Mutual Promotion Of Question Retrieval And Answer Ranking In Community Question Answering

In community-based question answering(CQA),a user posts a query and a large amount of question-answer(Q-A)pairs are retrieved and then the best answer of the most relevant question is returned.Generally,the CQA system consists of two key subtasks:(1)question retrieval is to detect similar questions from CQA archives for query question by estimating the degree of semantic similarity of Q-Q pair,and(2)answer ranking is to examine how well an answer responses to a given question and to rank the answers according to the degree of semantic relatedness of Q-A pair and to return the best answer of given question.Since the cost of building a Q&A knowledge base is huge,one alternative workable solution is to utilize Web resources to retrieve the possible answers.The first work of this thesis focuses on constructing an Web-aided CQA system with the help of Search Engines and our system ranked the 2th in TREC 2015 LiveQA track.Previous work usually regarded question retrieval and answer ranking as two individual and independent tasks.The second work of this thesis takes the interaction between these two tasks into consideration and designs new and effective features to improve the performance of two tasks.This work has been published on IJCNN 2016(CCF-C conference).One weakness of traditional CQA system using NLP features is its poor generalization resulting from domain expert and manually-designed linguistic features.Recently,with the rapid development of deep learning and its application to NLP tasks,the third work of this thesis is to automatically learn features for CQA using deep learning and this work has been applied and published on SemEval 2016.Inspired by above work,we further investigate the deep learning based CQA system and propose a gated mechanism based neural network model to improve the performance of question retrieval and answer ranking tasks by learning the mutual information of these two tasks.In a nutshell,this thesis extensively and deeply studies the CQA system with the mutual improvement of question retrieval and answer ranking whether using traditional NLP technology or using deep learning methods.The experimental results indicate that the methods we proposed are able to improve the CQA performance.