Research On Response Diversity And Proactive Methods For Chatbots

Building intelligent chatbots that can communicate naturally with humans has always been a big challenge in artificial intelligence.Especially,corpus and modeling methods have limited the open domain dialog system for a long time,so it is difficult to realize the free chat without any scene and topic constraints.With the development of deep learning and big data technology,a response generation method based on deep learning is proposed.The chatbot built by this method can chat with chatterers on any interesting topics and has good expansibility.However,the sequence-to-sequence modeling method based on deep learning model uses maximum likelihood probability to generate each word of the response sentence,which will result in the model generating many general responses,such as "I don't know.".This universal response is too single and seriously reduces the willingness of chatters to continue the conversation.Therefore,this thesis designs a new response generation model MSDialog to improve response diversity.First,by observing the dialogue in real life,it is found that for an input sentence,multiple response sentences are grammatically correct and semantically reasonable.that is,there may be various correspondence relations between input and response.First,based on the training corpus,the one-to-many mapping relationship between the input sentence and the real response sentence is modeled,and then multiple response sentences are generated based on the sequence-to-sequence model.Besides,by applying a topic derivation model to the input sentence to obtain the corresponding topic words,the occurrence probability of general response sentences can be further reduced when generating response words.The performance evaluation on the two public data sets of Cornell Movie and Daily Dialog,proves that the MSDialog model proposed in this thesis can generate diversified responses.On the other hand,the existing open-domain dialogue systems usually generate a response based on the input sentence of the chatterer,that is,the response is directly related to the input sentence,and the robot will not actively change the chat topic or further generate other chat contents that the user may be interested in.This thesis innovatively designs an active dialogue generation model Du Dialog,which can generate a response sentence and a suggestion sentence according to the chatter's input sentence.Prolonged rounds of the conversation by providing conversation cues with suggestion sentences.To realize this function,the model structure designed in this thesis contains two decoders to generate a response sentence and a suggestion sentence,respectively.In particular,the input sentence and the response sentence are fused by the fusion node during the generation of the suggestion sentence so that the generated suggestion statement is context-consistent with the input and response sentences.Since this is a new model of conversation,there is no corpus for active conversation.In this thesis,the training and evaluation of the Du Dialog model are realized by preprocessing multiple rounds of the Daily Dialog dataset and simulating the process of active dialogue.The experimental results show that the new model can generate response and suggestion sentences simultaneously,and the two sentences have a strong correlation with the input sentence.