Learning To Rank Answers In Domain Qusetion Answering System

Question-Answering(Q&A) System with the characteristics such as convenient, fast and efficient that could be better to meet the user retrieval needs, compared to the traditional search engines which rely on keyword matching, is more accurate to identify what was the user needed. Answers Ranking is the import issue to be solved in a Q&A system,its results directly determine the performance of the whole system.As the result,how to rank the answers efficiently for the different questions is the key problem in domain Q&A system.This paper discussed some techniques of answering ranking,dedicated to how to build efficient ranking model for different types of questions,thus improve the answer extraction results.Specifically we made a deep study in the following aspects and had gained some achievements.(1)We put forwards four types of features for questions and answers. Because of different characteristics could distinguish different contents of questions and answers, more effective use of domain knowledge, rich information of questions, question-answers, and the relationship of these evidentce to get the optimal ranking model, the features of questions and answers were divided into four categories: similarity features, translation features, density and words frequency, external knowledge features. Experimental results show that intergrate these features to ranking model can improve the ranking performance and results of answers extraction effectively.(2)We proposed a ranking model of domain entity answers based on question topics.For the characteristics of factoid and list answers in domain Q&A system,we built a ranking model of domain entity answers based on question topics.This method regards the candidate answers'documents as the bridge connecting the question topics and entity answers.Firstly,it retrieves the candidate answers' documents related to the questions,as well as the supporting answers documents corresponding to entity answers,and then integrate the relevance between supporting documents and question topic to obtain the relevance of entity answers and questions. The results demonstrate that using the ranking model leads to considerable improvments in accuracy and recall of answers.(3)We presented a ranking model based on discourse structure and Ranking-SVM.This method in view of discourse structure and the characteristics of complex answers, built a ranking model based on Ranking-SVM integrated discourse feature, converted the answers ranking to a binary classification problem. Experiments show that the method can effectively improve the quality and accuracy of the answer.(4)In order to verify the contribution of above answers ranking model to extract answers, with Yunnan tourism as the restricted domain, we designed and realized the domain Q&A testing prototype system based on the learning to rank theory, and finally analyse the performance of the system.