Research On Answer Summarization With Question Entity Expansion And Global Inference

With the popularity of the Internet, the need for people to seek help is becoming urgent. CQA service such as Yahoo ! Answer and Baidu Zhidao is becoming more important. Compared to traditional Question & Answering System and FAQ System, CQA system has a larger scale and more categories, greatly improves the resources of Internet, brings convenience to people for seeking information. But each answers in CQA may be incomplete or irrelevant to questions,this causes the quality degradation of question and answers pairs,re DUCing users’ satisfaction. Answer summarization is a solution to this problem. By summarizing all the sentences of the question,we can get a complete and relevant summary. This paper goes from question,takes use of knowledge graph to do entity expansion,finds all the sentences relevant to the question and finally gets the summary by mixed integer linear programming algorithm。This paper includes three aspects as follows:1. Research on question entity expansion based on knowledge. It’s difficult to predict answer’s content just based on question. Our intuition is that knowledge already has huge amounts of knowledge, we firstly find question’s entities, then expand question’s entities based on entities’ relations which may appear in the true answer. We estimate the weight of these entities. As a result, the expanded entities can accurately predict answer’s entities, this tells the expanded is question’s content.2. Research on the algorithm to get a complete, relevant summary by using the expanded entities. We use sentence compression and sentence filtering to remove the poor information sentence. For the remained sentences, we only consider the expanded entities, and build an object function. By maximizing the objective function, a readable summary is obtained. The sentence quality and missing entity weight estimation is used to optimize the algorithm.3. We built a CQA answer summarization system, which also contains question retrieval. The question retrieval module can get the closest question to user typed. Then the system display best answer and answer summarization together.