Entity Clustering Across Languages NAACL 2012 Montreal Spence - - PowerPoint PPT Presentation

▶
entity clustering across languages
SMART_READER_LITE
LIVE PREVIEW

Entity Clustering Across Languages NAACL 2012 Montreal Spence - - PowerPoint PPT Presentation

Feb 28, 2024 33 likes •533 views

Entity Clustering Across Languages NAACL 2012 Montreal Spence Green* Nicholas Andrews # Matthew R. Gormley # Mark Dredze # Christopher D. Manning* *Stanford University # Johns Hopkins University One Entity, Many Names Qaddafi, Muammar

slide-1
SLIDE 1

Entity Clustering Across Languages

NAACL 2012 — Montreal Spence Green* Nicholas Andrews# Matthew R. Gormley# Mark Dredze# Christopher D. Manning* *Stanford University

#Johns Hopkins University

slide-2
SLIDE 2
slide-3
SLIDE 3

One Entity, Many Names

Qaddafi, Muammar Al-Gathafi, Muammar al-Qadhafi, Muammar Al Qathafi, Mu’ammar Al Qathafi, Muammar El Gaddafi, Moamar El Kadhafi, Moammar El Kazzafi, Moamer El Qathafi, Mu’Ammar

3 / 41

slide-4
SLIDE 4

One Entity, Many Names

Qaddafi, Muammar Al-Gathafi, Muammar al-Qadhafi, Muammar Al Qathafi, Mu’ammar Al Qathafi, Muammar El Gaddafi, Moamar El Kadhafi, Moammar El Kazzafi, Moamer El Qathafi, Mu’Ammar

  • 3 / 41
slide-5
SLIDE 5

One Entity, Many Names

Qaddafi, Muammar Al-Gathafi, Muammar al-Qadhafi, Muammar Al Qathafi, Mu’ammar Al Qathafi, Muammar El Gaddafi, Moamar El Kadhafi, Moammar El Kazzafi, Moamer El Qathafi, Mu’Ammar

  • 3 / 41
slide-6
SLIDE 6

Basic Task: Entity Clustering

Cluster co-referent entity mentions across a corpus (documents and languages)

4 / 41

slide-7
SLIDE 7

Basic Task: Entity Clustering

Cluster co-referent entity mentions across a corpus (documents and languages) Clustering/disambiguation relies on:

◮ Mention similarity ◮ Context similarity

4 / 41

slide-8
SLIDE 8

Entity Disambiguation: Mention Similarity

Qaddafi, Muammar Al-Gathafi, Muammar al-Qadhafi, Muammar Al Qathafi, Mu’ammar Al Qathafi, Muammar El Gaddafi, Moamar El Kadhafi, Moammar El Kazzafi, Moamer El Qathafi, Mu’Ammar

  • 5 / 41
slide-9
SLIDE 9

Entity Disambiguation: Context Similarity Apple

6 / 41

slide-10
SLIDE 10

Entity Disambiguation: Context Similarity Apple

  • Apple Inc.
  • Apple Inc.
  • Apple Inc.
  • town in Lebanon
  • camel

6 / 41

slide-11
SLIDE 11

Entity Disambiguation: Context Similarity

The Apple chief executive was former Beatles road manager Neil Aspinall... Sentential context is usually required

7 / 41

slide-12
SLIDE 12

Old: Entity Clustering Tasks

Within-doc coref

Peter said to himself ...

8 / 41

slide-13
SLIDE 13

Old: Entity Clustering Tasks

Within-doc coref Cross-doc coref

doc1: Peter Jones said ... doc2: I told Mr. Jones ...

[Bagga and Baldwin, 1998] [Baron and Freedman, 2008]

Peter said to himself ...

9 / 41

slide-14
SLIDE 14

Old: Entity Clustering Tasks

Within-doc coref Cross-doc coref Entity Linking

doc1: Peter Jones said ... doc2: I told Mr. Jones ...

[Bagga and Baldwin, 1998] [Baron and Freedman, 2008] [McNamee et al., 2011]

Peter said to himself ... Peter Jones Peter

  • Mr. Jones

[Rao et al., 2011]

10 / 41

slide-15
SLIDE 15

New: Entity Clustering Across Languages

Within-doc coref Cross-doc coref This paper Entity Linking

doc1: Peter Jones said ... doc2: I told Mr. Jones ...

[Bagga and Baldwin, 1998] [Baron and Freedman, 2008] [McNamee et al., 2011]

Peter said to himself ... Peter Jones Peter

  • Mr. Jones

doc1: Peter Jones said ... doc2:

  • [Rao et al., 2011]

11 / 41

slide-16
SLIDE 16

New: Entity Clustering Across Languages

The Apple chief executive was former Beatles road manager Neil Aspinall...

  • No knowledge base of entities

12 / 41

slide-17
SLIDE 17

Why?

Crisis management Arab Spring (2011)

◮ French, Arabic dialects ◮ Facebook, Twitter, blog...

Haiti earthquake (2010)

◮ Kreyol, English, French ◮ SMS, Twitter, blog...

13 / 41

slide-18
SLIDE 18

Why?

Crisis management Arab Spring (2011)

◮ French, Arabic dialects ◮ Facebook, Twitter, blog...

Haiti earthquake (2010)

◮ Kreyol, English, French ◮ SMS, Twitter, blog...

Other applications: machine translation, name search

13 / 41

slide-19
SLIDE 19

Plan: Extend Existing Monolingual System

English Arabic English+Arabic 70 80 90 100 91.4 89.8 78.8

B3

14 / 41

slide-20
SLIDE 20

Cross-lingual Mention Similarity Cross-lingual Context Similarity Clustering Algorithms Evaluation

slide-21
SLIDE 21

Cross-lingual Mention Similarity Cross-lingual Context Similarity Clustering Algorithms Evaluation

slide-22
SLIDE 22

Within-language: Edit Distance

mi = Muammar Qaddafi mj = Moamer El Kazzafi sim(mi, mj) =?

17 / 41

slide-23
SLIDE 23

Within-language: Edit Distance

mi = Muammar Qaddafi mj = Moamer El Kazzafi sim(mi, mj) =? Algorithm: Sorted Jaro-Winkler

[Christen, 2006]

  • 1. Sort tokens
  • 2. Compute Jaro-Winkler distance in O(|mi| + |mj|)
  • 3. Evaluate sim(mi, mj) < β

17 / 41

slide-24
SLIDE 24

Cross-language: Binary classifier

mi = map(Apple) = abbl mj = map( ) = abl sim(mi, mj) =?

18 / 41

slide-25
SLIDE 25

Cross-language: Binary classifier

mi = map(Apple) = abbl mj = map( ) = abl sim(mi, mj) =? Algorithm: Phonetic mapping + classification

  • 1. Apply deterministic mapping map(·)
  • 2. Extract character-level features
  • 3. Classify (Maxent)

18 / 41

slide-26
SLIDE 26

Cross-language: Binary classifier

Training: parallel name list

◮ 97.1% accuracy on a held-out set

Phonetic mapping: think of Soundex Best features: character bigrams

19 / 41

slide-27
SLIDE 27

Cross-lingual Mention Similarity Cross-lingual Context Similarity Clustering Algorithms Evaluation

slide-28
SLIDE 28

Context Similarity: Mapping techniques

Method Resource-level Machine Translation High Lexicon Medium Polylingual Topic Model Low

21 / 41

slide-29
SLIDE 29

Context Similarity: Mapping techniques

Method Resource-level Machine Translation High Lexicon Medium Polylingual Topic Model Low

◮ MT: Phrasal with NIST-09 data

[Galley et al., 2009]

◮ Lexicon: 31k entries (web and LDC sources)

21 / 41

slide-30
SLIDE 30

Context Similarity: Polylingual Topic Model

Polylingual Topic Model (PLTM)

[Mimno et al., 2009]

◮ Words linked through cross-lingual priors ◮ Training: Wikipedia document tuples

Map context words to 1-best topics

22 / 41

slide-31
SLIDE 31

Context Similarity: Polylingual Topic Model

[Mimno et al., 2009]

23 / 41

slide-32
SLIDE 32

Context Similarity: Polylingual Topic Model

The Apple chief executive was former Beatles road manager Neil Aspinall...

  • 3 1 6 6 14 5 103 99 6 3 5...

2 1 14 99 7 7 103 79

24 / 41

slide-33
SLIDE 33

Context Similarity

Bag of words / smoothed unigram distributions Measure: Jensen-Shannon divergence

25 / 41

slide-34
SLIDE 34

Cross-lingual Mention Similarity Cross-lingual Context Similarity Clustering Algorithms Evaluation

slide-35
SLIDE 35

Constraint-Based Clustering

Two algorithms:

  • 1. Hierarchical clustering
  • 2. Dirichlet process mixture model

Setup:

◮ Mention similarity as a hard constraint ◮ Cluster distance: context similarity

27 / 41

slide-36
SLIDE 36

Constraint-Based Clustering

Muammar Qaddafi El Kazzafi Apple Apple Corps.

context context context

al-Qadhafi

28 / 41

slide-37
SLIDE 37

Constraint-Based Clustering

Muammar Qaddafi El Kazzafi Apple Apple Corps.

context context context

al-Qadhafi

0.40 0.31 0.15

29 / 41

slide-38
SLIDE 38

Constraint-Based Clustering

Muammar Qaddafi El Kazzafi Apple Apple Corps.

context context context

al-Qadhafi

0.31

30 / 41

slide-39
SLIDE 39

Cross-lingual Mention Similarity Cross-lingual Context Similarity Clustering Algorithms Evaluation

slide-40
SLIDE 40

Evaluation Corpus

Automatic Content Extraction (ACE) 2008 Arabic-English We annotated 216 cross-lingual entities Genres:

  • 1. broadcast

conversation

  • 2. broadcast news
  • 3. meeting
  • 4. newswire
  • 5. telephone
  • 6. usenet
  • 7. weblog

32 / 41

slide-41
SLIDE 41

ACE2008 Evaluation Corpus

Docs Tokens Entities Chains Mentions Arabic 412 178k 2.6k 4.2k 9.2k English 414 246k 2.3k 4.0k 9.1k

◮ Chain – set of mentions (within-doc) ◮ Entity – set of chains

33 / 41

slide-42
SLIDE 42

Within-document Processing

Our models cluster chains Evaluation:

◮ Gold chains ◮ Predicted chains from SERIF

[Ramshaw et al., 2011]

34 / 41

slide-43
SLIDE 43

Evaluation: Gold within-document processing

B3 B3 (cross-lingual only) 50 60 70 80 90 100 85.4 78.8 80.4 66.4 77.3 58.4 70.1 54.5 MT Lexicon PLTM Baseline

35 / 41

slide-44
SLIDE 44

Evaluation: Gold within-document processing

B3 B3 (cross-lingual only) 50 60 70 80 90 100 85.4 78.8 80.4 66.4 77.3 58.4 70.1 54.5 MT Lexicon PLTM Baseline

In paper: CEAF, NVI

35 / 41

slide-45
SLIDE 45

Automatic within-document processing

B3 65.0 70.0 75.0 80.0 76.7 76.0 75.3 67.0 MT Lexicon PLTM Baseline

36 / 41

slide-46
SLIDE 46

Lexicon Clustering

Correct tony blair

=

  • twny blyr

khaled mashaal

=

  • khld mshal

Incorrect NSA

=

CIA

(En and Ar mentions)

republican party

=

  • hzb jwmhwry

37 / 41

slide-47
SLIDE 47

MT Clustering

Correct NSA

=

CIA labour party

=

  • hzb aml

Incorrect hamed bin khalifa al- thani

=

  • hmd bn

khlfa thny

38 / 41

slide-48
SLIDE 48

Conclusion

Within-doc coref Cross-doc coref This paper Entity Linking

doc1: Peter Jones said ... doc2: I told Mr. Jones ...

[Bagga and Baldwin, 1998] [Baron and Freedman, 2008] [McNamee et al., 2011]

Peter said to himself ... Peter Jones Peter

  • Mr. Jones

doc1: Peter Jones said ... doc2:

  • [Rao et al., 2011]

39 / 41

slide-49
SLIDE 49

Code and corpus: spencegreen.com

thanks.

slide-50
SLIDE 50

Future Work

Pairwise models don’t scale

◮ See [Rao et al., 2010] and [Singh et al., 2011]

Model at mention level

◮ See Nick Andrews’ talk at EMNLP!

Unified similarity measure

◮ Logistic regression did not generalize

41 / 41