likes [him]F. SR FC @ : e like(j, @) Previous context 1. proof search
  35. How the related examples are derived: single wh-questions 37 Question

    Who does John like? An example of answers with anaphoric expressions Response John likes [him]F. SR FC @ : e like(j, @) 2. -elimination @ SR FC Previous context 1. proof search t : e like(j, t)
  36. How the related examples are derived: single wh-questions 38 Question

    Who does John like? An example of answers with anaphoric expressions Response John likes [him]F. SR FC @ : e like(j, @) 2. -elimination @ SR FC t : e like(j, t) Previous context 1. proof search The remaining procedures are the same as non-anaphoric examples
  37. 40 Question Who likes whom? Response [John]F likes [Mary]F, and

    [Taro]F likes [Susan]F. prediction: incongruent observation: congruent The current analysis cannot capture a pair-list answer Problematic cases (undergeneration) (pointed out by Daiki Matsuoka)
  38. 41 Question Who likes whom? like(@B 1 , @B 2

    ) SR FC @B 1 : e Response [John]F likes [Mary]F, and [Taro]F likes [Susan]F. like(j, m) ∧ like(m, s) SR FC prediction: incongruent @B 2 : e observation: congruent The current analysis cannot capture a pair-list answer Problematic cases (undergeneration) m : e j : e s : e t : e
  39. 42 Question Who likes whom? like(@B 1 , @B 2

    ) SR FC @B 1 : e Response [John]F likes [Mary]F, and [Taro]F likes [Susan]F. like(j, m) ∧ like(m, s) SR FC prediction: incongruent @B 2 : e observation: congruent The current analysis cannot capture a pair-list answer Problematic cases (undergeneration) m : e j : e s : e t : e -elimination @B like(j, m) proof search OK (others ommited)
  40. 43 Question Who does John like? Response John likes Mary,

    and [Mary]F is tall. prediction: congruent observation: incongruent FCs lacks positional information Problematic cases (overgeneration)
  41. 44 Question Who does John like? Response John likes Mary,

    and [Mary]F is tall. prediction: congruent observation: incongruent FCs lacks positional information Problematic cases (overgeneration) like(j, @B) SR FC @B : e like(j, m) ∧ tall(m) SR FC m : e
  42. Summary • Poposal: • Focus contexts Sequences of type assignments

    which represent 
 focus information • Backward underspecified terms Gaps in wh-questions which are expected to be filled 
 by an answer • Future tasks: Avoiding over/undergeneration, comparison with other systems, … 45
  43. References [Bekki 2014] “Representing anaphora with dependent types”. In: International

    conference on logical aspects of computational linguistics. Springer. 2014. [Rooth 1992] “A theory of focus interpretation”. In: Natural language semantics 1.1 (1992). [Watanabe+ 2019] “Questions in Dependent Type Semantics”. In: Proceedings of the Sixth Workshop on Natural Language and Computer Science. Gothenburg, Sweden: Association for Computational Linguistics, May 2019. 46
  44. Halliday’s generalization (Halliday, 1976) “An appropriate answer to a wh-question

    must have F marking on the constituent corresponding to the wh-phrase.” 47