Research And Implementation Of Knowledge-based Generative Dialogue System

Natural language processing is an important research direction of artificial intelligence,and dialogue systems have been widely used in people’s daily life.With the rapid development of the Internet and deep learning technologies,data-driven generative dialogue systems have become a hot research topic in recent years.However,traditional generative dialogue systems perform poorly in some scenarios and tend to generate lowinformation responses.For this reason,some researchers have proposed introducing external knowledge into dialogue generation models to improve the quality of responses.This thesis focuses on knowledge-based generative dialogue systems to improve the quality of responses in terms of external knowledge introduction,knowledge selection,and entity representation.The main research work of this thesis is as follows:(1)This thesis proposes a dialogue generation model that introduces external knowledge.The model uses dialogue memory and knowledge memory to store dialogue history and external knowledge separately,enabling the model to distinguish these two types of data more easily and thus introduce external knowledge more effectively.In the process of dealing with external knowledge,instead of using the traditional triplet form,the knowledge belonging to a subject is regarded as a whole,thus preserving the integrity of knowledge.A knowledge filtering mechanism is used to help the model filter out knowledge unrelated to the dialogue history to reduce the interference of irrelevant knowledge.Moreover,the traditional attention mechanism is combined with an end-toend memory network in the response generation stage,global and attention memory pointers improve the accuracy of selecting knowledge entities from global and local perspectives.Experimental results on Cam Rest and SMD datasets show that the model proposed in this thesis can introduce external knowledge more effectively and select knowledge more accurately than the baseline models.(2)This thesis proposes a dialogue generation model with enhanced entity representation.The model uses an enhanced entity representation to alleviate the problem that the representation of some knowledge entities is not accurate enough.For knowledge entities appearing in the context,the model incorporates information related to the dialogue context in its entity representation.For knowledge entities in external knowledge,the model includes information about other entities belonging to the same knowledge in its entity representation.It makes the entity representation more meaningful while reducing the dependence on word embedding vectors.In the process of selecting knowledge,the model uses a response-aware knowledge selection mechanism to make full use of the gold response information to help select the right knowledge.In addition,to better discriminate knowledge entities belonging to the same entity type,the model proposes an additional loss function to reduce the similarity between the word embedding vectors of knowledge entities belonging to the same entity type.Experimental results on the Multi-WOZ 2.1 dataset show that the model proposed in this thesis can generate higher-quality responses than the baseline models.(3)Based on the above dialogue generation models,this thesis designs and implements a knowledge-based human-machine dialogue system through which users can engage in dialogue with the models,thus more visually reflecting the performance of the models proposed in this thesis.