IKOMA at TAC2010: Textual Entailment System using Local-Novelty Detection

This paper reports the Recognizing Textual Entailment (RTE) system that our ‘IKOMA team’ developed at TAC 2010. We implemented a new method that identifies entailment sentences using local-novelty detection. This method determines whether a Hypothesis (H) is local-novel. ‘H is local-novel’ means that H is the new information first appeared in TH , and TH denotes the text (T) that entails H. If H is local-novel, the T that was published before TH can be recognized as no entailment. The experimental results show that this method detects T that does not entail H correctly without setting similarity threshold to be high, and raises precision as minimizing the decline of recall.