Learning in extended and approximate Rational Speech Acts models - - PowerPoint PPT Presentation

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Learning in extended and approximate Rational Speech Acts models

Christopher Potts

Stanford Linguistics

EMNLP 2016

Will Monroe

1 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Gricean pragmatics

• The cooperative principle: Make your

contribution as is required, when it is required, by the conversation in which you are engaged.

2 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Gricean pragmatics

• The cooperative principle: Make your

contribution as is required, when it is required, by the conversation in which you are engaged.

• Quality: Contribute only what you know to be true.

Do not say false things. Do not say things for which you lack evidence.

• Quantity: Make your contribution as informative as

is required. Do not say more than is required.

• Relation (Relevance): Make your contribution

relevant.

• Manner: (i) Avoid obscurity; (ii) avoid ambiguity;

(iii) be brief; (iv) be orderly.

• Politeness: Be polite, so be tactful, respectful,

generous, praising, modest, deferential, and

• sympathetic. (Leech)

2 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Gricean pragmatics

• The cooperative principle: Make your

contribution as is required, when it is required, by the conversation in which you are engaged.

• Quality: Contribute only what you know to be true.

Do not say false things. Do not say things for which you lack evidence.

• Quantity: Make your contribution as informative as

is required. Do not say more than is required.

• Relation (Relevance): Make your contribution

relevant.

• Manner: (i) Avoid obscurity; (ii) avoid ambiguity;

(iii) be brief; (iv) be orderly.

• Politeness: Be polite, so be tactful, respectful,

generous, praising, modest, deferential, and

• sympathetic. (Leech)

2 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Overview

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

3 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

4 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

John Stuart Mill: I saw some of your children to-day invites the inference that I didn’t see all of them some(X, Y) all(X, Y)

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

John Stuart Mill: I saw some of your children to-day invites the inference that I didn’t see all of them “not because the words mean it, but because, if I had seen them all, it is most likely that I should have said so.” some(X, Y) all(X, Y)

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

Generalization: Using a general term invites the inference that its more specific, salient alternatives are inappropriate. general specific

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

Generalization: Using a general term invites the inference that its more specific, salient alternatives are inappropriate. general specific

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

George Bush: “As I understand it, the current form asks the question ‘Did somebody use drugs within the last seven years?’, and I will be glad to answer that question, and the answer is ‘No’.” no drugs in 7 years never drugs

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

Chris Potts: Watching TV in your underwear – that’s a scalar implicature!

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

Chris Potts: Watching TV in your underwear – that’s a scalar implicature! underwear under&outer-wear

5 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar diversity

20 40 60 80 100

cheap/free sometimes/always some/all possible/certain may/will difficult/impossible rare/extinct may/have to warm/hot few/none low/depleted hard/unsolvable allowed/obligatory scarce/unavailable try/succeed palatable/delicious memorable/unforgettable like/love good/perfect good/excellent cool/cold hungry/starving adequate/good unsettling/horrific dislike/loathe believe/know start/finish participate/win wary/scared

• ld/ancient

big/enormous snug/tight attractive/stunning special/unique pretty/beautiful intelligent/brilliant funny/hilarious dark/black small/tiny ugly/hideous silly/ridiculous tired/exhausted content/happy

van Tiel, van Miltenburg, Zevakhina, and Geurts, ‘Scalar diversity’

6 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar diversity

20 40 60 80 100

cheap/free sometimes/always some/all possible/certain may/will difficult/impossible rare/extinct may/have to warm/hot few/none low/depleted hard/unsolvable allowed/obligatory scarce/unavailable try/succeed palatable/delicious memorable/unforgettable like/love good/perfect good/excellent cool/cold hungry/starving adequate/good unsettling/horrific dislike/loathe believe/know start/finish participate/win wary/scared

• ld/ancient

big/enormous snug/tight attractive/stunning special/unique pretty/beautiful intelligent/brilliant funny/hilarious dark/black small/tiny ugly/hideous silly/ridiculous tired/exhausted content/happy

van Tiel, van Miltenburg, Zevakhina, and Geurts, ‘Scalar diversity’ Also: Judith Degen, ‘Investigating the distribution

• f some (but not all)

implicatures using corpora and web-based methods’

6 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar diversity

6 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar diversity

6 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Partial-order implicature

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

Hirschberg 1985, A Theory of Scalar Implicature

7 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Partial-order implicature

A: Do you speak German? B: My husband does.

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

Hirschberg 1985, A Theory of Scalar Implicature

7 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Partial-order implicature

A: Do you speak German? B: My husband does.

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

A: Are you on your honeymoon? B: Well, I was.

128 AFFIRt-.t(B,rhird chaprer, BEL(B, read(B,third chapter»»

1\

ALT_SENT(read(B,ch3prec_one), read(B,thirdj:hapter), 'parts of a dissertation'»

=) SCALAR_IMP(B, A, I read the third, ...,BEL(B, read(B,chapter_one»,

Ci) There are no restrictions on those posers which support scalar implicamre. However, (ar least) one restriction does exist on which posers may be viewed as salient in a given exchange: Above (Section 5.1.6.3) I noted for most metrics that rank utterances. both a given metric and irs dual (converse) may be candidares for salience in an exchange. However, no metric (ji which orders values vi and Vj such that a) vi is higher than Vj and b} the truth of Vj entails the truth of vi can supporr scalar implicature -- for the simple reason in such a case, a sentence Pi ranked higher than a sentence Pj by (ji since then the implicature licensed would be inconsistent with the utterance licensing it. In terms of the fonnaIism presented in Chapter 2, such a meaning would not be reinforceable. Consider, for example. (212a): (212) A: Are you planning to buy a dog?

• a. B: A German Shepherd.
• b. B: I'm buying a Gennan Shepherd and I'm not buying a dog.

While one might identify either an ordering defined by 'isa' (i.e.• a Gennan Shepherd isa dog)

• r by 'subsumes' (i.e., a dog subsumes the subtype

Shepherd) as salient in this exchange, only the latter permits scalar implicature here. B cannot implicate that she is not buying a dog vla this response, since buying a Gennan Shepherd entails buying a dog. The attempted reinforcement of (212b) fails. However, we cannot rule out 'isa' relations as potential supporters of scalar imphcarore: In 213, for example. 8's response might evoke either an'isa' (213) A: Would you like a dog? 8: I'd like a German Shepherd. hierarchy - or irs dual. Apparently, any poser can support scaiar implicature, although other tests for conversational implicature may rule out some particuJ3r posers in panicular exchanges.

5.3.2.3. Representing Scalar Implicature Orderings as Pos.ets J have demonstrated above how part! whole re!arions can be represented. To demonstrate

that L'le other orderings discussed in Section 5.1 are accounted for by a poset condition. I w;i! describe how representative orderings can be accommodated by this condition so mat scalar implicatures are correctly predicted by ImPl_3' Rdations defined by ordering the non-null members of the power

• f some set x by

set-inclusion allow a poset representation of x and its non-null proper subsets a5 follows: Any non-null proper subset of a set m<lY be nnked as LOWER than the set which

• it. and

129 set, in consequence, will represent a HIGHER value in the ordering. Subs.ets which are neither included in, nor include, one another, will be ALTERNATE values in this poser. Consider how the salient ordering in the following exchange mighr be represented: (214) A: Do you speak: Portuguese? B: My husband does. The inclusion ordering which supports the implicature in 214 might be represented as follows:

So, {husband,wife.chiid} wi!! be the highest value in this ordering, with the alternate doubletons (husband,wife), (wife.child), and (husband,child) lower values and the alternate values, {husband}, (wife), and {child} lower values still in this poset. By the scalar implicature conventions, then, S may affinn. say, (husband.wife) to convey ...,BEL(S. (husband,wile.child)) as well as -,BEL(S. (husband,child}) and -,BEL(S, (wife.childJ). Note, particularly, that there may be some redundance in scalar implicatures predicted from this representation. Also, any subsets so represented may be lexica1ized in various ways -- as, the expression (husband.wife) might be lexicalized as •couple' or as 'husband and wife'. The theory presented in this thesis will not distinguish between these. 128 As noted in Sections 5.1.7, temporal orderings may also be represented as setS ofrernporal for the analysis of licensed scalar implicacures. So, these orderings too wilt be defined by set inclusion, as: {past, resent} {presenr,future} {past,furure}

r---:::--

{future} Posers defined by a type! subrype metric, such as that which supports 174, may be illustrated by me (parrial) classification hierarchy:

lZ1!Sut see {CorelIa 84, Ka!ita 84) for some approaches to thtS problem.

Hirschberg 1985, A Theory of Scalar Implicature

7 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Highly particularized implicature R1 R2 R3

“glasses”

8 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Reference games

Frank, Gómez, Peloquin, Goodman, and Potts 2016, 10 experiments, each N ≈ 600 (4,651 participants). The summary picture: https://github.com/langcog/pragmods

9 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Reference games

Frank, Gómez, Peloquin, Goodman, and Potts 2016, 10 experiments, each N ≈ 600 (4,651 participants). The summary picture: https://github.com/langcog/pragmods

9 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Reference games

Frank, Gómez, Peloquin, Goodman, and Potts 2016, 10 experiments, each N ≈ 600 (4,651 participants). The summary picture:

betting forced_choice likert 0.00 0.25 0.50 0.75 1.00 foil target logical foil target logical foil target logical

Target Normalized measure mean

https://github.com/langcog/pragmods

9 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat boat motorboat

canoe kayak sailboat

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

• 2. “boat or canoe”

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

• 2. “boat or canoe”
• 3. Kim is in France.

(in Paris)

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

• 2. “boat or canoe”
• 3. Kim is in France.

(in Paris)

• 4. “The nuptials will take place in either France or

Paris.”

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

• 2. “boat or canoe”
• 3. Kim is in France.

(in Paris)

• 4. “The nuptials will take place in either France or

Paris.”

• 5. I hit the button and it started.

(causation)

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

I-implicature

Levinson: “what is simply described is stereotypically exemplified”.

• 1. At a busy marina in water-skiing country:

“boat” interpreted as motorboat

• 2. “boat or canoe”
• 3. Kim is in France.

(in Paris)

• 4. “The nuptials will take place in either France or

Paris.”

• 5. I hit the button and it started.

(causation)

• 6. Sandy finished the book.

(reading)

10 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

M-implicature

Levinson: “What’s said in an abnormal way isn’t normal.” 1.

• a. T

urn on the car.

• b. Get the car to turn on.

2.

• a. Stop the car.
• b. Cause the car to stop.

11 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Sociolinguistic variables

12 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Sociolinguistic variables

Generalization

Where two forms are in salient contrast, the choice of

• ne will lead to inferences about the other.

12 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Sociolinguistic variables

Generalization

Where two forms are in salient contrast, the choice of

• ne will lead to inferences about the other.
• Community: Community members adopt a speech

style that is easily distinguished from the mainstream, enhancing solidarity.

12 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Sociolinguistic variables

Generalization

Where two forms are in salient contrast, the choice of

• ne will lead to inferences about the other.
• Community: Community members adopt a speech

style that is easily distinguished from the mainstream, enhancing solidarity.

• Individual: An individual systematically varies their

speech style by context to construct different personae.

12 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

13 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems
• Rabin 1990: recursive strategic signaling

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems
• Rabin 1990: recursive strategic signaling
• Camerer and Ho 2004: cognitive hierarchy models

for games of conflict and coordination

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems
• Rabin 1990: recursive strategic signaling
• Camerer and Ho 2004: cognitive hierarchy models

for games of conflict and coordination

• Michael Franke and Gerhard Jäger: iterated best

response

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems
• Rabin 1990: recursive strategic signaling
• Camerer and Ho 2004: cognitive hierarchy models

for games of conflict and coordination

• Michael Franke and Gerhard Jäger: iterated best

response

• Golland, Liang, and Klein 2010 (EMNLP): pragmatic

listeners and probabilistic compositionality

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Origin story

• Rosenberg and Cohen 1964: early Bayesian model
• f production and comprehension
• Lewis 1969: signaling systems
• Rabin 1990: recursive strategic signaling
• Camerer and Ho 2004: cognitive hierarchy models

for games of conflict and coordination

• Michael Franke and Gerhard Jäger: iterated best

response

• Golland, Liang, and Klein 2010 (EMNLP): pragmatic

listeners and probabilistic compositionality

• Frank and Goodman 2012 (Science): very

sophisticated pragmatic agents and a new Bayesian foundation

14 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic listeners

15 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic listeners

Literal listener

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

15 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic listeners

Pragmatic speaker

s1(msg | w, Lex) ∝ exp λ (logl0(w | msg, Lex) − C(msg))

Literal listener

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

15 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic listeners

Pragmatic listener

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

Pragmatic speaker

s1(msg | w, Lex) ∝ exp λ (logl0(w | msg, Lex) − C(msg))

Literal listener

l0(w | msg, Lex) ∝ Lex(msg, w)P(w)

15 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic listeners

Pragmatic listener

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

Pragmatic speaker

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

Literal listener

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

15 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA listener example

beard 1 glasses 1 1 tie 1 1

l1 s1 l0 Lex

16 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA listener example

beard

1

glasses .5 .5 tie .5 .5

l1 s1 l0 Lex

16 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA listener example

beard glasses tie

.67

.33

1

0 1

l1 s1 l0 Lex

16 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA listener example

beard

1

glasses .25 .75 tie

1

l1 s1 l0 Lex

16 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic speakers

17 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic speakers

Literal speaker

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

17 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic speakers

Pragmatic listener

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

Literal speaker

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

17 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic speakers

Pragmatic speaker

s1(msg | w, Lex) ∝ exp λ (logl1(w | msg, Lex) − C(msg))

Pragmatic listener

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

Literal speaker

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

17 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pragmatic speakers

Pragmatic speaker

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

Pragmatic listener

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

Literal speaker

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

17 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA speaker example

beard glasses tie 1 1 1 1 1

s1 l1 s0 Lex

18 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA speaker example

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

s1 l1 s0 Lex

18 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA speaker example

beard

1

glasses .5 .5 tie .33 .67

s1 l1 s0 Lex

18 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

RSA speaker example

beard glasses tie

.67

.33

.6

.4 0 1

s1 l1 s0 Lex

18 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Joint reasoning

L(w, Context | msg) ∝ P(w)PC(Context)s1(msg | w, Context)

19 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Joint reasoning

L(w, Context | msg) ∝ P(w)PC(Context)s1(msg | w, Context) L(w | msg) ∝ P(w)

• Context∈C

PC(Context)s1(msg | w, Context)

19 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

• M-implicatures

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

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

• M-implicatures

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

• I-implicatures and implicature blocking

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

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

• M-implicatures

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

• I-implicatures and implicature blocking

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

• Implicatures and compositionality

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

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

• M-implicatures

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

• I-implicatures and implicature blocking

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

• Implicatures and compositionality

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

• Hyperbole

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

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Achievements

• M-implicatures

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

• I-implicatures and implicature blocking

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

• Implicatures and compositionality

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

• Hyperbole

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

• Metaphor

Kao, Bergen, Goodman, ‘Formalizing the pragmatics of metaphor understanding’

20 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Limitations

• Hand-specified lexicon
• High-bias model; few chances to learn from data
• Cognitive demands limit speaker rationality
• Speaker preferences
• Scalability

21 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

Will Monroe

22 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA furniture example

23 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA furniture example

Utterance: “blue fan small”

23 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA 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]

24 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA people example

25 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA people example

Utterance: “The bald man with a beard”

25 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA people example

age:old hairColour:light hasBeard:1 hasGlasses:0 hasHair:0 hasShirt:1 hasSuit:0 hasTie:0 type:person age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0 type:person age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:1 type:person age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1 type:person age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1 type:person age:young hairColour:dark hasBeard:1 hasGlasses:0 hasHair:1 hasShirt:1 hasSuit:0 hasTie:0 type:person age:young hairColour:dark hasBeard:0 hasGlasses:0 hasHair:1 hasShirt:0 hasSuit:1 hasTie:1 type:person

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

26 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Feature representations

  

colour:blue

• rientation:left

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

,

[colour:blue] [size:small] [type:fan]

  

27 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Feature representations

  

colour:blue

• rientation:left

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

,

[colour:blue] [size:small] [type:fan]

   Cross-product features colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

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

. . .

27 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Feature representations

  

colour:blue

• rientation:left

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

,

[colour:blue] [size:small] [type:fan]

   Cross-product features colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

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

. . . Generation features color type + color color + ¬size attribute-count = 3 . . .

27 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Feature representations

  

colour:blue

• rientation:left

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

,

[colour:blue] [size:small] [type:fan]

   Cross-product features colour:blue ∧ [colour:blue] colour:blue ∧ [size:small] colour:blue ∧ [type:fan]

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

. . . Generation features color type + color color + ¬size attribute-count = 3 . . . type ≫ orientation ≫ color ≫ size

27 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

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)

28 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Optimization

“guy with the beard”

⊙ ϕ θ

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

∂ ∂θ log S1(m|t ,θ)

29 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

Goal Features

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

Goal Features Avoid hand-built lexicon Cross-product features

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Addressing the drawbacks of RSA

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

30 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Example

Train [person] [glasses] [person] [beard] T est

31 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

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

32 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

∅ .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

32 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

∅ .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

32 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

∅ .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

32 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA Results

0.0 0.2 0.4 0.6 0.8 1.0

Mean Dice furniture people

33 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA 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

33 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA 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

33 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA 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

33 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

TUNA 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

* *

33 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Error analysis

50 100 150 200 250 300 350

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

(Lower is better!)

34 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Error analysis

50 100 150 200 250 300 350

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

(Lower is better!)

34 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

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!)

34 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

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!)

34 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Limitations

• Hand-specified lexicon
• High-bias model; few chances to learn from data
• Cognitive demands limit speaker rationality
• Speaker preferences
• Scalability

35 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Limitations

• Hand-specified lexicon
• High-bias model; few chances to learn from data
• Cognitive demands limit speaker rationality
• Speaker preferences
• Scalability

35 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

Robert Hawkins Will Monroe Noah Goodman

36 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Color reference

Color Utterance xxxx violet xxxx blue xxxx dark green xxxx the best color in the freakin’ world!!!

T able: Examples from the xkcd color survey

Color papers at this conference, Friday: Monroe et al. (Session 8A) and Kawakami et al. (Session P8)

37 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

Context Utterance xxxx xxxx xxxx blue

T able: Example from the Colors in Context corpus from the Stanford Computation & Cognition Lab

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

Context Utterance xxxx xxxx xxxx blue xxxx xxxx xxxx The darker blue one

T able: Example from the Colors in Context corpus from the Stanford Computation & Cognition Lab

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

Context Utterance xxxx xxxx xxxx blue xxxx xxxx xxxx The darker blue one xxxx xxxx xxxx dull pink not the super bright one

T able: Example from the Colors in Context corpus from the Stanford Computation & Cognition Lab

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

Context Utterance xxxx xxxx xxxx blue xxxx xxxx xxxx The darker blue one xxxx xxxx xxxx dull pink not the super bright one xxxx xxxx xxxx Purple

T able: Example from the Colors in Context corpus from the Stanford Computation & Cognition Lab

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Colors in context

Context Utterance xxxx xxxx xxxx blue xxxx xxxx xxxx The darker blue one xxxx xxxx xxxx dull pink not the super bright one xxxx xxxx xxxx Purple xxxx xxxx xxxx blue

T able: Example from the Colors in Context corpus from the Stanford Computation & Cognition Lab

38 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Literal neural speaker S0

c1 c2 cT h h; 〈s〉 h;x1 h;x2 x1 x2 〈/s〉 LSTM Fully connected softmax

39 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural literal listener L0

x1 x2 x3 (μ, Σ) c1 c2 c3

• c3

Embedding LSTM Softmax

40 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural pragmatic agents

41 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural pragmatic agents

Neural pragmatic speaker (Andreas & Klein, here!)

S1(msg | c, C; θ) = L0(c | msg, C; θ)

• msg′∈X L0(c | msg′, C; θ)

41 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural pragmatic agents

Neural pragmatic speaker (Andreas & Klein, here!)

S1(msg | c, C; θ) = L0(c | msg, C; θ)

• msg′∈X L0(c | msg′, C; θ)

where X is a sample from S0(msg | c, C; θ) such that msg∗ ∈ X.

41 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural pragmatic agents

Neural pragmatic speaker (Andreas & Klein, here!)

S1(msg | c, C; θ) = L0(c | msg, C; θ)

• msg′∈X L0(c | msg′, C; θ)

where X is a sample from S0(msg | c, C; θ) such that msg∗ ∈ X.

Neural pragmatic listener

L1(c | msg, C; θ) ∝ S1(msg | c, C; θ)

41 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Neural pragmatic agents

Neural pragmatic speaker (Andreas & Klein, here!)

S1(msg | c, C; θ) = L0(c | msg, C; θ)

• msg′∈X L0(c | msg′, C; θ)

where X is a sample from S0(msg | c, C; θ) such that msg∗ ∈ X.

Neural pragmatic listener

L1(c | msg, C; θ) ∝ S1(msg | c, C; θ)

Blended neural pragmatic listener

Weighted combination of L0 and L1.

41 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Language and action

• 1. Meaning from a communicative tension
• 2. The Rational Speech Acts (RSA) model
• 3. Learning in the Rational Speech Acts Model
• 4. Neural RSA
• 5. Language and action

Adam Vogel Dan Jurafsky

42 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

The Cards task

You are on 2D Yellow boxes mark cards in your line of sight. Task description: Six consecutive cards of the same suit TYPE HERE The cards you are holding Move with the arrow keys or these buttons.

43 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

The Cards task

Gather six consecutive cards of the same suit (decide which suit together) or determine that this is

• impossible. Each of you can hold only three cards at a

time, so you’ll have to coordinate your efforts. You can talk all you want, but you can make only a limited number of moves.

43 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

The Cards task

Gather six consecutive cards of the same suit (decide which suit together) or determine that this is

• impossible. Each of you can hold only three cards at a

time, so you’ll have to coordinate your efforts. You can talk all you want, but you can make only a limited number of moves. What’s going on? ⇓ Which suit should we pursue? ⇓ Which sequence should we pursue? ⇓ Where is card X?

43 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

T ask-oriented dialogue corpora

Corpus T ask type Domain T ask-orient. Docs. Format Switchboard discussion open very loose 2,400 aud/txt SCARE search 3d world tight 15 aud/vid/txt TRAINS routes map tight 120 aud/txt Map T ask routes map tight 128 aud/vid/txt Columbia Games games maps tight 12 aud/txt Settlers strategy board tight 40 txt Cards search 2d grid tight 1,266 txt

Chief selling points for Cards:

• Pretty large
• Controlled enough that similar things happen often
• Very highly structured

44 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Simplified cards scenario

Both agents must find the ace of spades.

45 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

• A POMDP agent that learns to navigate its world

and interpret language.

• Driven by its small negative reward for not having

the card and its large positive reward for finding it.

• No sensitivity to the other player.
• Literal listeners: each message msg denotes

P(w | msg) estimated from the Cards corpus.

• Bayes rule to incorporate these as observations.

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

• A POMDP agent that learns to navigate its world

and interpret language.

• Driven by its small negative reward for not having

the card and its large positive reward for finding it.

• No sensitivity to the other player.
• Literal listeners: each message msg denotes

P(w | msg) estimated from the Cards corpus.

• Bayes rule to incorporate these as observations.

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

“it’s on the left side” ⇒ board(left) ⇒

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

ListenerBot

“it’s on the left side” ⇒ board(left) ⇒

46 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

DialogBot

A strict extension of Listener Bot:

• The set of states is now all combinations of

◮ both players’ positions ◮ the card’s region ◮ the region the other player believes the card to

be in

• The set of actions now includes dialogue actions.
• Same basic reward structure as for Listenerbot,

except now also sensitive to whether the other player has found the card.

• Speech actions are modeled in terms of how they

affect the agent’s estimation of the belief state of the other agent.

47 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

DialogBot

A strict extension of Listener Bot:

• The set of states is now all combinations of

◮ both players’ positions ◮ the card’s region ◮ the region the other player believes the card to

be in

• The set of actions now includes dialogue actions.
• Same basic reward structure as for Listenerbot,

except now also sensitive to whether the other player has found the card.

• Speech actions are modeled in terms of how they

affect the agent’s estimation of the belief state of the other agent.

47 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Relationship to RSA

48 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pursuing the ideal of Gricean pragmatics

• The cooperative principle: Make your

contribution as is required, when it is required, by the conversation in which you are engaged.

• Quality: Contribute only what you know to be true.

Do not say false things. Do not say things for which you lack evidence.

• Quantity: Make your contribution as informative as

is required. Do not say more than is required.

• Relation (Relevance): Make your contribution

relevant.

• Manner: (i) Avoid obscurity; (ii) avoid ambiguity;

(iii) be brief; (iv) be orderly.

• Politeness: Be polite, so be tactful, respectful,

generous, praising, modest, deferential, and

• sympathetic. (Leech)

49 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Pursuing the ideal of Gricean pragmatics

• The cooperative principle: Make your

contribution as is required, when it is required, by the conversation in which you are engaged.

• Quality: Contribute only what you know to be true.

Do not say false things. Do not say things for which you lack evidence.

• Quantity: Make your contribution as informative as

is required. Do not say more than is required.

• Relation (Relevance): Make your contribution

relevant.

• Manner: (i) Avoid obscurity; (ii) avoid ambiguity;

(iii) be brief; (iv) be orderly.

• Politeness: Be polite, so be tactful, respectful,

generous, praising, modest, deferential, and

• sympathetic. (Leech)

49 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Emergent pragmatics

50 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Emergent pragmatics

Quality

• Very roughly, “Be truthful”.
• For DialogBot, this emerges from the decision

problem: false information is (typically) more costly.

• DialogBot would lie if he thought it would move

them toward the objective.

50 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Emergent pragmatics

Quality

• Very roughly, “Be truthful”.
• For DialogBot, this emerges from the decision

problem: false information is (typically) more costly.

• DialogBot would lie if he thought it would move

them toward the objective.

Quantity and Relevance

• Favor informative, timely contributions.
• When DialogBot finds the card, it communicates its

location, not because it is hard-coded to do so, but rather because it will help the other agent.

50 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Grown-up DialogBots

51 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Baby DialogBots

52 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Experimental results

Agents % Success Average Moves ListenerBot & ListenerBot 84.4% 19.8 ListenerBot & DialogBot 87.2% 17.5 DialogBot & DialogBot 90.6% 16.6

T able: The evaluation for each combination of agents. 500 random initial states per agent combination.

53 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

top right top left bottom right bottom left top right left bottom middle

✱ ✱ ✱ ✱ ✱ ✱ ❭ ❭ ❭ ❭ ❭ ✚✚✚✚✚✚ ❙ ❙ ❙ ❙ ❙ ✦✦✦✦✦✦✦✦✦✦✦ ✦ ❙ ❙ ❙ ❙ ❙ ✧✧✧✧✧✧✧ ✧

54 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

top left top top right left middle right bottom left bottom bottom right

Figure: Human literal interpretations

54 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

top left top top right left middle right bottom left bottom bottom right

Figure: Human pragmatic interpretations

54 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Scalar implicature

top left top top right left middle right bottom left bottom bottom right

Figure: DialogBot interpretations

54 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Limitations

States Card location 231 × Agent location 231 × Partner location 231 × Partner’s card beliefs 231 Total ≈3 billion

• Exact solutions are out of the question.
• State-of-the-art approximate POMDP solutions can

solve problems with around 20K states.

55 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

56 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

• 1. The RSA insight L(S(L)) is a powerful tool for

achieving pragmatic language understanding.

56 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

• 1. The RSA insight L(S(L)) is a powerful tool for

achieving pragmatic language understanding.

• 2. RSA can be instantiated as a learned classifier.

56 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

• 1. The RSA insight L(S(L)) is a powerful tool for

achieving pragmatic language understanding.

• 2. RSA can be instantiated as a learned classifier.
• 3. The intractability of these models traces to the

inherent intractability of pragmatic reasoning.

56 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

• 1. The RSA insight L(S(L)) is a powerful tool for

achieving pragmatic language understanding.

• 2. RSA can be instantiated as a learned classifier.
• 3. The intractability of these models traces to the

inherent intractability of pragmatic reasoning.

• 4. Computational and cognitive considerations should

lead us to effective approximations.

56 / 56

A Gricean ideal Implicatures RSA Learned RSA Neural RSA Language and action Prospects

Conclusion and prospects

• 1. The RSA insight L(S(L)) is a powerful tool for

achieving pragmatic language understanding.

• 2. RSA can be instantiated as a learned classifier.
• 3. The intractability of these models traces to the

inherent intractability of pragmatic reasoning.

• 4. Computational and cognitive considerations should

lead us to effective approximations.

Thanks!

56 / 56