Research On Community Q&A Expert Finding Method Based On Graph Convolutional Neural Network

The rapid development of the Internet has brought great convenience to users.Users can obtain various information from the Internet,but the explosive growth of information has made it more difficult for users to acquire knowledge accurately.Generally speaking,users will search for the answers they want through search engines,but at this stage,search engines only search information through keywords,and basically do not consider the semantic information of the questions.Therefore,it often happens that users cannot find the answers they want.At the same time,with the strengthening of users’ awareness of participation,users have gradually changed from acquirers of knowledge to producers and sharers.Therefore,community Q&A websites such as Quora and Zhihu came into being.But as the number of users on community Q&A websites increases,the question asked by the questioner may not be answered for a long time or even never be answered.Therefore,it is very important for a community Q&A website to recommend experts who are capable of answering questions asked by other users.Researchers have proposed a variety of expert recommendation methods,but most of the current expert recommendation methods ignore the dynamic changes in user knowledge reserves over time and the dynamic shift of user interests.In response to the shortcomings of existing research methods,the specific work done in this thesis is as follows.(1)This thesis proposes a time-sensitive expert finding method for community Q&A websites,which constructs a multi-relational co-answer network based on co-answer and following relations.The question and answer text sequence is constructed by week.For each respondent,the method records the answer text every week.If the respondent does not participate in answering any question in a certain week,the answer text of the previous week will be used.The semantic features of these question and answer texts are extracted,and the topology of the multi-relational co-answer network is learned to capture spatial social features.The question and answer texts that change dynamically over time are learned to capture temporal semantic features,finally,they are classified through a fully connected layer.(2)This thesis proposes an expert recommendation method for community Q&A websites based on long-term and short-term interests.The method calculates the question feature through the pretrained language model.The language model and the recurrent neural network will be used to process the user’s historical answer text to learn the user’s short-term interest.The graph neural network in the multi-relational answer network will be used to capture the user’s long-term interest.Long-term interests and short-term interests are combined to form user feature vectors.A list of recommended experts is obtained by calculating the cosine similarity between user features and question features.(3)Finally,a prototype system for community Q&A websites is built,and the "time-sensitive expert finding method" and " long-term and short-term interests based expert recommendation model" proposed in this thesis are applied to the prototype system to verify availability and effectiveness of both methods.Through the specific application,it is shown that the methods in this thesis are effective for the expert recommendation on community Q&A websites.