Paraphrase Generation from Latent-Variable PCFGs for Semantic - - PowerPoint PPT Presentation

Paraphrase Generation from Latent-Variable PCFGs for Semantic Parsing

Shashi Narayan, Siva Reddy, Shay B. Cohen

School of Informatics, University of Edinburgh

INLG, September 2016

1 / 51

Semantic Parsing for Question Answering

Semantically parsing questions into Freebase logical forms for the goal of question answering

◮ task-specific grammars (Berant et al., 2013) ◮ strongly-typed CCG grammars (Kwiatkowski et al., 2013;

Reddy et al., 2014, 2016)

◮ neural networks without requiring any grammar (Yih et al.,

2015)

2 / 51

Semantic Parsing for Question Answering

Semantically parsing questions into Freebase logical forms for the goal of question answering

◮ task-specific grammars (Berant et al., 2013) ◮ strongly-typed CCG grammars (Kwiatkowski et al., 2013;

Reddy et al., 2014, 2016)

◮ neural networks without requiring any grammar (Yih et al.,

2015) Sensitive to words used in a question and their word order Vulnerable to unseen words and phrases

2 / 51

Semantic Parsing for Question Answering: An Example

What language do people in Czech Republic speak?

3 / 51

Semantic Parsing for Question Answering: An Example

What language do people in Czech Republic speak? Freebase knowledge graph

language .human language target Czech m

Czech Republic

location.country .official language.2 location.country .official language.1

type 3 / 51

Graph Matching Problem

What language do people in Czech Republic speak?

language target people x e1 y e2 Czech Republic speak .arg2 speak .arg1 people .in.arg1 people .in.arg2 type type

Freebase knowledge graph

language .human language target Czech m

Czech Republic

location.country .official language.2 location.country .official language.1

type 4 / 51

Graph Matching Problem

What language do people in Czech Republic speak?

language target people x e1 y e2 Czech Republic speak .arg2 speak .arg1 people .in.arg1 people .in.arg2 type type

Freebase knowledge graph

language .human language target Czech m

Czech Republic

location.country .official language.2 location.country .official language.1

type 4 / 51

Graph Matching Problem with Paraphrases

What is Czech Republic’s language?

language target x e1 Czech Republic language .’s.arg1 language .’s.arg2 type

Freebase knowledge graph

language .human language target Czech m

Czech Republic

location.country .official language.2 location.country .official language.1

type 5 / 51

Graph Matching Problem with Paraphrases

What language do people speak in Czech Republic?

people y target e1 x e1 language e1 Czech Republic speak.arg2 s p e a k . a r g 1 speak.in speak.arg1 speak.arg2 speak.in type type

6 / 51

Question Answering with Paraphrases

Paraphrasing with phrase-based machine translation for text-based QA (Duboue and Chu-Carroll, 2006; Riezler et al., 2007) Paraphrasing with hand annotated grammars for KB-based QA (Berant and Liang, 2014)

7 / 51

This talk ...

Paraphrase Generation with Latent-Variable PCFGs (L-PCFGs)

8 / 51

This talk ...

Paraphrase Generation with Latent-Variable PCFGs (L-PCFGs)

◮ Uses spectral method of Narayan and Cohen (EMNLP 2015)

to learn sparse and robust grammar to sample paraphrases, and

8 / 51

This talk ...

Paraphrase Generation with Latent-Variable PCFGs (L-PCFGs)

◮ Uses spectral method of Narayan and Cohen (EMNLP 2015)

to learn sparse and robust grammar to sample paraphrases, and

◮ generates lexically and syntactically diverse paraphrases

8 / 51

This talk ...

Paraphrase Generation with Latent-Variable PCFGs (L-PCFGs)

◮ Uses spectral method of Narayan and Cohen (EMNLP 2015)

to learn sparse and robust grammar to sample paraphrases, and

◮ generates lexically and syntactically diverse paraphrases

Improving semantic parsing of questions into Freebase logical forms using paraphrases

8 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

9 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

10 / 51

Probabilistic CFGs with Latent States (Matsuzaki et al., 2005;

Prescher 2005) S VP NP N cat D the V saw NP N dog D the

⇒

S1 VP2

NP5

N4 cat D1 the V4 saw

NP3

N2 dog D1 the

Latent states play the role of nonterminal subcategorization, e.g., NP → {NP1, NP2, . . . , NP24}

◮ analogous to syntactic heads as in lexicalization (Charniak 1997)

? They are not part of the observed data in the treebank

11 / 51

Estimating PCFGs with Latent States (L-PCFGs)

EM Algorithm (Matsuzaki et al., 2005; Petrov et al., 2006) ⇓ Problems with local maxima; it fails to provide certain type

• f theoretical guarantees as it doesn’t find global maximum of

the log-likelihood

12 / 51

Estimating PCFGs with Latent States (L-PCFGs)

EM Algorithm (Matsuzaki et al., 2005; Petrov et al., 2006) ⇓ Problems with local maxima; it fails to provide certain type

• f theoretical guarantees as it doesn’t find global maximum of

the log-likelihood Spectral Algorithm (Cohen et al., 2012, 2014, Narayan and Cohen, 2015,

2016)

⇑ Statistically consistent algorithms that make use of spectral decomposition ⇑ Much faster training than the EM algorithm

12 / 51

Intuition behind the Spectral Algorithm

Inside and outside trees At node VP: S VP P him V saw NP N dog D the

Outside tree o =

S VP NP N dog D the

Inside tree t =

VP P him V saw

Conditionally independent given the label and the hidden state p(o, t|VP, h) = p(o|VP, h) × p(t|VP, h)

13 / 51

Inside Features used

Consider the VP node in the following tree:

S VP NP N dog D the V saw NP N cat D the

The inside features consist of:

◮ The pairs (VP, V) and (VP, NP) ◮ The rule VP → V NP ◮ The tree fragment (VP (V saw) NP) ◮ The tree fragment (VP V (NP D N)) ◮ The pair of head part-of-speech tag with VP: (VP, V)

14 / 51

Outside Features used

Consider the D node in the following tree:

S VP NP N dog D the V saw NP N cat D the

The outside features consist of:

◮ The pairs (D, NP) and (D, NP, VP) ◮ The pair of head part-of-speech tag with D: (D, N) ◮ The tree fragments

NP N D*

,

VP NP N D* V

and

S VP NP N D* V NP

15 / 51

Recent Advances in Spectral Estimation

= Singular value decomposition (SVD) of cross-covariance matrix for each nonterminal

16 / 51

Recent Advances in Spectral Estimation

= SVD Step

Method of moments (Cohen et al., 2012, 2014)

◮ Averaging with SVD parameters ⇒ Dense estimates

17 / 51

Recent Advances in Spectral Estimation

= SVD Step

Method of moments (Cohen et al., 2012, 2014)

◮ Averaging with SVD parameters ⇒ Dense estimates

Clustering variants (Narayan and Cohen 2015)

(1, 1, 0, 1, . . .)

S VP N w3 V w2 NP N w1 D w0 S[1] VP[3] N[1] w3 V[1] w2 NP[4] N[4] w1 D[7] w0

Sparse estimates

17 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

18 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

19 / 51

Paraphrase Generation Algorithm

Given an input sentence

◮ Word lattice construction to constrain our paraphrases to a

specific choice of words and phrases

what kind just what what exactly what what sort language linguistic do people members of the public human beings people ’s the population the citizens in Czech Republic Czech the Czech Republic Czech Cze Republic speak talking about express itself talk about to talk is speaking ?

What language do people in Czech Republic speak?

20 / 51

Paraphrase Generation Algorithm

Given an input sentence

◮ Word lattice construction to constrain our paraphrases to a

specific choice of words and phrases

◮ Sampling paraphrases using L-PCFGs, constrained by the

word lattice

20 / 51

Paraphrase Generation Algorithm

Given an input sentence

◮ Word lattice construction to constrain our paraphrases to a

specific choice of words and phrases

◮ Sampling paraphrases using L-PCFGs, constrained by the

word lattice

◮ Paraphrase classification to improve precision

20 / 51

L-PCFG Estimation for Sampling Paraphrases

The Paralex Corpus, 18m paraphrase pairs with 2.4M distinct questions (Fader et. al. 2013)

21 / 51

L-PCFG Estimation for Sampling Paraphrases

The Paralex Corpus, 18m paraphrase pairs with 2.4M distinct questions (Fader et. al. 2013) Parse all the questions using the BLLIP Parser (Charniak and Johnson, 2005) Estimate a robust and sparse L-PCFG Gsyn with m = 24 (Narayan and Cohen 2015)

21 / 51

Sampling Sentential Paraphrases using L-PCFG Gsyn

Given an input word lattice and a grammar Gsyn: Lexical pruning: Extract a grammar G ′

syn from Gsyn which is

constrained to the lattice

what kind just what what exactly what what sort language linguistic do people members of the public human beings people ’s the population the citizens in Czech Republic Czech the Czech Republic Czech Cze Republic speak talking about express itself talk about to talk is speaking ?

What language do people in Czech Republic speak?

22 / 51

Sampling Sentential Paraphrases using L-PCFG Gsyn

Given an input word lattice and a grammar Gsyn: Controlled sampling: Sample a question from G ′

syn by recursively

sampling nodes in the derivation tree, together with their latent states, over the lattice

what kind just what what exactly what what sort language linguistic do people members of the public human beings people ’s the population the citizens in Czech Republic Czech the Czech Republic Czech Cze Republic speak talking about express itself talk about to talk is speaking ?

(what, language, do, people ’s, in, Czech, Republic, is speaking, ?) ⇓ what is Czech Republic ’s language?

22 / 51

Paraphrase Classification

Filter with a classifier to improve the precision of the generated paraphrases MT metrics for paraphrase identification (Madnani et al. 2012)

23 / 51

Word Lattice Construction

Two approaches:

• 1. Lexical and phrasal paraphrase rules from the Paraphrase

Database (Ganitkevitch et al., 2013)

what kind just what what exactly what what sort language linguistic do people members of the public human beings people ’s the population the citizens in Czech Republic Czech the Czech Republic Czech Cze Republic speak talking about express itself talk about to talk is speaking ?

What language do people in Czech Republic speak?

24 / 51

Word Lattice Construction

Two approaches:

• 1. Lexical and phrasal paraphrase rules from the Paraphrase

Database (Ganitkevitch et al., 2013)

• 2. Lexical paraphrases from Bi-layered L-PCFG

24 / 51

Inducing Paraphrases with a Bi-layered L-PCFG

L-PCFG Glayered with two layers of latent states:

• ne layer is intended to capture the usual syntactic information

(traditional Gsyn with m = 24), and the other aims to capture semantic and topical information by using a large set of states (Gpar with m = 1000)

25 / 51

Training trees for Bi-layered L-PCFG Training

SBARQ-33-403 SQ-8-925 NN-41-854 nochebuena AUX-22-300 is WHNP-7-291 NN-45-142 day WP-7-254 what SBARQ-30-403 SQ-8-709 NN-41-854 nochebuena AUX-12-300 is WRB-42-707 when SBARQ-24-403 SQ-17-709 JJ-18-579 celebrated SQ-15-931 NN-30-854 nochebuena AUX-29-300 is WRB-42-707 when 26 / 51

Features for Second Layer

Design feature functions ψ and φ:

S VP NP N dog D the VP P him V saw

Outside tree o ⇒ Inside tree t ⇒

ψ(o) = [0, 1, 0, 0, . . . , 0, 1] ∈ Rd′ φ(t) = [1, 0, 0, 0, . . . , 1, 0] ∈ Rd

Bag of aligned words Bag of aligned words (the, dog, pet, . . . ) (saw, him, notice, . . . )

27 / 51

Training a Bi-layered L-PCFG

The Paralex Corpus, 18m paraphrase pairs (Fader et. al. 2013)

28 / 51

Inducing Paraphrases with a Bi-layered L-PCFG

SBARQ-33-403 SQ-8-925 NN-41-854 nochebuena AUX-22-300 is WHNP-7-291 NN-45-142 day WP-7-254 what 29 / 51

Inducing Paraphrases with a Bi-layered L-PCFG

SBARQ-33-403 SQ-8-925 NN-41-854 nochebuena AUX-22-300 is WHNP-7-291 NN-45-142 day WP-7-254 what SBARQ-30-403 SQ-8-709 NN-41-854 nochebuena AUX-12-300 is WRB-42-707 when SBARQ-24-403 SQ-17-709 JJ-18-579 celebrated SQ-15-931 NN-30-854 nochebuena AUX-29-300 is WRB-42-707 when 29 / 51

Inducing Paraphrases with a Bi-layered L-PCFG

SBARQ-33-403 SQ-8-925 NN-41-854 nochebuena AUX-22-300 is WHNP-7-291 NN-45-142 day WP-7-254 what SBARQ-30-403 SQ-8-709 NN-41-854 nochebuena AUX-12-300 is WRB-42-707 when SBARQ-24-403 SQ-17-709 JJ-18-579 celebrated SQ-15-931 NN-30-854 nochebuena AUX-29-300 is WRB-42-707 when SBARQ-1-403 SQ-22-809 VBN-29-682 born SQ-21-910 NP-24-60 NNP-21-290 montez NNP-21-567 gabuella AUX-10-866 was WRB-23-103 where 29 / 51

Inducing Paraphrases with a Bi-layered L-PCFG

SBARQ-33-403 SQ-8-925 NN-41-854 nochebuena AUX-22-300 is WHNP-7-291 NN-45-142 day WP-7-254 what SBARQ-30-403 SQ-8-709 NN-41-854 nochebuena AUX-12-300 is WRB-42-707 when SBARQ-24-403 SQ-17-709 JJ-18-579 celebrated SQ-15-931 NN-30-854 nochebuena AUX-29-300 is WRB-42-707 when SBARQ-1-403 SQ-22-809 VBN-29-682 born SQ-21-910 NP-24-60 NNP-21-290 montez NNP-21-567 gabuella AUX-10-866 was WRB-23-103 where

Inducing lexical paraphrases only

29 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

30 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

31 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak?

32 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak?

What language do people in Czech Republic speak? What is Czech Republic’s language? What language do people speak in Czech Republic? . . . 32 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak?

What language do people in Czech Republic speak? What is Czech Republic’s language? What language do people speak in Czech Republic? . . . λe.speak.arg1(e, people) ∧speak.arg2(e, language?) ∧speak.in(e, CzechRepublic)

CCG

32 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak?

What language do people in Czech Republic speak? What is Czech Republic’s language? What language do people speak in Czech Republic? . . . λe.speak.arg1(e, people) ∧speak.arg2(e, language?) ∧speak.in(e, CzechRepublic)

people y target e1 x e1 language e1 Czech Republic speak.arg2 speak.arg1 speak.in speak.arg1 s p e a k . a r g 2 speak.in type type

CCG

32 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak?

32 / 51

Semantic Parsing using Paraphrases (Reddy et. al., 2014)

What language do people in Czech Republic speak? (ˆ p, ˆ u, ˆ g) = arg max

(p,u,g)

θ · Φ(p, u, g, q, K) where, Φ(p, u, g, q, K) ∈ Rn denotes the features for the tuple of paraphrase p, ungrounded u and grounded g graphs

32 / 51

Model

Structured Perceptron: Ranks a tuple of paraphrase, grounded and ungrounded graph. (ˆ p, ˆ u, ˆ g) = arg max

(p,u,g)

θ · Φ(p, u, g, q, K) Features: Φ is defined over sentence, grounded and ungrounded graph. Training: Use surrogate gold graph to update weights θt+1 ← θt + Φ(p+, u+, g+, q, K) − Φ(ˆ p, ˆ u, ˆ g, q, K) , More details: We use Margin-Sensitive Averaged Peceptron.

33 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

34 / 51

Outline of this talk

Spectral Learning of Latent-variable PCFGs Paraphrase Generation using L-PCFGs Semantic Parsing using Paraphrases Results and Discussion

35 / 51

Experimental Setup

WebQuestions (Berant et al., 2013)

◮ Google search queries starting with wh question words ◮ 5,810 question-answer pairs (3,778 training and 2,032 test) ◮ Development experiments: held-out data consisting of

30% training questions

36 / 51

Experimental Setup

WebQuestions (Berant et al., 2013)

◮ Google search queries starting with wh question words ◮ 5,810 question-answer pairs (3,778 training and 2,032 test) ◮ Development experiments: held-out data consisting of

30% training questions Evaluation metric

◮ Average precision, average recall and average F1 (Berant et

al., 2013)

36 / 51

Experimental Setup

Our systems

◮ naive: Word lattice representing the input sentence itself ◮ ppdb: Word lattice constructed using the PPDB rules ◮ bilayered: Word lattice constructed using the bi-layered

L-PCFG

37 / 51

Experimental Setup

Our systems

◮ naive: Word lattice representing the input sentence itself ◮ ppdb: Word lattice constructed using the PPDB rules ◮ bilayered: Word lattice constructed using the bi-layered

L-PCFG Baselines

◮ original: Semantic parser (Reddy et. al., 2014) without

paraphrases

◮ mt: Monolingual machine translation based model for

paraphrase generation (Quirk et al., 2004; Wubben et al., 2010)

37 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

38 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

39 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

40 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

41 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

42 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

43 / 51

Results on the Development Set

Oracle statistics and Average F1 Scores

Method avg oracle F1 # oracle graphs avg F1

• riginal

65.1 11.0 44.7 mt 71.5 77.2 47.0 naive 71.2 53.6 47.5 ppdb 71.8 59.8 47.9 bilayered 71.6 55.0 47.1

44 / 51

Results on the Test Set

Method avg P. avg R. avg F1

• riginal

53.2 54.2 45.0 mt 48.0 56.9 47.1 naive 48.1 57.7 47.2 ppdb 48.4 58.1 47.7 bilayered 47.0 57.6 47.2

45 / 51

Results on the Test Set

Method avg P. avg R. avg F1

• riginal

53.2 54.2 45.0 mt 48.0 56.9 47.1 naive 48.1 57.7 47.2 ppdb 48.4 58.1 47.7 bilayered 47.0 57.6 47.2 Others Berant and Liang ’14 40.5 46.6 39.9 Bordes et al. ’14

• 39.2

Dong et al. ’15

• 40.8

Yao ’15 52.6 54.5 44.3 Bao et al. ’15 44.7 52.5 45.3 Bast and Haussmann ’15 49.8 60.4 49.4 Berant and Liang ’15 50.4 55.7 49.7 Yih et al. ’15 52.8 60.7 52.5

46 / 51

Results on the Test Set

Method avg P. avg R. avg F1

• riginal

53.2 54.2 45.0 mt 48.0 56.9 47.1 naive 48.1 57.7 47.2 ppdb 48.4 58.1 47.7 bilayered 47.0 57.6 47.2 Others Berant and Liang ’14 40.5 46.6 39.9 Bordes et al. ’14

• 39.2

Dong et al. ’15

• 40.8

Yao ’15 52.6 54.5 44.3 Bao et al. ’15 44.7 52.5 45.3 Bast and Haussmann ’15 49.8 60.4 49.4 Berant and Liang ’15 50.4 55.7 49.7 Yih et al. ’15 52.8 60.7 52.5

47 / 51

Results on the Test Set

Method avg P. avg R. avg F1

• riginal

53.2 54.2 45.0 mt 48.0 56.9 47.1 naive 48.1 57.7 47.2 ppdb 48.4 58.1 47.7 bilayered 47.0 57.6 47.2 Others Berant and Liang ’14 40.5 46.6 39.9 Bordes et al. ’14

• 39.2

Dong et al. ’15

• 40.8

Yao ’15 52.6 54.5 44.3 Bao et al. ’15 44.7 52.5 45.3 Bast and Haussmann ’15 49.8 60.4 49.4 Berant and Liang ’15 50.4 55.7 49.7 Yih et al. ’15 52.8 60.7 52.5

48 / 51

Results on the Test Set

Method avg P. avg R. avg F1

• riginal

53.2 54.2 45.0 mt 48.0 56.9 47.1 naive 48.1 57.7 47.2 ppdb 48.4 58.1 47.7 bilayered 47.0 57.6 47.2 Others Berant and Liang ’14 40.5 46.6 39.9 Bordes et al. ’14

• 39.2

Dong et al. ’15

• 40.8

Yao ’15 52.6 54.5 44.3 Bao et al. ’15 44.7 52.5 45.3 Bast and Haussmann ’15 49.8 60.4 49.4 Berant and Liang ’15 50.4 55.7 49.7 Yih et al. ’15 52.8 60.7 52.5

49 / 51

Error Mining

78.4% of the errors are partially correct answers occurring due to incomplete gold answer annotations or partially correct groundings 13.5% are due to bad paraphrases, and the rest 8.1% are due to wrong entity annotations

50 / 51

Conclusion

Our method is rather generic and can be applied to any question answering system Bi-layered L-PCFG for semantic similarity tasks

51 / 51