Research On Deep Learning Based Open Domain Multi-turn Dialogue System

Building a dialogue system that can communicate with human nature is a long-term vision in the field of artificial intelligence.In recent years,the research of dialogue system has made rapid development.On the one hand,it benefits from the development of deep learning technology and the increase of Internet dialogue data.On the other hand,it benefits from the application prospect and commercial value brought by the landing of intelligent assistant and chatting robot.Therefore,the research of dialogue system is of great significance to the development of technology and industry.This paper mainly studies the open domain based multi-turn dialogue system,combining retrieval based and generation based technologies to solve the key problems faced by multi-turn dialogue scenarios.The first part of this paper studies the multi-turn dialog candidate response selection in the retrieval dialogue system,and proposes a multi-turn response matching model based on memory network to extract context-related clue information and match the context with the response.Specifically,this paper first uses a sequence matching network to model the context of multi-turn dialogue,then constructs an effective memory network to incorporate the semantic information and sequence relationship information,and uses a multiple-attention mechanism to extract clue information from the context.Finally,it combines the features of traditional natural language processing to improve the performance of candidate response selection.The retrieval system can retrieve informative responses from the pre-built retrieval database.However,the limitation of the retrieval database capacity will lead to the problem that the retrieved response does not match the current context.Therefore,the second part of this paper proposes a multi-turn dialogue response rewriting model based on matching enhancement,combined with the generation-based method to rewrite the retrieved responses.Specifically,this paper first uses neural matching networks to extract the context-sensitive words in the candidate responses as skeletons,and then uses a hierarchical encoder-decoder model to generate appropriate responses based on the skeletons and contexts.The experimental results show that the rewritten responses improve the relevance to the context while retaining information.Dialogues in real scenes usually need to rely on context-related external knowledge.The first two parts of the work only learn semantic interactions from the dialogue corpus and it is difficult to adapt to the real dialogue scene.Therefore,the third part of this paper proposes a knowledge-driven multi-turn dialogue response generation model,using external knowledge related to the dialogue to improve the quality of the generated responses.Specifically,based on the hierarchical encoder-decoder model,this paper integrates external knowledge into the semantic representation of the context in the encoding process,and dynamically pays attention to the external knowledge during the decoding process,using the external knowledge to guide the model to generate more diverse responses.