TG Network: A Model that More Effectively Identifies the Use of the Auxiliary Word "DE"

Chuang Liu, Hongying Zan, Xuemin Duan, Kunli Zhang, Yingjie Han

Abstract: In the knowledge base of function word usage of “trinity”, the auxiliary word “DE” has the characteristics of high frequency and flexible usage. In this paper, a neural network model (TG network) is proposed to automatically recognize the usage of “DE”. In this network, the self-attention mechanism is firstly adopted as the first-layer feature encoder and GRU (gated recurrent unit) as the second-layer semantic extractor, and the recognition accuracy rate reaches 82.8%. Experiments show that the recognition effect of TG network is better than that of previous methods. In further experiments, the larger the window, the better the effect of the model is proved by setting different windows. At the same time, the fine-grained analysis of each usage category is carried out. In the future, it is expected that this model will automatically recognize more function words and the recognition results can be applied to other natural language processing tasks.