Identifying Experts in Community Question Answering Website Based on Graph Convolutional Neural Network

Abstract

High-quality answers, usually given by experts, play an important role in community question answering (CQA) websites. Therefore, experts in these websites are defined as those who provide high-quality answers. We propose two semi-supervised learning models based on the graph convolution neural network (GCN) to identify them. Both models comprehensively extract features from the social behavior network, user profiles, and question and answer text. Specifically, we construct a social behavior network according to the co-answering relationship among answerers, which means every two answerers are connected if they answer the same questions. The difference between these two models is the methods of extracting text features. One model named GCN-Doc uses Doc2vec to get text vectors before training. The other model named GCN-Lstm with a long short term memory (LSTM) network extracts text features while training. Experiments using real-world data from Zhihu.com, one of the largest Chinese CQA websites, show that both GCN-Doc and GCN-Lstm can identify experts effectively comparing with baselines of PageRank and other GCN based neural network models. Besides, GCN-Lstm performs better than GCN-Doc.