a compositional and interpretable semantic space alona fyshe, leila wehbe, partha talukdar, brian...

40
A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University [email protected] 1

Upload: jasmin-powell

Post on 21-Dec-2015

220 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

1

A Compositional and Interpretable Semantic Space

Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell

Carnegie Mellon University

[email protected]

Page 2: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

2

pear

lettuce

orange

apple

carrots

VSMs and Composition

Page 3: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

How to Make a VSM

CountDim.

ReductionCorpus

Statistics

VSM

3

Many cols Few cols

Page 4: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

4

pear

lettuce

orange

apple

carrots

seedless orange

VSMs and Composition

Page 5: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

5

VSMs and Composition

f( , )

=adjective noun estimate

observed

Stats for seedless Stats for orange

Observed stats for “seedless orange”

Page 6: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

6

Previous Work

• What is “f”?(Mitchell & Lapata, 2010; Baroni and Zamparelli, 2010; Blacoe and Lapata, 2012; Socher et al., 2012; Dinu et al., 2013; Hermann & Blunsom, 2013)

• Which VSMs are best for composition?(Turney, 2012, 2013; Fyshe et al., 2013; Baroni et al., 2014)

Page 7: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

7

Our Contributions

• Can we learn a VSM that – is aware of composition function?– is interpretable?

FFIs

edib

le

Page 8: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

How to make a VSM

• Corpus– 16 billion words– 50 million documents

• Count dependencies arcs in sentences• MALT dependency parser

• Point-wise Positive Mutual Information

8

Page 9: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Matrix Factorization in VSMs

X A

D

Corpus Stats (c)

Words

9

VSM

Page 10: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Interpretability

10

A

Latent Dims

Words

Page 11: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Interpretability

11

• SVD (Fyshe 2013)– well, long, if, year, watch – plan, engine, e, rock, very – get, no, features, music, via

• Word2vec (pretrained on Google News)– pleasantries, draft_picks, chairman_Harley_Hotchkiss,

windstorm, Vermont_Yankee– Programme_Producers_AMPTPP, ###/mt, Al_Mehwar, NCWS,

Whereas– Ubiquitous_Sensor_Networks, KTO, discussing,

Hibernia_Terra_Nova, NASDAQ_ENWV

Page 12: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Non-Negative Sparse Embeddings

12

X A

D

(Murphy 2012)

Page 13: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Interpretability

13

• SVD– well, long, if, year, watch – plan, engine, e, rock, very – get, no, features, music, via

• NNSE– inhibitor, inhibitors, antagonists, receptors,

inhibition – bristol, thames, southampton, brighton, poole – delhi, india, bombay, chennai, madras

Page 14: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

14

A Composition-aware VSM

Page 15: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

15

Modeling Composition

• Rows of X are words– Can also be phrases

X APhrases Phrases

Adjectives

Nouns

Adjectives

Nouns

Page 16: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

16

Modeling Composition

• Additional constraint for composition

APhrases

Adjectives w1w2

p

p = [w1 w2]

Nouns

Page 17: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

17

Weighted Addition

Page 18: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

18

Modeling Composition

Page 19: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

19

Modeling Composition

• Reformulate loss with square matrix B

AB

α β -1

adj. col. noun col. phrase col

Page 20: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

20

Modeling Composition

Page 21: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Optimization

• Online Dictionary Learning Algorithm(Mairal 2010)

• Solve for D with gradient descent• Solve for A with ADMM– Alternating Direction Method of Multipliers

21

Page 22: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Testing Composition

• W. add

• W. NNSE

• CNNSE

22

A

w1w2

p

SVDw1w2

p

A

w1w2

p

Page 23: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

23

Phrase Estimation

• Predict phrase vector• Sort test phrases by distance to estimate

•Rank (r/N*100)•Reciprocal rank (1/r)•Percent Perfect (δ(r==1))

r

N

Page 24: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

24

Phrase Estimation

Chance 50 ~ 0.05 1%

Page 25: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

25

Interpretable Dimensions

Page 26: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

26

Interpretability

Page 27: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Testing Interpretability

• SVD

• NNSE

• CNNSE

27

A

w1w2

p

SVDw1w2

p

A

w1w2

p

Page 28: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

28

Interpretability

• Select the word that does not belong:• crunchy• gooey• fluffy• crispy• colt• creamy

Page 29: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

29

Interpretability

Page 30: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Phrase Representations

30

A

phrase

top scoringwords/phrases

top scoringdimension

Page 31: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

31

Phrase Representations

Choose list of words/phrases most associated with target phrase “digital computers”• aesthetic, American music, architectural style• cellphones, laptops, monitors• both• neither

Page 32: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

32

Phrase Representation

Page 33: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Testing Phrase Similarity• 108 adjective-noun phrase pairs

• Human judgments of similarity [1…7]

• E.g. Important part : significant role (very similar)

Northern region : early age (not similar)

33

(Mitchell & Lapata 2010)

Page 34: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Correlation of Distances

34

Behavioral Data

Model A

Model B

Page 35: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Testing Phrase Similarity

35

Page 36: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

36

Interpretability

Page 39: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

Summary

• Composition awareness improves VSMs– Closer to behavioral measure of phrase similarity– Better phrase representations

• Interpretable dimensions– Helps to debug composition failures

39

Page 40: A Compositional and Interpretable Semantic Space Alona Fyshe, Leila Wehbe, Partha Talukdar, Brian Murphy, and Tom Mitchell Carnegie Mellon University amfyshe@gmail.com

40

Thanks!

www.cs.cmu.edu/~fmri/papers/naacl2015/

[email protected]