slides

Two Related Lexico-Syntactic
Approaches to Entailment
Vasile Rus
Institute for Intelligent Systems
Department of Computer Science
http://www.cs.memphis.edu/~vrus
TODAY- Outline
• General strategy
– Map T and H into lexico-syntactic graphs
– Perform graph subsumption between graph-T and
graph-H
– Additive strategy
• Not cascaded
• Two approaches
– Lexico-syntactic approach
– Dependency-based approach
• Results, Comparison, Conclusions
The Two Approaches
• Lexico-syntactic approach
–
–
–
–
–
Lexical component
Syntactic component
Dependencies derived from phrase-based parse trees
Negation
thesaurus
• Dependency-based approach
– Dependencies from MINIPAR
– Lexical component by default
– Postprocessing (thanks to Vivi Nastase)
• To eliminate unused information
• To retain only dependencies among content words
Graph Subsumption
• Map nodes and edges in H-graph to nodes
and edges in T-graph
• complex mapping based on
– Named Entity Inferences: Overture Services
Inc -- Overture
– Word-level entailment / equivalence: take over
– buy
– Syntactic Info:
• Yahoo is the agent of buying
From Sentences to Graph
Representation
• vertices represent content words
• edges represent dependencies
– local dependencies (intra-phrase) are
straightforwardly obtained from a parse tree
– remote dependencies are obtained using an
extended functional tagger
– Or from MINIPAR (for the second approach)
The Entailment Score
•The score is so defined to be non-reflexive:
entail(T, H) ≠ entail(H,T)
Score is also used as confidence
The Parameters
• the following parameters worked best on
development
α=.5 β =.5 γ=0
Negation
• Explicit
– Clue phrases
• no, not, neither … nor
• shortened forms: ‘nt
• Implicit
– Antonymy in WordNet
• Hypothetical sentences:
“a possible visit by Clinton to China”
does not entail
“Clinton visited China”
– a form of negation
Results – Lexico-Syntactic
Approach
System
Accuracy Average precision
Lexico-syntactic
0.5900
0.6047
Lex
0.5663
0.5823
Lex-cnt-words
0.5875
0.5725
Lex+syn
0.5737
0.5841
Lex+syn+neg
0.5800
0.6096
Lex+syn+synt
0.5813
0.5941
lex+syn+synt+neg 0.5900
0.6047
Comparison
System
Accuracy Average precision
Lexico-syntactic
0.5900
0.6047
Lex+syn+synt
0.5813
0.5941
Dependency-based
0.5837
0.5785
Conclusions
• Lexical information significantly helps
• The other components (synonymy,
dependencies, negation) add value but not
significantly
Missed Opportunities
• Linguistic Level
– Five = 5
– Tuscany province = province of Tuscany
• Current subsumption algorithm is weak
• T: Besancon is the capital of France’s watch and clockmaking industry and of high precision engineering.
• H: Besancon is the capital of France.
Solution: matching with more complex
structures
• World Knowledge
More Conclusions
• Our system is light
– Good for interactive environment such as
Intelligent Tutoring Systems
• No training involved
– Just development to tune few parameters
One More Conclusion
• It is not clear whether there is a difference
among the two ways to obtain
dependencies!
Two Related Lexico-Syntactic
Approaches to Entailment
Thank you everyone !