Emotional Dialogue Response Generation Via Reinforcement Learning

In recent years,with the continuous update and improvement of deep learning algorithms and the improvement of computing power of computer hardware devices,research on chatbot has also made great progress and breakthroughs.Generative chatbots have broad application prospects due to their good mobility and generalization.This article mainly studies open-domain generative chatbots,and is committed to adding emotional intelligence to conversations,while improving the quality of conversation generation,and establishing emotional fetters with users to make the conversation response content smoother and more diverse.The main work of this article is as follows:1.Aiming at the problems of safe responses and lack of emotional factors in opendomain generative dialogues,this paper proposes an algorithm for generating emotional dialogue responses based on reinforcement learning.Firstly,based on the analysis of sentiment data,a multi-emotion transfer matrix in the dialogue is obtained,and the multiemotion classification model is used to add the emotional supervision information to the dialogue corpus.Combining reinforcement learning algorithms,construct a reward function from two aspects of content quality and emotion,and then optimize its strategy so that it can choose the most emotional and smooth response sentence with the greatest future benefit.Experiments show that compared with the benchmark method,the content of the emotional dialogue response generated by the proposed method is clearer and smoother,and the emotion is more abundant.2.Aiming at the problem that the public sentiment classification dataset has a small impact on the effect of subsequent emotional dialogue generation,this paper proposes a fusion model of sentiment word vector classification and reinforcement learning algorithms.First,a multi-task joint sentiment word vector model is proposed,which combines language model and sentiment classification model,weighs semantic representation and sentiment representation,and embeds finer-grained sentiment information into the word vector.Then,the emotion classifier and the reinforcement model are fused,and an emotional diverse cluster search algorithm is added in the decoding stage,which effectively avoids the secure reply and increases the emotional color in the dialog reply.Experiments show that compared with the benchmark method,the proposed method has stronger generalization ability,and the sentimental tendency of generating responses is more obvious.In this paper,emotion is integrated into the dialogue through reinforcement learning,and combined with experiments to explore and verify the emotion dialogue reply generation algorithm,the dialogue emotion generated by the proposed model is more abundant,and has important reference value for the subsequent dialogue research.