Utilizing Co-Occurrence of Answers in Question Answering

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 1169–1176, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Utilizing Co-Occurrence of Answers in Question Answering Abstract In this paper, we discuss how to utilize the co-occurrence of answers in building an automatic question answering system that answers a series of questions on a specific topic in a batch mode. Experi- ments show that the answers to the many of the questions in the series usually have a high degree of co-occurrence in rele- vant document passages. This feature sometimes can’t be easily utilized in an automatic QA system which processes questions independently. However it can be utilized in a QA system that processes questions in a batch mode. We have used our pervious TREC QA system as base- line and augmented it with new answer clustering and co-occurrence maximiza- tion components to build the batch QA system. The experiment results show that the QA system running under the batch mode get significant performance im- provement over our baseline TREC QA system. 1 Introduction Question answering of a series of questions on one topic has gained more and more research interest in the recent years. The current TREC QA test set contains factoid and list questions grouped into different series, where each series has the target of a definition associated with it (Overview of the TREC 2004 Question Answer- ing Track, Voorhees 2005). Usually, the target is also called “topic” by QA researchers. One of the restrictions of TREC QA is that “questions within a series must be processed in order, with- out looking ahead.” That is, systems are allowed to use answers to earlier questions to help answer later questions in the same series, but can not use later questions to help answer earlier questions. This requirement models the dialogue discourse between the user and the QA system. However our experiments on interactive QA system show that some impatient QA users will throw a bunch of questions to the system and waiting for the answers returned in all. This prompted us to con- sider building a QA system which can accept as many questions as possible from users once in all and utilizing the relations between these ques- tions to help find answers. We would also like to know the performance difference between the QA system processing the question series in an order and the QA system processing the question series as a whole. We call the second type of QA system as batch QA system to avoid the ambigu- ity in the following description in this paper. What kind of relations between questions could be utilized is a key problem in building the batch QA system. By observing the test ques- tions of TREC QA, we found that the questions given under the same topic are not independent at all. Figure-1 shows a series of three questions proposed under the topic “Russian submarine Kursk Sinks” and some relevant passages to this …

Content truncated to 500 words