GHGA-Net: Global Heterogeneous Graph Attention Network for Chinese Short Text Classification

As an important research content in the field of natural language processing, Chinese short text classification task has been facing two challenges: (i) existing methods rely on Chinese word segmentation and have insufficient semantic understanding of short texts; (ii) there is lacking of annotated training data in practical applications. In this paper, we propose the Global Heterogeneous Graph Attention Network (GHGA-Net) for few-shot Chinese short text classification. First, we construct the global character and keyword graph representations from the entire original corpus to collect more text information and make full use of the unlabeled data. Then, the hierarchical graph attention network is used to learn the contribution of different graph nodes and reduce the noise interference. Finally, we concatenate embedding with text vector and fuse the keyword and character features to enrich the Chinese semantics. Our method is evaluated on the Chinese few-shot learning benchmark FewCLUE. Extensive experiments show that our method has achieved impressive results in the classification tasks of news text and sentiment analysis, especially in minimal sample learning. Compared with existing methods, our method has an average performance improvement of 5% and less training consumption, which provides a new idea for few-shot Chinese natural language processing without relying on pre-training.