Exploring semantic awareness via graph representation for text classification

Text classification is a fundamental problem in natural language processing. Nowadays, text classification based on GNN attracts the attention of researchers. However, the existing works not fulfill well the transmission of contextual semantic information, and they pay more attention to capturing the local features instead of global. Such methods ignore the importance of keyword information features, so they can not fully mine the text-level semantic representation. To relieve such problems, we propose the GText model for discovering the basic features with words and establishing a deeper relationship representation between words and documents. Specially, we utilize semantic features graphs to achieve text semantic representation. Meanwhile, we propose semantic information passing(SIP) mechanism to transmit contextual semantic information, which can enhance the semantic representation from multi-views. In addition, the gate mechanism can further mine the explicit keywords of the whole document. With GText, the test accuracy on MR improved about 2% and on Ohsumed at most 9%, which illustrates GText can better achieve the mining and transmission of text semantic information. Experiments on several authoritative datasets show that our method is superior to the existing text classification methods.