Research On Question-type Sensitive Answer Summarization In Community Question Answering

With the rapid development of web interactive technologies, Community-basedQuestion Answering Portals(CQA), such as Baidu Zhidao and Yahoo!Answer, haveaccumulated a lot of question-answer pairs(QA pairs) and provided new ways tosolve open questions. However question-answer pairs in CQA are always unreliable,noisy and redundant, it is difficult to re-use these resources. Mining high-quality QApairs from CQA is an significant task in CQA. Many studies get high-qualityanswers by quality evaluation of answers and rarely take single answer incompleteproblem into account. In order to get complete,high-quality answers, this thesisproposes answer summarization algorithms for different kinds of question byconsidering the set of answers of question. This paper mainly has the followingthree aspects:(1) In order to design more effective answer summarization algorithms fordifferent types of questions, this thesis investigates question classification task inCQA. Firstly, we propose a two-layer community-oriented question taxonomy.Secondly, we analyze the difference between factoid question and CQA question，then introduce community features into question classification. Finally, we selectbest combination of features by adding feature incrementally and use active learningto enhance the perfermance of question classification.(2) We introduce topic phrase to represent answers and apply traditional topicextraction methods to answer topic phrase extraction. Then we measure coverage ofanswers, content quality of answers and the relevance between answer and questionas real-value feature and use them in answer summarization process. Finally, weanalyze the characteristics of various types of questions and present answersummarization algorithms for advice, opinion and polling questions. Experimentresults show that the proposed algorithms improve quality of answer summary,especially the coverage of answer collection, and reflect the effectiveness of usingtopic phrase describing answers.(3) This thesis implements a question-type sensitive answer summary system bycombination of question classification and answer summarization methods. In thesystem, people retrieves question then get answer summary corresponding to thequestion, this pattern greatly improves the efficiency of user’s access to information.Meanwhile, the system improves presentation style of answer summary: answers ofpolling&survey questioin presented to the user in the form of charts; answers ofopinion question presented in accordance with the type of emotion polar. Thesepresentation improvements enhance the readability of the answer summary.