Research On Expert Finding Method For CQA Service Based On User-Tag Network

Community Question Answering(CQA)such as Quora,Yahoo! Answers and Stack Exchange provides users with a knowledge sharing platform.CQA meets the needs of users learn and share knowledge.Thus,CQA has attracted a large number of users from various industries and develops rapidly.In CQA,askers post their questions and wait for others to answer.If the questions cannot be answered in time,the askers may not to trust CQA.This will result in the loss of users and affect CQA's further development.At the same time,the user's areas of expertise and levels of professional knowledge in CQA are different.Therefore,CQA needs a way to find experts who can provide high-quality answers for the posted questions.This thesis studies the expert finding methods in CQA.The main contents include:1)In CQA,it is difficult for askers to create the most related tags for questions,resulting in that tags are too detailed or have multiple names.This thesis proposes a method for measuring the similarity of tags,and combines similar tags with the Markov Clustering Algorithm.The experiments on the CQA site Stack Exchange show that proposed method performs well.Further,this thesis discusses the relationship between users and tags,and constructs a network with user and tag as nodes.The network embedding method is applied to generate user representations.Therefore,this user representations contain tag information and structure information at the same time.2)Facing that each question has one best answer,this thesis proposes an expert finding method based on deep learning(EF-DL).First,EF-DL combines the title,body and tags of the question to construct the question text.After several data cleaning steps for the question text,EF-DL can get a sequence of words consisting of important words,and the question representations is generated by word2 vec.Then EF-DL constructs two DNNs with the same structure but different parameters to extract feature from the user representations and the question representations,and compares the cosine similarity of user feature and question feature to predict expert list.Finally,experiments on the Stack Exchange dataset show that EF-DL has better performance than other methods.3)Facing that each question has several high-quality answers provided by experts,this thesis regards the expert finding task in CQA as the expert ranking task and proposes an expert finding method based on reinforcement learn to rank(EF-RLTR).First,EF-RLTR formalizes the expert finding task in CQA as the Markov decision process,and then uses policy gradient algorithm to learn the parameters of the model.Finally,the thesis conducts extensive experiments on the well-known CQA site Stack Exchange.And experiments show that EF-RLTR can achieve better performance on expert finding tasks.