COGEX at RTE3

Abstract

This paper reports on LCC’s participation at the Third PASCAL Recognizing Textual Entailment Challenge. First, we summarize our semantic logical-based approach which proved successful in the previous two challenges. Then we highlight this year’s innovations which contributed to an overall accuracy of 72.25% for the RTE 3 test data. The novelties include new resources, such as eXtended WordNet KB which provides a large number of world knowledge axioms, event and temporal information provided by the TARSQI toolkit, logic form representations of events, negation, coreference and context, and new improvements of lexical chain axiom generation. Finally, the system’s performance and error analysis are discussed.