UNED at PASCAL RTE-2 Challenge

1
UNED at PASCAL RTE-2 Challenge

• IRNLP Group at UNED
• nlp.uned.es
• Jesús Herrera
• Anselmo Peñas
• Álvaro Rodrigo
• Felisa Verdejo

2
Outline

• Underlying thesis
• Overall architecture
• Lexical Entailment
• Entailment decision
• Results
• Conclusion future Work

3
Underlying Thesis

• Compositional meaning of the statements
• All components of the hypothesis
• must be entailed
• Definition of statement component
• Our Roadmap
• Word multiword level
• NE, numeric temporal expressions level
• Phrase level paraphrase (in progress)

4
Overall Architecture
5
WN-based Lexical Entailment

• WordNet based entailment
• Synonymy
• Hyponymy (Herrera et al. 2005)
• Antonymy
• Verb Entailment
• Part meronymy Italy entails Europe
• Pertainymy Italian entails Italy
• (Glosses)
• Search of entailment paths
• Depth
• Breadth
• Path Length

6
WN-based Lexical Entailment
7
Numerical Entailment
Text
Hypothesis
17 million citizens
More than 15 million people
recognize
8
Numerical Entailment
9
Entailment Decision

• Support Vector Machines
• of hypothesis lemmas entailed
• of hypothesis words coincidence
• Existence of numeric expressions not entailed in
  the hypothesis
• Existence of NE not entailed in the hypothesis
  (after submission)

10
Results
Run 1 Run 2 (word coincidence) (NE entailment)
IE 49.0 52.0 51.0
QA 56.5 52.0 53.5
IR 64.5 57.0 63.5
SUM 69.0 74.5 74.0
Overall 59.75 58.87 60.5
11
Conclusion Future Work

• WN is nice but doesnt help too much with an
  indiscriminate use (Check by source task)
• Different processing for each source task
• NE numerical entailment seem promising
• Pay more attention to the entailment decision
• Future
• Phrase level entailment
• Paraphrase