Short Text Classification via Hypergraph Convolution Network

Effective classification of short texts is a key task in many areas, such as web search, literature recommendation and emotion analysis and so on. It can effectively help users find relevant information. With the increasing number of short texts, how to classify short texts with short content and sparse features has become a popular research theme in recent years. We propose a short text classification method via hypergraph convolution network, which can flexibly capture the complex high-order relationships of words in the short text. Specifically, firstly, the method models an individual text as hyperedges, so that all words are connected with each other. Different hyperedges are connected by shared words and form a hypergraph containing high-order correlation. Secondly, a hypergraph convolution network is performed to learn the feature representation of words and short text. Finally, the classification model is used to achieve short text classification. The results of the experiments suggest that this method can successfully classify short texts. Further experimental analysis verifies the importance of using hypergraph to model short text to alleviate the problems of sparse data and limited labeled data, and proves the rationality and effectiveness of this method.