An end-to-end information fusion model for task-oriented dialogue systems

There are many aspects to consider when building an end-to-end task-oriented dialogue system, but recent studies have ignored the importance of dialogue history and textual representation. Although existing neural networks can effectively construct a linear relationship between dialogue history and knowledge base information, it is difficult to improve the dialogue quality. In this paper, we propose an end-to-end neural network model that reuses dialogue history information for task-oriented dialogue systems. We also reconstructed a large-scale word vector data set for use in NLP research. Our model integrates the implicit forgetting information and selection information of the dialogue history into the dialogue generation results through LSTM cell units, and uses Bert pre-trained word vectors for feature transformation to enhance entity representation capabilities. Finally, our model achieved a maximum increase of 2.4% in F1 value and a maximum 4% increase in Bleu value on the public data sets Incar and CamRest.