The Study On Expert Recommendation In Community Question Answering

The rapid development of the Internet has changed the way people communicate,more and more people rely on their Internet community to obtain information and consult expertise,community question answering(CQA)websites have gained widespread popularity among the public.With a growing number of questions have been answered,the community question answering sites have created a large-scale knowledge repository that is free to acquire knowledge.Which can meet the needs of the questioners,and provide valuable information for the majority of the community at the meantime.High-quality answers not only meet the needs of current questioners,but also continue to produce value.This thesis enhances the quality and effectiveness of the CQA system by recommending the new questions to the appropriate experts to raise the quality of the answers.In the case of a single domain expert recommendation,this thesis constructs a recurrent neural network classification model.The best answerer of the question as a positive example,the other people as negative cases,and then a variety of recurrent neural network models are used to predict the expert users,and the introduction of attention mechanism,by weighting the text feature to enhance the weight of more important features for the classification.Experimental results shows the effectiveness of the recurrent neural network for expert recommendation,and the bidirectional recurrent neural network with attention mechanism shows unique advantages.When performing cross domain expert recommendation,two methods are used,including the similarity sorting and text classification methods.Co-occurrence information of the user mentions with the question words in the same context is assumed to be evidence of expertise.However,the question in CQA usually too short to get enough information for dealing with the word-matching between posted questions and users' profile.That is,there are semantic gap between questions and users' profile,in this thesis,distributed representation is used to tackle this problem.Experimental results show that the distributed representation can capture meaningful syntactic and semantic information,which can improve the performance of the system,and the convolutional neural network has achieved good results.We have performed experiments in real-world datasets,both the specific area and multidisciplinary areas of Stack Overflow,the experimental results show that the effectiveness of the proposed method.