The Endeavour to Advance Short Text Classification: Using Heterogeneous Graph Neural Network via Building Sememe-relationships

Abstract

Short Text Classification (STC) is one of the fundamental tasks in natural language processing. The lack of grammatical structure and contextual information causes it challenging. One approach is to improve the STC by introducing the label information of entities via the entity knowledge base to build a hierarchical heterogeneous graph. However, the previous entity knowledge bases do not consider the complex semantic relationships of entities, and the number of entities in the articles is too large, affecting the computational resources. This paper proposes using sememes instead of entities to exploit the deeper semantic relations between words better to build heterogeneous graph networks. As the smallest semantic unit, the sememe consists of a finite number of words. We utilized Self-attention to find the sememe in short texts and the weight parameter between them. Extensive experiments results have demonstrated that our proposed method outperforms state-of-the-art methods on the Snippets dataset for STC.