ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch,...

15
ISCOL 2011 – Bar Ilan University / 151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University

Transcript of ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch,...

Page 1: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /151

A Probabilistic Model for Lexical Entailment

Eyal Shnarch, Jacob Goldberger, Ido Dagan

Bar Ilan University

Page 2: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /152

Textual Entailment is a common task

Obama gave a speech last night in

the Israeli lobby conference ...

Obama gave a speech last night in

the Israeli lobby conference ...

In his speech at the American Israel Public

Affairs Committee yesterday, the president

challenged …

In his speech at the American Israel Public

Affairs Committee yesterday, the president

challenged … Barack Obama’s AIPAC

address ...Barack Obama’s AIPAC

address ...AIPAC

Israeli lobby

American Israel Public Affairs Committee

address

speech

Barack Obama the president

Obama

Page 3: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /153

Textual Entailment

AIPAC Israeli lobby speech address

Page 4: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /154

The president’s car got stuck in Ireland, surrounded by many peopleThe president’s car got stuck in Ireland, surrounded by many people

Obama’s Cadillac got stuck in Dublin in a large Irish crowdObama’s Cadillac got stuck in Dublin in a large Irish crowd

social group

social group

Modeling entailment at the lexical level

Page 5: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /155

rule2

rule1

The president’s car got stuck in Ireland, surrounded by many peopleThe president’s car got stuck in Ireland, surrounded by many people

Obama’s Cadillac got stuck in Dublin in a large Irish crowdObama’s Cadillac got stuck in Dublin in a large Irish crowd

social group

social group

Terminology

rule

lexical resource

chain

Page 6: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /156

The president’s car got stuck in Ireland, surrounded by many peopleThe president’s car got stuck in Ireland, surrounded by many people

Obama’s Cadillac got stuck in Dublin in a large Irish crowdObama’s Cadillac got stuck in Dublin in a large Irish crowd

social group

social group

Goals

p( ) p( ) p( ) Distinguish resources’ reliability levels

Consider transitive chains length

Consider multiple evidence

Page 7: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /157

Probabilistic model for Lexical Entailment

t1 tmti

h1 hnhj

t’

AND

y

OR

chain

… …

……

validity probability of the resource which produces r

(ACL 2011 short paper)

Page 8: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /158

Results on RTE are nice, but…

F1 %Model

RTE 6 RTE 5

33.8 30.5 Avg. of all systems

38.5 36.2 Base Prob.

47.6 44.4 Best lexical system

48.0 45.6 Best full system

Page 9: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /159

Extension 1: relaxing with noisy-AND

Page 10: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1510

Better results on RTE with extension 1

F1 %Model

RTE 6 RTE 5

33.8 30.5 Avg. of all systems

38.5 36.2 Base Prob.

43.1 44.6 Base Prob. + noisy-AND

47.6 44.4 Best lexical system

48.0 45.6 Best full system

Page 11: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1511

Extension 2: considering coverage

Page 12: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1512

Same (better) results on RTE with extension 2

F1 %Model

RTE 6 RTE 5

33.8 30.5 Avg. of all systems

38.5 36.2 Base Prob.

43.1 44.6 Base Prob. + noisy-AND

44.7 42.8 Base Prob. + coverage normalization

47.6 44.4 Best lexical system

48.0 45.6 Best full system

Page 13: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1513

Putting it all together is best

F1 %Model

RTE 6 RTE 5

33.8 30.5 Avg. of all systems

38.5 36.2 Base Prob.

43.1 44.6 Base Prob. + noisy-AND

44.7 42.8 Base Prob. + coverage normalization

45.6 48.3 Full Prob. model (noisy-AND + coverage norm)

47.6 44.4 Best lexical system

48.0 45.6 Best full system

Negative result: F1 usually decreases when allowing chains

Page 14: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1514

Future work

• Better model for transitivity

• noisy-AND for chains too

• Verify rule application in a specific context

• Test with other application data sets• passage retrieval for QA

• Integrate into a full entailment system

Page 15: ISCOL 2011 – Bar Ilan University /151 A Probabilistic Model for Lexical Entailment Eyal Shnarch, Jacob Goldberger, Ido Dagan Bar Ilan University.

ISCOL 2011 – Bar Ilan University /1515

Summary

• Learn for each lexical resource an individual

reliability value

• Consider multiple evidence and chain length

• Probabilistic method to relax the strict AND

demand

• Taking into account the number of covered

terms when modeling entailment probability

A first probabilistic model:

noisy-