30
It’s a Contradiction? No It’s Not! A Case Study Using Functional Relations Alan Ritter, Doug Downey, Stephen Soderland, Oren Etzioni Turing Center University of Washington 1

It’s a Contradiction? No It’s Not! A Case Study Using Functional Relations

  • Upload
    aggie

  • View
    36

  • Download
    0

Embed Size (px)

DESCRIPTION

It’s a Contradiction? No It’s Not! A Case Study Using Functional Relations. Alan Ritter, Doug Downey, Stephen Soderland , Oren Etzioni Turing Center University of Washington. Outline. Contradiction Detection AuContraire : Contradiction Detection with Functions Experiments Conclusions. - PowerPoint PPT Presentation

Citation preview

Page 1: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

1

It’s a Contradiction? No It’s Not!

A Case Study Using Functional Relations

Alan Ritter, Doug Downey, Stephen Soderland, Oren EtzioniTuring Center

University of Washington

Page 2: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

2

Outline

1. Contradiction Detection2. AuContraire:

Contradiction Detection with Functions3. Experiments4. Conclusions

Page 3: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

3

Detecting Contradictions

1.BornIn– Mozart was born in Salzburg– Mozart was born in Vienna

2.Visited– Mozart visited Salzburg– Mozart visited Vienna

BornIn is Functional, but Visited is not.

Mozart was born in Austria

Page 4: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

4

Background KnowledgeFind (T,H) which contradict with high probability

HT Background Knowledge is Key:

HTK

Not a Contradiction!

PartOf(Salzburg, Austria)

Example:• Mozart was born in Salzburg• Mozart was born in Austria

Page 5: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

5

Motivating Applications

• Political Analysis– Highlight controversial facts

• Analyze Scientific Literature– Find inconsistencies

• Fact Checker– Similar to Spell Checker, or Grammar Checker

Page 6: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

6

Related Work

• Condrovardi et. al. 2003– First proposed contradiction detection

• Harabagiu et al. 2006– First empirical results– Manually negated entailments from RTE (Recognizing

Textual Entailment)• de Marneffe et al. 2008– Annotated RTE data for contradictions– Wide variety of contradiction types– 23% precision 19% recall

Page 7: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

7

Our Contributions

1. Finding contradictions “in the wild”– RTE contains hand-selected pairs– “Balanced data” – high proportion of contradictions

2. Focus on Functional Relations– Previous work mostly dealt with “negation” and

“antonyms”3. Background Knowledge is Key– E.g. Salzburg does not conflict with Austria

Page 8: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

8

Outline

1. Contradiction Detection2. AuContraire:

Contradiction Detection with Functions1. Detecting Contradictions2. Identifying Functional Relations and Ambiguous

Arguments3. Experiments4. Conclusions

Page 9: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

9

AuContraire – A Contradiction Detection System Based on Functions

1. Begin with (Subject, Verb, Object) triples– TextRunner (Banko et. al. 2007)

2. Identify Functional Relations3. Generate “apparent contradictions”4. Sift out genuine contradictions

Page 10: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

10

How to Find Contradictions Using Functions?

• Intuition: single correct value

Invented(Basketball, <PERSON>)

<PERSON> Frequency of Extraction

James Naismith 16

Aztecs 1

Page 11: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

11

Problem: Many Seeming Contradictions

• Meronyms• Synonyms• Type mismatch• Ambiguous Arguments

We Need Background Knowledge!

Page 12: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

12

Meronyms

Example:– BornIn(Mozart, Salzburg)– BornIn(Mozart, Austria)

AuContraire’s Sources of MeronymsTipster GazetteerWordNet

Page 13: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

13

Synonyms

Example:– DiedFrom(Mozart, kidney failure)– DiedFrom(Mozart, renal failure)

AuContraire’s Sources of Synonyms:WordNetRESOLVER (Yates and Etzioni 2007)Token based string similarity (Cohen et al. 2003)

Page 14: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

14

Type Mismatch

Example:– BornIn(Mozart, Salzburg)– BornIn(Mozart, 1756)

Use a Named Entity Tagger to assign high level types• Person• Location• Date• Other

Page 15: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

15

Ambiguity

Example:– BornIn(John Smith, 1850)– BornIn(John Smith, 1737)

Later…Ambiguity is not an issue in RTE data

Page 16: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

16

Outline

1. Contradiction Detection2. AuContraire:

Contradiction Detection with Functions1. Detecting Contradictions2. Identifying Functional Relations and Ambiguous

Arguments3. Experiments4. Conclusions

Page 17: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

17

Relation Functionality1. Compute Probability of Functionality– URNS model (Downey et. al. 2005)

2. Average across all arguments

Y Frequency of Extraction

James Naismith 16

Aztecs 1

Invented(Basketball, Y)

Y Frequency of Extraction

Germany 6

Princeton 6

Berlin 5

America 3

…17

LivedIn(Einstein, Y)

Page 18: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

Argument Ambiguity

18

<PERSON> Frequency of Extraction

Brooklyn 5

Texas 5

New York 4

St. Louis 3

Pittsburgh 3

Toronto 2

Boston 2

… (15 more)

<LOCATION> Frequency of Extraction

Germany 80

Switzerland 2

BornIn(<PERSON>, <LOCATION>)

<PERSON>=Einstein <PERSON>=Smith

Page 19: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

19

Relation Functionality

Only works when unambiguous

Y Frequency of Extraction

James Naismith 16

Aztecs 1

Invented(Basketball, Y)Y Frequency of

ExtractionBouchage 8

Charles Goodyear 8

Henry Bessemer 8

Daniel Berg 8

…19

Invented(a process, Y)

Page 20: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

20

Detecting Functionality and Ambiguity• Functionality: Only use unambiguous args• Ambiguity: Only use functional relations• Expectation Maximization-like process– Alternately update two disjoint sets of parameters

EstimateFunctionali

ty

EstimateAmbiguity

Page 21: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

21

Outline1. Contradiction Detection2. AuContraire:

Contradiction Detection with Functions3. Experiments

1. Functionality/Ambiguity2. Contradictions

4. Conclusions

Page 22: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

22

Experimental Setup:Functionality and Ambiguity

• 1000 most frequent relations– Sufficient evidence to estimate functionality

• Computed functionality and ambiguity1. No EM2. EM: Converged after 5 iterations

• Hand labeled test set– Relations– A sample of arguments

Page 23: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

25

Results: Functionality and Ambiguity

19% boost in AUC

31% boost in AUC

Page 24: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

26

Experimental Setup:Contradiction Detection

• Keep top 20 relations out of 1000• Hand-Tagged a sample of contradictions– 10% rel-arg pairs

• 1.2% true contradictions (without filtering)• Use Logistic Regression to trade precision for recall– 10 fold cross validation

• Features:– Functionality/Ambiguity– String similarity between y values (Synonyms)– Etc…

Page 25: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

27

Contradiction Detection

Page 26: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

28

Error Analysis

Page 27: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

29

Future Work

• Ambiguity– Cross-Document CoReference Resolution– Focus on a specific domain (e.g. political news)

• Sparsity– Move beyond the 1000 most frequent relations

• Which sentence is correct?– How trustworthy is the source of information?

Page 28: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

30

Conclusions• AuContraire– Contradiction Detection with Functions– First work to do CD “in the wild”– Web corpus

• EM-like algorithm for detecting:– Functionality– Ambiguity

• Much external knowledge needed

Page 29: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

31

THANK YOU!

Page 30: It’s a Contradiction?  No It’s Not! A Case Study Using Functional Relations

32