BiLGAT: Bidirectional lattice graph attention network for chinese short text classification

Chinese short text classification approaches based on lexicon information and pretrained language models have yielded state-of-the-art results. However, they simply use the pretrained language model as an embedding layer and fuse lexicon features while not fully utilizing the advantages of either. In this paper, we propose a new model, the bidirectional lattice graph attention network (BiLGAT). It enhances the representation of characters by aggregating the features of different hidden states of BERT. The lexicon features in the lattice graph are fused into character features with the powerful representation capability of the graph attention network, and the problem of word segmentation error propagation is solved at the same time. The experimental results on three Chinese short text classification datasets demonstrate the superior performance of this method. Among these datasets, 94.75% accuracy was achieved on THUCNEWS, 70.71% accuracy was achieved on TNEWS, and 86.49% accuracy was achieved on CNT.