A novel topic clustering algorithm based on graph neural network for question topic diversity

In community question answering, many questions have no topic labeling or the topic labeling is very diverse, which has become the biggest obstacle to building the bridge between users and posts. Topic clustering methods could alleviate this issue. However, existing research employed words as topic representation units and could not express topic semantic relevance. In this paper, we propose a novel Topic Clustering framework based on the Graph Neural Network (called TCGNN) to alleviate topic diversity in Community Question Answering. Firstly, we separately consider the relationship representation of existing topics and unlabeled topics. For manually labeled topics, we count the frequency of topics in community questions and construct a topic cooccurrence matrix to represent the topic relation. For unmarked topics, we extract the core phrases from community questions and employ them to indicate the topics of questions. Then, we transform the topic co-occurrence matrix into a topic relation graph, optimizing the topic relevance and improving presentation efficiency. Next, we employ a graph neural network for embedding the topic connection graph and get the vector representation of each topic. Finally, an improved K-mean method is proposed for topic clustering based on the distance of topic vectors. Additionally, we briefly discuss the extended effect of topic clustering methods in other domains (bibliographic information and reviews). In the literature we have, it is a primary work that conders topic clustering in multiple situations and offers innovative cogitation to apply graph neural networks in topic clustering. Our experiment compared prevalent clustering methods and some combination methods of text representation and graph embedding. The outcome of experiments on four extensive and varied datasets (Stack Overflow, DBLP, Yelp, and Zhihu) illustrate that TCGNN leads the prevalent baseline in Entropy and Purity.