基于兴趣和专业度建模的CQA专家发现方法

CQA Expert Discovery Method Based on Interest and Expertise Modeling

现有问答社区专家发现方法通过学习用户解答的问题序列单向信息建模用户兴趣,忽略了用户兴趣的波动性,对于解答过较少问题的用户建模准确度将受到影响,此外,未考虑历史回答与问题的语义相关性对评估用户表现的作用。论文提出基于兴趣和专业度建模的CQA专家发现方法,首先,使用BERT4Rec学习用户近期解答的问题序列双向信息得到近期动态兴趣表示;其次,构建用户社交网络,使用DeepWalk算法学习网络结构特征,得到用户长期兴趣表示;再次,构建用户专业度评估网络,依据用户回答与问题的语义相关性及反馈信息计算权重,对相应问题进行加权,引入注意力机制,重点关注用户在与新问题相近问题上的表现,得到用户专业度表示;最后,综合用户近期动态兴趣、长期兴趣和专业度表示与新问题进行匹配打分,为新问题找出有意愿接受邀请并能提供优质回答的用户。实验表明,该方法取得了较好表现,较基线方法在英语、3D打印和天涯问答数据集的MRR评价指标上分别提升了5.2%、2.7%、16.1%。

The existing question answering community expert discovery methods model user interest by learning the one-way information of the question sequence answered by users,ignoring the volatility of user interest,which will affect the accuracy of modeling for users who have answered fewer questions.In addition,the role of semantic relevance of historical answers and questions in evaluating user performance is not considered.Therefore,in this research a CQA expert discovery method based on interest and expertise modeling is proposed.First,BERT4Rec is used to learn the two-way information of the recent question sequence answered by users to obtain the recent dynamic interest representation.Secondly,this research builds a user social network,and gets the long-term interest expression of users using DeepWalk algorithm to learn the network structure characteristics.Then,the user professionalism evaluation network is constructed,and weighting corresponding questions is calculated according to the semantic correlation between user answers and questions and feedback information.The attention mechanism is also introduced to focus on the user’s performance on issues similar to the new questions,and the user professionalism is expressed.Finally,the user’s recent dynamic interest,long-term interest and professional expression are combined to match with new questions for scoring,so as to identify users who are willing to accept the invitation and can provide high-quality answers to new questions.The experiment shows that this method has achieved good performance:compared with the baseline method,the MRR evaluation indexes of English,3dprinting and Tianya Q&A datasets are improved by 5.2%,2.7%and 16.1%respectively.