Short text classification is an important natural language processing task due to the prevalence of short text on the internet and social media platforms.In this paper, we propose a novel graph-based short text classification method named GBBM (Graph-BERT-BTM Model) that leverages the powerful representation ability of graph data to capture the structural features of short text.In this work, we incorporate topic information to enrich and expand the feature space for the short text and compare our proposed method on five publicly available short text datasets with five existing models.Experimental results indicate the superiority of our proposed method.

As for standard topic model, such as LDA (Latent Dirichlet Allocation), each topic is generally depicted by a weighted word set, where the high-ranked words are deemed more representative.Meanwhile, the probability of each word is considered as the ability to represent the semantic contribution for the topic.However, few efforts are focused on enhancing the representative ability of the topic to support fine grained topic representation.In this paper, we propose a Word Topic Ware (WTW) model to take word inherent diversity characteristic into consideration, in order to screen out and enhance the more representative words for topic representation.Experimental results on three large datasets show that our proposed method can increase the representative ability for topic representation.In addition, our work will positively affect improving the quality of topic content analysis.