Learning in the Rational Speech Acts Model Christopher Potts - - PowerPoint PPT Presentation

▶
learning in the rational speech acts model
SMART_READER_LITE
LIVE PREVIEW

Learning in the Rational Speech Acts Model Christopher Potts - - PowerPoint PPT Presentation

Nov 29, 2022 357 likes •1.72k views

Overview RSA TUNA Learned RSA Experiments Conclusion Learning in the Rational Speech Acts Model Christopher Potts Stanford Linguistics Paper: http://arxiv.org/abs/1510.06807 Will Monroe 1 / 29 Overview RSA TUNA Learned RSA Experiments

slide-1
SLIDE 1

Overview RSA TUNA Learned RSA Experiments Conclusion

Learning in the Rational Speech Acts Model

Christopher Potts

Stanford Linguistics

Paper: http://arxiv.org/abs/1510.06807

Will Monroe

1 / 29

slide-2
SLIDE 2

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

2 / 29

slide-3
SLIDE 3

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

Bayesian pragmatic models

2 / 29

slide-4
SLIDE 4

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

Bayesian pragmatic models

  • Analytic insights
  • Small problems
  • Idealized agents
  • Constrained by model

design

2 / 29

slide-5
SLIDE 5

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

Bayesian pragmatic models

  • Analytic insights
  • Small problems
  • Idealized agents
  • Constrained by model

design Machine learning

2 / 29

slide-6
SLIDE 6

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

Bayesian pragmatic models

  • Analytic insights
  • Small problems
  • Idealized agents
  • Constrained by model

design Machine learning

  • Coverage
  • Huge, unruly problems
  • Fallible agents
  • Constrained by

experiences in training

2 / 29

slide-7
SLIDE 7

Overview RSA TUNA Learned RSA Experiments Conclusion

Bringing machine learning and pragmatics together

Bayesian pragmatic models

  • Analytic insights
  • Small problems
  • Idealized agents
  • Constrained by model

design Machine learning

  • Coverage
  • Huge, unruly problems
  • Fallible agents
  • Constrained by

experiences in training Goal: combine the best aspects of both to achieve broader coverage and novel pragmatic insights

2 / 29

slide-8
SLIDE 8

Overview RSA TUNA Learned RSA Experiments Conclusion

Plan

1 The Rational Speech Acts (RSA) model 2 TUNA 3 Learned RSA 4 Experiments

3 / 29

slide-9
SLIDE 9

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-10
SLIDE 10

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-11
SLIDE 11

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-12
SLIDE 12

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-13
SLIDE 13

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-14
SLIDE 14

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-15
SLIDE 15

Overview RSA TUNA Learned RSA Experiments Conclusion

Bias and variance

4 / 29

slide-16
SLIDE 16

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

5 / 29

slide-17
SLIDE 17

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

Chat80:

  • 5 / 29
slide-18
SLIDE 18

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’

5 / 29

slide-19
SLIDE 19

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ N → dog : dog V → dog : dogv N → dog : cat N → cat : cat N → cat : dog V → jump : dog V → jump : jump

5 / 29

slide-20
SLIDE 20

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 1 N → dog : dog V → dog : dogv N → dog : cat N → cat : cat N → cat : dog V → jump : dog V → jump : jump

N dog : dog

5 / 29

slide-21
SLIDE 21

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 1 N → dog : dog 1 V → dog : dogv N → dog : cat N → cat : cat N → cat : dog V → jump : dog V → jump : jump

N dog : dog V dog : dogv

5 / 29

slide-22
SLIDE 22

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 1 N → dog : dog 1 V → dog : dogv N → dog : cat 1 N → cat : cat N → cat : dog V → jump : dog V → jump : jump

N dog : dog V dog : dogv N cat : cat

5 / 29

slide-23
SLIDE 23

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 2 N → dog : dog 1 V → dog : dogv N → dog : cat 1 N → cat : cat N → cat : dog V → jump : dog V → jump : jump

N dog : dog V dog : dogv N cat : cat N dog : dog

5 / 29

slide-24
SLIDE 24

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 3 N → dog : dog 1 V → dog : dogv N → dog : cat 1 N → cat : cat N → cat : dog V → jump : dog V → jump : jump

N dog : dog N dog : dog V dog : dogv N cat : cat N dog : dog

5 / 29

slide-25
SLIDE 25

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 3 N → dog : dog 1 V → dog : dogv N → dog : cat 1 N → cat : cat N → cat : dog V → jump : dog 1 V → jump : jump

N dog : dog N dog : dog V dog : dogv V jump : jump N cat : cat N dog : dog

5 / 29

slide-26
SLIDE 26

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 3 N → dog : dog 1 V → dog : dogv N → dog : cat 2 N → cat : cat N → cat : dog V → jump : dog 1 V → jump : jump

N dog : dog N dog : dog V dog : dogv V jump : jump N cat : cat N cat : cat N dog : dog

5 / 29

slide-27
SLIDE 27

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’ 3 N → dog : dog 1 V → dog : dogv N → dog : cat 2 N → cat : cat N → cat : dog V → jump : dog 1 V → jump : jump

N dog : dog N dog : dog V dog : dogv V jump : jump N cat : cat N cat : cat N dog : dog

5 / 29

slide-28
SLIDE 28

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

1 for w ∈ Words 2 for X ∈ Categories 3 for d ∈ Domain 4 yield ‘X → w : d’

  • 5 / 29
slide-29
SLIDE 29

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

Zettlemoyer & Collins 2005:

Rules Categories produced from logical form Input Trigger Output Category arg max(λx.state(x) ^ borders(x, texas), λx.size(x)) constant c NP : c NP : texas arity one predicate p1 N : λx.p1(x) N : λx.state(x) arity one predicate p1 S\NP : λx.p1(x) S\NP : λx.state(x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(y, x) (S\NP )/NP : λx.λy.borders(y, x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(x, y) (S\NP )/NP : λx.λy.borders(x, y) arity one predicate p1 N/N : λg.λx.p1(x) ^ g(x) N/N : λg.λx.state(x) ^ g(x) literal with arity two predicate p2 and constant second argument c N/N : λg.λx.p2(x, c) ^ g(x) N/N : λg.λx.borders(x, texas) ^ g(x) arity two predicate p2 (N\N)/NP : λx.λg.λy.p2(x, y) ^ g(x) (N\N)/NP : λg.λx.λy.borders(x, y) ^ g(x) an arg max / min with second argument arity one function f NP/N : λg. arg max / min(g, λx.f(x)) NP/N : λg. arg max(g, λx.size(x)) an arity one numeric-ranged function f S/NP : λx.f(x) S/NP : λx.size(x)

5 / 29

slide-30
SLIDE 30

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

Zettlemoyer & Collins 2005:

Rules Categories produced from logical form Input Trigger Output Category arg max(λx.state(x) ^ borders(x, texas), λx.size(x)) constant c NP : c NP : texas arity one predicate p1 N : λx.p1(x) N : λx.state(x) arity one predicate p1 S\NP : λx.p1(x) S\NP : λx.state(x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(y, x) (S\NP )/NP : λx.λy.borders(y, x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(x, y) (S\NP )/NP : λx.λy.borders(x, y) arity one predicate p1 N/N : λg.λx.p1(x) ^ g(x) N/N : λg.λx.state(x) ^ g(x) literal with arity two predicate p2 and constant second argument c N/N : λg.λx.p2(x, c) ^ g(x) N/N : λg.λx.borders(x, texas) ^ g(x) arity two predicate p2 (N\N)/NP : λx.λg.λy.p2(x, y) ^ g(x) (N\N)/NP : λg.λx.λy.borders(x, y) ^ g(x) an arg max / min with second argument arity one function f NP/N : λg. arg max / min(g, λx.f(x)) NP/N : λg. arg max(g, λx.size(x)) an arity one numeric-ranged function f S/NP : λx.f(x) S/NP : λx.size(x)

constant c NP : c arity one predicate

5 / 29

slide-31
SLIDE 31

Overview RSA TUNA Learned RSA Experiments Conclusion

Semantic parsing

Zettlemoyer & Collins 2005:

Rules Categories produced from logical form Input Trigger Output Category arg max(λx.state(x) ^ borders(x, texas), λx.size(x)) constant c NP : c NP : texas arity one predicate p1 N : λx.p1(x) N : λx.state(x) arity one predicate p1 S\NP : λx.p1(x) S\NP : λx.state(x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(y, x) (S\NP )/NP : λx.λy.borders(y, x) arity two predicate p2 (S\NP )/NP : λx.λy.p2(x, y) (S\NP )/NP : λx.λy.borders(x, y) arity one predicate p1 N/N : λg.λx.p1(x) ^ g(x) N/N : λg.λx.state(x) ^ g(x) literal with arity two predicate p2 and constant second argument c N/N : λg.λx.p2(x, c) ^ g(x) N/N : λg.λx.borders(x, texas) ^ g(x) arity two predicate p2 (N\N)/NP : λx.λg.λy.p2(x, y) ^ g(x) (N\N)/NP : λg.λx.λy.borders(x, y) ^ g(x) an arg max / min with second argument arity one function f NP/N : λg. arg max / min(g, λx.f(x)) NP/N : λg. arg max(g, λx.size(x)) an arity one numeric-ranged function f S/NP : λx.f(x) S/NP : λx.size(x)

arity one predicate p1 N : λx.p1(x) arity one predicate p1 S\NP : λx.p1(x) arity two predicate

5 / 29

slide-32
SLIDE 32

Overview RSA TUNA Learned RSA Experiments Conclusion

The Rational Speech Acts Model (RSA)

1 The Rational Speech Acts (RSA) model 2 TUNA 3 Learned RSA 4 Experiments

6 / 29

slide-33
SLIDE 33

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

7 / 29

slide-34
SLIDE 34

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner

7 / 29

slide-35
SLIDE 35

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner
  • Politeness

7 / 29

slide-36
SLIDE 36

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner
  • Politeness
  • Stylishness

7 / 29

slide-37
SLIDE 37

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner
  • Politeness
  • Stylishness
  • Q principle
  • R principle

7 / 29

slide-38
SLIDE 38

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner
  • Politeness
  • Stylishness
  • Q principle
  • R principle
  • Q heuristic
  • I heuristic
  • M heuristic

7 / 29

slide-39
SLIDE 39

Overview RSA TUNA Learned RSA Experiments Conclusion

Grice

Cooperative principle: Make your contribution as is required, when it is required, by the conversation in which you are engaged.

  • Quantity
  • Quality
  • Relevance
  • Manner
  • Politeness
  • Stylishness
  • Q principle
  • R principle
  • Q heuristic
  • I heuristic
  • M heuristic

7 / 29

slide-40
SLIDE 40

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

8 / 29

slide-41
SLIDE 41

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

8 / 29

slide-42
SLIDE 42

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative.

8 / 29

slide-43
SLIDE 43

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q.

8 / 29

slide-44
SLIDE 44

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

8 / 29

slide-45
SLIDE 45

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

8 / 29

slide-46
SLIDE 46

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

Ann: What city does Paul live in? Bob: Hmm . . . he lives in California.

8 / 29

slide-47
SLIDE 47

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

Ann: What city does Paul live in? Bob: Hmm . . . he lives in California. (A) Assume Bob is cooperative.

8 / 29

slide-48
SLIDE 48

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

Ann: What city does Paul live in? Bob: Hmm . . . he lives in California. (A) Assume Bob is cooperative. (B) Bob supplied less information than required; clash with (A).

8 / 29

slide-49
SLIDE 49

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

Ann: What city does Paul live in? Bob: Hmm . . . he lives in California. (A) Assume Bob is cooperative. (B) Bob supplied less information than required; clash with (A). (C) Assume Bob does not know which city Paul lives in.

8 / 29

slide-50
SLIDE 50

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Example

Ann: What city does Paul live in? Bob: Hmm . . . he lives in California. (A) Assume Bob is cooperative. (B) Bob supplied less information than required; clash with (A). (C) Assume Bob does not know which city Paul lives in. (D) Then Bob’s answer is optimal given his evidence.

8 / 29

slide-51
SLIDE 51

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Implicature as

  • Rooted in cooperativity
  • Social, interactional
  • Cognitively complex
  • Error-driven

8 / 29

slide-52
SLIDE 52

Overview RSA TUNA Learned RSA Experiments Conclusion

Conversational implicature

Definition

Speaker S saying U to listener L conversationally implicates q iff

1 S and L mutually, publicly presume that S is cooperative. 2 To maintain 1 given U, it must be supposed that S thinks q. 3 S thinks that both S and L mutually, publicly presume that L is

willing and able to work out that

2 holds.

Implicature as

  • Rooted in cooperativity
  • Social, interactional
  • Cognitively complex
  • Error-driven

8 / 29

slide-53
SLIDE 53

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic listeners

9 / 29

slide-54
SLIDE 54

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic listeners

Definition (Literal listener)

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

9 / 29

slide-55
SLIDE 55

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic listeners

Definition (Pragmatic speaker)

s1(msg | w, Lex) ∝ exp λ (log l0(w | msg, Lex) − C(msg))

Definition (Literal listener)

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

9 / 29

slide-56
SLIDE 56

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic listeners

Definition (Pragmatic listener)

l1(w | msg, Lex) ∝ s1(msg | w, Lex)P(w)

Definition (Pragmatic speaker)

s1(msg | w, Lex) ∝ exp λ (log l0(w | msg, Lex) − C(msg))

Definition (Literal listener)

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

9 / 29

slide-57
SLIDE 57

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic listeners

Definition (Pragmatic listener)

l1(w | msg, Lex) = pragmatic speaker × state prior

Definition (Pragmatic speaker)

s1(msg | w, Lex) = literal listener − message costs

Definition (Literal listener)

l0(w | msg, Lex) = lexicon × state prior

9 / 29

slide-58
SLIDE 58

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA listener example

beard 1 glasses 1 1 tie 1 1

l1 s1 l0 Lex

10 / 29

slide-59
SLIDE 59

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA listener example

beard

1

glasses .5 .5 tie .5 .5

l1 s1 l0 Lex

10 / 29

slide-60
SLIDE 60

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA listener example

beard glasses tie

.67

.33

1

0 1

l1 s1 l0 Lex

10 / 29

slide-61
SLIDE 61

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA listener example

beard

1

glasses .25

.75

tie

1

l1 s1 l0 Lex

10 / 29

slide-62
SLIDE 62

Overview RSA TUNA Learned RSA Experiments Conclusion

More Gricean terrain

11 / 29

slide-63
SLIDE 63

Overview RSA TUNA Learned RSA Experiments Conclusion

More Gricean terrain

  • Lexical uncertainty and the division of pragmatic labor

Bergen, Levy, Goodman, ‘Pragmatic reasoning through semantic inference’

11 / 29

slide-64
SLIDE 64

Overview RSA TUNA Learned RSA Experiments Conclusion

More Gricean terrain

  • Lexical uncertainty and the division of pragmatic labor

Bergen, Levy, Goodman, ‘Pragmatic reasoning through semantic inference’

  • Lexical uncertainty and embedded implicatures

Potts, Lassiter, Levy, Frank, ‘Embedded implicatures as pragmatic inferences under compositional lexical uncertainty’

11 / 29

slide-65
SLIDE 65

Overview RSA TUNA Learned RSA Experiments Conclusion

More Gricean terrain

  • Lexical uncertainty and the division of pragmatic labor

Bergen, Levy, Goodman, ‘Pragmatic reasoning through semantic inference’

  • Lexical uncertainty and embedded implicatures

Potts, Lassiter, Levy, Frank, ‘Embedded implicatures as pragmatic inferences under compositional lexical uncertainty’

  • Pragmatic free variables and non-literal language

Kao, Wu, Bergen, Goodman, ‘Nonliteral understanding of number words’

11 / 29

slide-66
SLIDE 66

Overview RSA TUNA Learned RSA Experiments Conclusion

More Gricean terrain

  • Lexical uncertainty and the division of pragmatic labor

Bergen, Levy, Goodman, ‘Pragmatic reasoning through semantic inference’

  • Lexical uncertainty and embedded implicatures

Potts, Lassiter, Levy, Frank, ‘Embedded implicatures as pragmatic inferences under compositional lexical uncertainty’

  • Pragmatic free variables and non-literal language

Kao, Wu, Bergen, Goodman, ‘Nonliteral understanding of number words’

  • Implicature blocking by higher-level agents

Potts & Levy, ‘Negotiating lexical uncertainty and speaker expertise with disjunction’

11 / 29

slide-67
SLIDE 67

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic speakers

12 / 29

slide-68
SLIDE 68

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic speakers

Definition (Literal speaker)

s0(msg | w, Lex) ∝ exp λ (log Lex(msg, w) − C(msg))

12 / 29

slide-69
SLIDE 69

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic speakers

Definition (Pragmatic listener)

l1(w | msg, Lex) ∝ s0(msg | w, Lex)P(w)

Definition (Literal speaker)

s0(msg | w, Lex) ∝ exp λ (log Lex(msg, w) − C(msg))

12 / 29

slide-70
SLIDE 70

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic speakers

Definition (Pragmatic speaker)

s1(msg | w, Lex) ∝ exp λ (log l1(w | msg, Lex) − C(msg))

Definition (Pragmatic listener)

l1(w | msg, Lex) ∝ s0(msg | w, Lex)P(w)

Definition (Literal speaker)

s0(msg | w, Lex) ∝ exp λ (log Lex(msg, w) − C(msg))

12 / 29

slide-71
SLIDE 71

Overview RSA TUNA Learned RSA Experiments Conclusion

Pragmatic speakers

Definition (Pragmatic speaker)

s1(msg | w, Lex) = pragmatic listener − message costs

Definition (Pragmatic listener)

l1(w | msg, Lex) = literal speaker × state prior

Definition (Literal speaker)

s0(msg | w, Lex) = lexicon − message costs

12 / 29

slide-72
SLIDE 72

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA speaker example

beard glasses tie 1 1 1 1 1

s1 l1 s0 Lex

13 / 29

slide-73
SLIDE 73

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA speaker example

beard glasses tie .5 .5 .5 .5 0 1

s1 l1 s0 Lex

13 / 29

slide-74
SLIDE 74

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA speaker example

beard

1

glasses .5 .5 tie .33

.67

s1 l1 s0 Lex

13 / 29

slide-75
SLIDE 75

Overview RSA TUNA Learned RSA Experiments Conclusion

RSA speaker example

beard glasses tie

.67

.33

.6

.4 0 1

s1 l1 s0 Lex

13 / 29

slide-76
SLIDE 76

Overview RSA TUNA Learned RSA Experiments Conclusion

Achievements and drawbacks

beard glasses tie

.67

.33

.6

.4 0 1

14 / 29

slide-77
SLIDE 77

Overview RSA TUNA Learned RSA Experiments Conclusion

Achievements and drawbacks

beard glasses tie

.67

.33

.6

.4 0 1

  • Cognitive demands limit

speaker rationality

14 / 29

slide-78
SLIDE 78

Overview RSA TUNA Learned RSA Experiments Conclusion

Achievements and drawbacks

beard glasses tie

.67

.33

.6

.4 0 1

  • Cognitive demands limit

speaker rationality

  • Speaker preferences

14 / 29

slide-79
SLIDE 79

Overview RSA TUNA Learned RSA Experiments Conclusion

Achievements and drawbacks

beard glasses tie

.67

.33

.6

.4 0 1

  • Cognitive demands limit

speaker rationality

  • Speaker preferences
  • Hand-specified lexicon

14 / 29

slide-80
SLIDE 80

Overview RSA TUNA Learned RSA Experiments Conclusion

Achievements and drawbacks

beard glasses tie

.67

.33

.6

.4 0 1

  • Cognitive demands limit

speaker rationality

  • Speaker preferences
  • Hand-specified lexicon
  • High-bias model; few

chances to learn from data

14 / 29

slide-81
SLIDE 81

Overview RSA TUNA Learned RSA Experiments Conclusion

TUNA

1 The Rational Speech Acts (RSA) model 2 TUNA 3 Learned RSA 4 Experiments

15 / 29

slide-82
SLIDE 82

Overview RSA TUNA Learned RSA Experiments Conclusion

Furniture example

colour:green

  • rientation:left

size:small type:fan x-dimension:1 y-dimension:1 colour:green

  • rientation:left

size:small type:sofa x-dimension:1 y-dimension:2 colour:red

  • rientation:back

size:large type:fan x-dimension:1 y-dimension:3 colour:red

  • rientation:back

size:large type:sofa x-dimension:2 y-dimension:1 colour:blue

  • rientation:left

size:large type:fan x-dimension:2 y-dimension:2 colour:blue

  • rientation:left

size:large type:sofa x-dimension:3 y-dimension:1 colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3

Utterance: “blue fan small” Utterance attributes: [colour:blue]; [size:small]; [type:fan]

16 / 29

slide-83
SLIDE 83

Overview RSA TUNA Learned RSA Experiments Conclusion

Furniture example

colour:green

  • rientation:left

size:small type:fan x-dimension:1 y-dimension:1 colour:green

  • rientation:left

size:small type:sofa x-dimension:1 y-dimension:2 colour:red

  • rientation:back

size:large type:fan x-dimension:1 y-dimension:3 colour:red

  • rientation:back

size:large type:sofa x-dimension:2 y-dimension:1 colour:blue

  • rientation:left

size:large type:fan x-dimension:2 y-dimension:2 colour:blue

  • rientation:left

size:large type:sofa x-dimension:3 y-dimension:1 colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3

Utterance: “blue fan small” Utterance attributes: [colour:blue]; [size:small]; [type:fan]

16 / 29

slide-84
SLIDE 84

Overview RSA TUNA Learned RSA Experiments Conclusion

Furniture example

colour:green

  • rientation:left

size:small type:fan x-dimension:1 y-dimension:1 colour:green

  • rientation:left

size:small type:sofa x-dimension:1 y-dimension:2 colour:red

  • rientation:back

size:large type:fan x-dimension:1 y-dimension:3 colour:red

  • rientation:back

size:large type:sofa x-dimension:2 y-dimension:1 colour:blue

  • rientation:left

size:large type:fan x-dimension:2 y-dimension:2 colour:blue

  • rientation:left

size:large type:sofa x-dimension:3 y-dimension:1 colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3

Utterance: “blue fan small” Utterance attributes: [colour:blue]; [size:small]; [type:fan]

16 / 29

slide-85
SLIDE 85

Overview RSA TUNA Learned RSA Experiments Conclusion

People example

age:old hairColour:light hasBeard:1 hasGlasses:0 hasHair:0 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:left

type:person x-dimension:1 y-dimension:1 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:1 y-dimension:2 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:3 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:3

Utterance: The bald man with a beard [hasBeard:1]; [hasHair:0]; [type:person]

17 / 29

slide-86
SLIDE 86

Overview RSA TUNA Learned RSA Experiments Conclusion

People example

age:old hairColour:light hasBeard:1 hasGlasses:0 hasHair:0 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:left

type:person x-dimension:1 y-dimension:1 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:1 y-dimension:2 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:3 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:3

Utterance: The bald man with a beard [hasBeard:1]; [hasHair:0]; [type:person]

17 / 29

slide-87
SLIDE 87

Overview RSA TUNA Learned RSA Experiments Conclusion

People example

age:old hairColour:light hasBeard:1 hasGlasses:0 hasHair:0 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:left

type:person x-dimension:1 y-dimension:1 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:1 y-dimension:2 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:2 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:1 age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0

  • rientation:front

type:person x-dimension:3 y-dimension:2 age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1

  • rientation:front

type:person x-dimension:3 y-dimension:3

Utterance: The bald man with a beard [hasBeard:1]; [hasHair:0]; [type:person]

17 / 29

slide-88
SLIDE 88

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

2

x∈D min

  • Za(msgi)(x), Za(msgj)(x)
  • |a(msgi)| + |a(msgj)|

18 / 29

slide-89
SLIDE 89

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

Multiset intersection cardinality Multiset union cardinality

18 / 29

slide-90
SLIDE 90

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

Multiset intersection cardinality Multiset union cardinality

  • predicted: [a

b c a] actual: [a b c a] = 1

18 / 29

slide-91
SLIDE 91

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

Multiset intersection cardinality Multiset union cardinality

  • predicted: [a

b c a] actual: [a b c a] = 1

  • predicted: [a

b c a] actual: [a b c

] = .86

18 / 29

slide-92
SLIDE 92

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

Multiset intersection cardinality Multiset union cardinality

  • predicted: [a

b c a] actual: [a b c a] = 1

  • predicted: [a

b c a] actual: [a b c

] = .86

  • predicted: [a

b c

]

actual: [a b c a] = .86

18 / 29

slide-93
SLIDE 93

Overview RSA TUNA Learned RSA Experiments Conclusion

Multiset Dice coefficient

Definition

Multiset intersection cardinality Multiset union cardinality

  • predicted: [a

b c a] actual: [a b c a] = 1

  • predicted: [a

b c a] actual: [a b c

] = .86

  • predicted: [a

b c

]

actual: [a b c a] = .86

  • predicted: [a

]

actual: [a b c a] = .4

18 / 29

slide-94
SLIDE 94

Overview RSA TUNA Learned RSA Experiments Conclusion

Learned RSA

1 The Rational Speech Acts (RSA) model 2 TUNA 3 Learned RSA 4 Experiments

19 / 29

slide-95
SLIDE 95

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

20 / 29

slide-96
SLIDE 96

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

Target Utterance attributes Features colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3 [colour:blue] [size:small] [type:fan]

20 / 29

slide-97
SLIDE 97

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

Target Utterance attributes Features colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3 [colour:blue] [size:small] [type:fan] colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

  • rientation:left ∧ [colour:blue]
  • rientation:left ∧ [size:small]
  • rientation:left ∧ [type:fan]

. . .

20 / 29

slide-98
SLIDE 98

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

Target Utterance attributes Features colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3 [colour:blue] [size:small] [type:fan] colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

  • rientation:left ∧ [colour:blue]
  • rientation:left ∧ [size:small]
  • rientation:left ∧ [type:fan]

. . . Generation features        color

20 / 29

slide-99
SLIDE 99

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

Target Utterance attributes Features colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3 [colour:blue] [size:small] [type:fan] colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

  • rientation:left ∧ [colour:blue]
  • rientation:left ∧ [size:small]
  • rientation:left ∧ [type:fan]

. . . Generation features        color type + color color + ¬size type ≫ color ≫ size type ≫ orientation ≫ color ≫ size

20 / 29

slide-100
SLIDE 100

Overview RSA TUNA Learned RSA Experiments Conclusion

Feature representations

Target Utterance attributes Features colour:blue

  • rientation:left

size:small type:fan x-dimension:3 y-dimension:3 [colour:blue] [size:small] [type:fan] colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

  • rientation:left ∧ [colour:blue]
  • rientation:left ∧ [size:small]
  • rientation:left ∧ [type:fan]

. . . Generation features        color type + color color + ¬size attribute-count = 3 type ≫ color ≫ size type ≫ orientation ≫ color ≫ size

20 / 29

slide-101
SLIDE 101

Overview RSA TUNA Learned RSA Experiments Conclusion

Model definition

⊙ ϕ θ

“beard” “guy with the beard” “guy with glasses” ...

S0(m|t ,θ)∝exp[θ

T ϕ(t ,m)]

L1(t|m ,θ)∝S0(m|t ,θ) S1(m|t ,θ)∝ L1(t|m ,θ)

21 / 29

slide-102
SLIDE 102

Overview RSA TUNA Learned RSA Experiments Conclusion

Optimization

“guy with the beard”

⊙ ϕ θ

“beard” “guy with the beard” “guy with glasses” ...

∂ ∂θ log S1(m|t ,θ)

22 / 29

slide-103
SLIDE 103

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features

23 / 29

slide-104
SLIDE 104

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon

23 / 29

slide-105
SLIDE 105

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features

23 / 29

slide-106
SLIDE 106

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production

23 / 29

slide-107
SLIDE 107

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color

23 / 29

slide-108
SLIDE 108

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color Learn attribute hierarchies

23 / 29

slide-109
SLIDE 109

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color Learn attribute hierarchies Attribute-pair features like color + ¬size

23 / 29

slide-110
SLIDE 110

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color Learn attribute hierarchies Attribute-pair features like color + ¬size Learn message costs

23 / 29

slide-111
SLIDE 111

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color Learn attribute hierarchies Attribute-pair features like color + ¬size Learn message costs Length features and others

23 / 29

slide-112
SLIDE 112

Overview RSA TUNA Learned RSA Experiments Conclusion

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features Learn quirks of production Features like color Learn attribute hierarchies Attribute-pair features like color + ¬size Learn message costs Length features and others Cognitive and linguistic insights combined with learning

23 / 29

slide-113
SLIDE 113

Overview RSA TUNA Learned RSA Experiments Conclusion

Example

Train

[person] [glasses] [person] [beard]

Test

24 / 29

slide-114
SLIDE 114

Overview RSA TUNA Learned RSA Experiments Conclusion

Example

∅ [person] [glasses] [beard] [person];[glasses] [person];[beard] [glasses];[beard] [all]

24 / 29

slide-115
SLIDE 115

Overview RSA TUNA Learned RSA Experiments Conclusion

Example

∅

.08 .25

[person]

.08 .25

[glasses]

.17 .00

[beard]

.08 .25

[person];[glasses]

.17 .00

[person];[beard]

.08 .25

[glasses];[beard]

.17 .00

[all]

.17 .00 RSA

24 / 29

slide-116
SLIDE 116

Overview RSA TUNA Learned RSA Experiments Conclusion

Example

∅

.08 .25 .03 .00

[person]

.08 .25 .22 .10

[glasses]

.17 .00 .03 .00

[beard]

.08 .25 .03 .04

[person];[glasses]

.17 .00 .22 .01

[person];[beard]

.08 .25 .22 .74

[glasses];[beard]

.17 .00 .03 .00

[all]

.17 .00 .22 .10 RSA Learned S0

24 / 29

slide-117
SLIDE 117

Overview RSA TUNA Learned RSA Experiments Conclusion

Example

∅

.08 .25 .03 .00 .10 .11

[person]

.08 .25 .22 .10 .16 .13

[glasses]

.17 .00 .03 .00 .11 .07

[beard]

.08 .25 .03 .04 .08 .17

[person];[glasses]

.17 .00 .22 .01 .18 .08

[person];[beard]

.08 .25 .22 .74 .12 .19

[glasses];[beard]

.17 .00 .03 .00 .10 .11

[all]

.17 .00 .22 .10 .16 .11 RSA Learned S0 Learned S1

24 / 29

slide-118
SLIDE 118

Overview RSA TUNA Learned RSA Experiments Conclusion

Experiments

1 The Rational Speech Acts (RSA) model 2 TUNA 3 Learned RSA 4 Experiments

25 / 29

slide-119
SLIDE 119

Overview RSA TUNA Learned RSA Experiments Conclusion

Experimental set-up

  • TUNA furniture: 420 trials, 176 distinct referents
  • TUNA people: 360 trials, 228 distinct referents
  • Five-fold cross-validation for all models
  • RSA cross-validation sets λ and the cost function
  • Learned RSA optimized via AdaGrad with ℓ2 regularization

26 / 29

slide-120
SLIDE 120

Overview RSA TUNA Learned RSA Experiments Conclusion

Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people

27 / 29

slide-121
SLIDE 121

Overview RSA TUNA Learned RSA Experiments Conclusion

Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people RSA s1

0.522

RSA s1

0.254

27 / 29

slide-122
SLIDE 122

Overview RSA TUNA Learned RSA Experiments Conclusion

Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people RSA s1

0.522

Learned S0

0.812

RSA s1

0.254

Learned S0

0.73

27 / 29

slide-123
SLIDE 123

Overview RSA TUNA Learned RSA Experiments Conclusion

Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people RSA s1

0.522

Learned S0

0.812

Learned S1

0.788

RSA s1

0.254

Learned S0

0.73

Learned S1

0.764

27 / 29

slide-124
SLIDE 124

Overview RSA TUNA Learned RSA Experiments Conclusion

Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people RSA s1

0.522

Learned S0

0.812

Learned S1

0.788

RSA s1

0.254

Learned S0

0.73

Learned S1

0.764

* *

27 / 29

slide-125
SLIDE 125

Overview RSA TUNA Learned RSA Experiments Conclusion

Error analysis

50 100 150 200 250 300 350

Underproductions of attribute [type:person] [hasBeard:true]

(Lower is better!)

28 / 29

slide-126
SLIDE 126

Overview RSA TUNA Learned RSA Experiments Conclusion

Error analysis

50 100 150 200 250 300 350

Underproductions of attribute [type:person] [hasBeard:true] RSA s1 RSA s1

(Lower is better!)

28 / 29

slide-127
SLIDE 127

Overview RSA TUNA Learned RSA Experiments Conclusion

Error analysis

50 100 150 200 250 300 350

Underproductions of attribute [type:person] [hasBeard:true] RSA s1 Learned S0 RSA s1 Learned S0

(Lower is better!)

28 / 29

slide-128
SLIDE 128

Overview RSA TUNA Learned RSA Experiments Conclusion

Error analysis

50 100 150 200 250 300 350

Underproductions of attribute [type:person] [hasBeard:true] RSA s1 Learned S0 Learned S1 RSA s1 Learned S0 Learned S1

(Lower is better!)

28 / 29

slide-129
SLIDE 129

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

29 / 29

slide-130
SLIDE 130

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

29 / 29

slide-131
SLIDE 131

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

  • Cognitive and linguistic insights combined with learning

29 / 29

slide-132
SLIDE 132

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

  • Cognitive and linguistic insights combined with learning
  • Learning a lexicon, speaker preferences, and

context-dependent disambiguation in one process

29 / 29

slide-133
SLIDE 133

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

  • Cognitive and linguistic insights combined with learning
  • Learning a lexicon, speaker preferences, and

context-dependent disambiguation in one process

  • Next: learning in extended versions of RSA

29 / 29

slide-134
SLIDE 134

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

  • Cognitive and linguistic insights combined with learning
  • Learning a lexicon, speaker preferences, and

context-dependent disambiguation in one process

  • Next: learning in extended versions of RSA
  • Next: increased scalability via variational inference

29 / 29

slide-135
SLIDE 135

Overview RSA TUNA Learned RSA Experiments Conclusion

Summary and general lessons

  • Best model of production: established insights about natural

language generation synthesized with RSA

  • Cognitive and linguistic insights combined with learning
  • Learning a lexicon, speaker preferences, and

context-dependent disambiguation in one process

  • Next: learning in extended versions of RSA
  • Next: increased scalability via variational inference

Thanks!

29 / 29