Improved method of dialogue model based on character and word fusion

Abstract

An important part of the dialogue robot is the understanding of semantics and the analysis of sentences. At present, most of the models are trained based on character vector and word vector. Compared with traditional models, this method can fully consider the long and irregular language characteristics of daily chit-chat dialogues, and make full use of character and word fusion to improve the dialogue performance of the model. At the same time, the word embedding matrix trained by the pre-trained model is calculated to obtain the same dimension as the character embedding. Then, the character vector and word vector are fused, and the transformer model is used for dialogue training. The experimental results show that the proposed method can better improve the understanding ability of the dialogue model.