Research On Reinforcement Learning For Open Domain Chatbot Dialogue Generation

With the development of deep learning algorithm,a generate algorithm of chatbot is accessible,which make the dialog between chatbot and customers no longer rely on template matching and retrieval.For the generate algorithm possess better migration and generalization performance in the contrast,it has a wider application scenarios.This paper focus on open domain chatbot,differing from task-oriented chatbot which aim at assist people do something appointed.Our purpose is not to help the user as soon as possible to complete the assigned task,but to keep the user`s interest and let the users talk with the chatbot as long as possible.Therefore,we regard the dialogue length as the ultimate goal and try to extend it.Firstly,an experimental verification test was completed in the field of mainstream generative chat robot encoder-decoder algorithm model based on RNN Meanwhile,LSTM and GRU neural network unit statements influence on the results was also compared.Then add the attention model which will increase the generation effect,finally we use beam search algorithm to produce more diverse results.For the current seq2 seq generative algorithm utilizes the maximum likelihood method to estimate the output,it is quite easy to generate a large number of meaningless security replies.This paper propose to use reinforcement learning method to assess future reward,making the answer no longer for the current input only and just a simple selection of the maximum possible solution,but the reply considering the future rewards which can promote the sustainable development of the chat.We believe that a more intelligent robot should be emotional.In the reward function of reinforcement learning,this paper not only verify the fluency of the sentence generated,and also join the verification of the dialogue emotion to help the chatbot choose the reply sentences which have a positive emotion factor and achieve a continuing dialogue.Additionally,this paper regard the emotional information as a supervision signal added in the generation process to study the emotion transfer distribution.Then,using the emotion transfer distribution data as a signal to guide chatbot.Finally,we compare the methods mentioned above and several scholars` previous work by auto-evaluation and human appraise to verify the effectiveness of our proposed model.