Answering Contextual Questions Based on the Cohesion with the Knowledge

In this paper, we introduce the notion “the cohe-sion with the knowledge”, and, based on it, propose a question answering system to answer contextual ques-tions using a non-contextual QA system. The con-textual questions usually have some cohesive relation to their context like reference expressions. Therefore, systems have to detect the cohesion and resolve the reference relations to answer the questions. Previous works usually address this problem in terms of the co-hesion with the context. On the other hand, we address the problem by using the notion “the cohesion with the knowledge.” First of all, the proposed method detects reference expressions in a given question. Second, it generates all possible completed question candidates by gather-ing antecedent candidates corresponding to reference expressions using I) the selectional restriction based on a case-frame dictionary and a thesaurus, or II) a modified version of the centering theory and I). As the degree of cohesion with the knowledge for each question candidate, we adopt the score of answer for each question candidate produced by a non-contextual QA system. The experimental results show that Strat-egy I is effective to improve the accuracy of answering the question series of the gathering type, on the other hand, Strategy II is effective for the question series of the browsing type