Research On Persona-based Open Domain Dialogue Generation Method

With the vigorous development of the Internet and the maturity of deep learning technology,the open domain dialogue system has attracted more and more researchers’ attention.The end-to-end generative dialogue system has achieved great success,but it still faces the challenge of generating generic responses and lacking speaker’s characteristics.Although the researchers have introduced the character’s personas,there are still two factors of this method that lead to poor responses:(1)Insufficient breadth of information utilization: Existing models emphasize the use of given explicit personas,but ignore the implicit personas that derived from inference of given information.There are still not enough researches on inference and fuse of two kinds of personas to improve the quality of responses.(2)Insufficient depth of information utilization: Researches show that only part of given personas in real conversations are related to the conversation history,and the deep features of semantic and topic have not been fully mined.Therefore,how to accurately select given personas and extract deep features such as semantic and topic still need to explore further.In the open domain dialogue methods based on personas,this paper aims to improve the quality of responses from the perspective of the breadth and depth of personas utilization.The main contributions of this paper are as follows:(1)Aim at the insufficient breadth of utilization of personas,this paper studies the extraction and utilization of implicit personas,and proposes an open domain dialogue system that combines explicit and implicit personas.The model designs an implicit personas reasoner to sample latent representation as implicit personas,and fuses two kinds of personas for reponse generation.This is the use of both explicit and implicit personas in a personal-based generative dialogue model,and the improvement of response quality is verified on ConvAI2.(2)Aim at the insufficient depth of utilization of personas,this paper studies the accurate selection of multiple personas in multi-turn dialogue and the use of its deep features of semantic and topic.This model designs a personas selection mechanism,a semantictopic features sampling mechanism,and a dynamical attention mechanism which can combines the above information dynamically in the response generation.Finally this model can generate a response containing context-related personas and deep features of semantic and topic.Compared with other methods,this model proves the effectiveness of personas information selection and the diversity of responses.To sum up,this article gradually develops the researches from different perspectives.First,we proposes an open domain dialogue system that integrates explicit and implicit personas;then we studies the accurate selection of personas and the mining of deep features about semantic and topic;we can conclude that the quality of responses can be improved from two perspectives,and it also be verified that have high theoretical and practical application value.