Answer ranking based on language phenomenon recognition

Abstract

Answer Ranking is one of the core tasks in Question Answering, which greatly depends on the performance of answer ranking. This paper introduces an approach of answer ranking based on language phenomenon identification, that is, identifying language phenomena between a question and its answer sentence candidates, then computing entailment confidence score between the question and each candidate. Finally, an answer ranking is made according to such scores. This paper also introduces a joint model for both language phenomenon identification and entailment recognition task, in order to avoid error propagation to some extent, and make the two tasks learn to each other for a better overall performance as well. Experimental results show that the joint learning of language phenomenon identification and entailment recognition is an effective way for answer ranking.