Research And Implementation Of Key Technologies Of Document-Driven Human-Machine Dialogue

Human-machine dialogue system is one of the important applications of natural language processing technology.Most of the current human-machine dialogues are at the two end-points of a spectrum of dialogues.The task-driven dialogue is on one end and chit-chat is on the other end.However,in real life,human dialogue is between task-driven dialogue and chit-chat,often a mixture of the two.Document-driven dialogue provides a bridge between them.It can not only respond to user's opinions in a chit-chat manner,but also exchange information about the topic of the conversation with the user in a task-driven manner.Therefore,document-driven dialogue has important value.Document-driven dialogue systems are designed to enable machines and users to interact around a specified document.There are still some problems in the current generative document-driven dialogue model.For example,the key information,which is conversation-related,in the document is underutilized;The model can't understand the user's intention by relying only on the user's current input,so it is difficult for the model to generate a good response;The model does not make full use of important entity knowledge such as person names and movie names in the document,and it is difficult to effectively construct a response associated with the document.To solve the above problems,this paper conducts research on document-driven dialogue model on the basis of extensive research.The specific contents are as follows:A dialogue model based on bidirectional-attention-encoded document(BDGen)is proposed.The model is based on the "encoder-decoder" framework and incorporates the method of document and user input interaction in machine reading comprehension technology.Experiments show that the model has better performance than existing document-driven dialogue models.In terms of answering user questions,the accuracy of BDGen model is improved by more than five percentage points.A dialogue model based on encoding and decoding fusion document information(D3G)is proposed.This model is further extended based on BDGen model.It introduces dialogue history information.And it improves the method of locating the key content of the document and decoding method,respectively proposes Doc-Reader mechanism and Multi-Copy mechanism.Experimental results prove that D3G model can generate more document-related and informative sentences.Based on the above model,a document-driven human-machine dialogue system in the field of film is implemented.Users can interact with the system through dialogue to obtain movie-related information.