Research On Chatbot In Combination Of Retrieval And Generation

As the smart speaker and online customer service enter many fields of our daily life,the frequency of people's dialogue with machines has increased significantly,and open domain chatbots are more and more important.With the goal of imitating human conversation,chatbots can not only reduce the cost of customer service for enterprises,but also meet the needs of users' daily chat.Therefore,chatbots have become a research focus in the field of computer science.At present,most of the open domain chatbots are data-driven models,and they can develop from rule-driven ones mainly because of the popularity of social media and the development of computer hardware.Social media generate large-scale dialog corpus,while the development of computer hardware accelerates the calculation of large throughput.From the perspective of implementation methods,data-driven chatbots can be classified into retrieval-based and generation-based methods.Retrieval-based chatbots match replies through context similarity,while generation-based models generate replies based on context.There are advantages and disadvantages in the above two methods.For example,retrieval reply has rich information but lacks relevance,while generative reply has high relevance but the information is not rich enough.Therefore,learning from the strong points of the retrieval-based method and close the gap of the generation-based method,this paper focuses on the construction of a model combining retrieval and generation.In particular,the main contents of this paper are as follows:(1)An open domain dialogue generation model based on multi-view adversarial learning is put forward.The model consists of two parts:a generator for generating a realistic reply and a discriminator for identifying the generated reply.The highlight of this model is that it uses a binary discriminator for multi-view confrontation training.The simple sentence discriminator uses CNN(Convolutional Neural Networks)to model the generated sentence and then calculates the probability that this sentence is judged to be false".The dialogue discriminator use LSTM(Long Short-Term Memory)to model the context information and the generated response at different granularities and then calculate the probability that this set of dialogs will be judged as "false".Verification has been conducted on the open Douban Chinese dialog corpus,and the experimental results show that the model can indeed improve the informativeness and relevance of generative reply(2)We propose a retrieval-polished(RP)model for response generation that polishes a draft response based on a retrieved prototype is put forward.The model polishes the generated responses based on retrieved responses to get smoother,more informative responses.We first adopt a prototype selector to retrieve a contextually similar prototype.Then,a generation-based polisher is designed to obtain a polished response.Finally,we introduce a polished response filter to select the final reply between the retrieved response and the polished response.Experiments have been conducted on the open Douban Chinese dialog corpus,and the model has achieved good results in diversity,relevance and information quantity,whether it is automatic or manual testing.(3)The above models are deployed to a real application,and thus a chatbot based on WeChat official account is developed.It is convenient to access and easy to use the chatbot.Users just need to log on to WeChat,and then can achieve a human-machine conversation in the official account chat interface.