Response Generation In Multi-turn Dialogue Using Topic Prediction

In recent years,the research of the open domain dialogue system has received extensive attention from academia and industry.On the one hand,human-machine dialogue system is an important form of artificial intelligence,whose in-depth study is beneficial to promote the development of artificial intelligence and natural language.On the other hand,open domain dialogue system is of great application value.It can meet the users' need of communication,and at the same time,its application in automatic customer service can smooth the dialogue process,improve user stickiness and reduce labor costs.As a major technical route of the dialogue system,response generation model has also become a focus and hotspot of current research.Although the core technology of response generation is constantly evolving,various systems tends to give lower semantic richness such as "I do" and "I don't know" in practice,which is named as "generic response",and therefore it is difficult to guarantee the attraction and sustainability of the dialogue.In addition,most research work has been carried out for single-turn dialogue,which is difficult to meet the requirement of the context information in the real dialogue scene.This paper mainly conducts the following three research contents:(1)Topic model building for short text conversations.Owing to the problem that the social media conversation text is short and the colloquialism is serious,it is difficult to mine the implicit topic information.This paper uses "pseudo-long document" method to construct the training corpus of the topic model,combines the TextRank method to remove the meaningless words in dialogue text,and uses the perplexity to get the proper number of corpus topics.At last,a topic model for social media short text conversations is constructed.(2)Topic prediction for response in multi-turn dialogue.A topic prediction model combining recurrent neural network and convolutional neural network is proposed,in which the convolutional layer and pooling function are optimized.This paper predicts the topic of response by mining the trend of the topic in the dialogue,and the proposed model performs well on the experimental test set.(3)Response generation in multi-turn dialogue using topic prediction.Based on the research contents of the two chapters before,a model for multi-turn dialogue response generation with topic prediction(TPJA-Seq2Seq)is proposed,which uses the hierarchical model combined with the self-attention mechanism to express the dialogue history,and uses the joint attention of dialogue topic,dialogue context and query to generate the response of dialogue.The model performs well both on the automatic evaluation index and the samples of response,which is in line with the expected effect of this paper.