Research On Human-computer Conversation Generation Model Integrating External Knowledg

In recent years,artificial intelligence technology has developed rapidly.Among them,the human-machine conversation system is the hot spot of current research.Traditional conversation generation technique is mainly based on the sequence to sequence framework,the framework from the massive data set learning conversation context semantic information and generate related reply,but based on the framework generated reply generated common response and contains less information.This problem affects the user experience to a certain extent and reduces the continuity of the human-computer conversation.Using external knowledge can deepen semantic understanding,enrich response expression,provide semantic information beyond the context of the conversation,and enable the model to generate more diverse responses.Therefore,this paper conducts research on the above issues,and the main research contents are as follows:(1)This paper studies how to select appropriate external knowledge and integrate knowledge into response generation,and proposes a knowledge-driven conversation generation model: Siamese Network based Posterior Knowledge Selection Model for Knowledge Driven Conversation Generation(SPK-CG).We design a new knowledge selection mechanism to obtain knowledge information that is highly relevant to the context of conversation.The posterior knowledge distribution is used as soft label to keep the prior distribution consistent with the posterior distribution in training process.At the same time,in order to narrow the gap between prior and posterior distribution and improve the accuracy of knowledge selection,a multi-granularity matching module is designed for knowledge selection using siamese network.Compared with the previous knowledge-based conversation generation model,our method can select more appropriate knowledge and use the selected knowledge to generate response that are more relevant to the conversation context.Results from both automated and manual evaluations show that our model has advantages over previous baseline models.(2)To solve the problem that knowledge selection and response generation are divided into two processes,resulting in insufficient knowledge fusion in decoder,this paper proposes Adaptive Knowledge Selection Interactive Conversation Generation Model(AKS-CG).The process of knowledge selection is integrated into the response generation process of decoder to improve the degree of combination of knowledge selection and response generation.The knowledge selection is dynamically updated through the current decoding state to obtain knowledge information more relevant to the current decoding state.At the same time,a double copy replication mechanism is designed according to the idea of pointer network.Flexible copying of entity words from input context and selected knowledge information can alleviate OOV(out-of-vocabulary)problems and make the generated response contain more useful information.The results of automatic and manual evaluation show that the proposed model can improve the accuracy of knowledge selection and the quality of generated response.