Analysis Of Open Domain Dialogue Structure Under Chat Robot

As China’s population aging continues to aggravate,spiritual companionship is becoming more and more important for the elderly,Chatting robot can reduce loneliness and improve the quality of life for the elderly.The core technology of the chatting robot is dialogue system.As the core technology in the field of natural language processing,it promotes the continuous development of human-computer interaction,which has great research significance and application value.Dialogue system can be divided into task-based dialogue and open domain dialogue according to domain classification.Task-based dialogue usually has a clear goal and a specific scope of knowledge,most of which can be developed quickly by using relatively mature technology.However,open domain dialogue has no limited topic or clear goal,and man-machine can have free dialogue on any topic,which poses many new challenges to the research and development of relevant dialogue technology.Currently open domain dialogue generation technique are the main framework of the sequence to sequence and its improved model,which has the trouble with the model characteristics and application scenario is not completely matching.Therefore,this dissertation is mainly based on the open domain dialogue under the chatting robot,the exploration and research of its conversational structure are carried out in depth which lays the foundation for subsequent research on innovative dialogue generated processing architecture and Promotes the transformation of dialogue generation technology model paradigm.This dissertation comprehensively summarizes the development of open domain dialogue technology,and sorts out the research status and achievements at home and abroad from the four dimensions of reply quality optimization technology,attribute control technology,topic continuation and transformation control technology and dialogue technology improvement direction.To study session structure analysis from turns,adjacent pairs,topic clues which the topic hierarchy a description of the relationship between dimensions from bottom to top in turn dialogue scenes and speaker tagging and labeling for automatic segmentation of turns,dialogue action tags and adjacent pairs forecasting,the segmentation of topic module and topic module application,topic continuity and transition.Which is proposed based on BERT-TCN-CRF turn segmentation method,the adjacent pairs to the core function of BERT-BIGRU-CRF prediction method,the adjacent pairs dependence relationship of BERT-TCN-CRF prediction method,the topic segmentation of BERT-TCN-CRF prediction method,the WAE and its improved model topics clues subject classification,the semantic similarity analysis method of Siamese BIGRUAttention,the clustering of the topic of Singlepass improvement model.The main work is as follows:(1)For the problems that films and television shows have subtitles lack turns structure,the researchers always impossible to judge whether the captions of the two consecutive sentences belong to the same speaker.In this paper,BERT-TCN-CRF turn segmentation method is proposed,can realize the accuracy tag of the different speakers.(2)For the problems that how to realize the dialogue process of comprehensive control according to how to predict to the different session in the process of dialogue to adjacent pairs to before and after a round of core function and dependence relationship.The key is proposed based on BERT-BIGRU-CRF adjacent pairs to the core function prediction model,and based on BERT-TCN-CRF adjacent pairs to the dependence relationship prediction model,implement the adjacent pairs to the basic words of before and after the parts form turns structure unit forecasts.(3)For the problems that the lack of boundaries between different topics clues in the dialogue process,the subtitles of the dialogue in the form of text flow exist,the different topics clues continuously replaced,there is no clear boundary between them,it is not convenient to retrieve and utilize.This paper proposes a topic segmentation method based on BERT-TCNCRF,which can achieve accurate segmentation of different topic clues in the process of dialogue.(4)For the subtitle scene stream discourse message string fragments marked with conversationist ID as input,each topic clue can be automatically recognized and the accurate before and after boundary position of topic clue can be determined.The topic classification of topic clue based on WAE and its improved model is proposed to achieve fine adjustment of the boundary position of each topic clue fragment.(5)For the identification problems of topic continuation and transition from the perspective of "semantic relationship between before and after turns",and then effectively realize the control of topic clues in the dialogue process.This paper proposes a semantic similarity analysis method based on Siamese-Bigru-Attention.The semantic similarity calculation method is based on twin network and has strong semantic similarity calculation ability.(6)For the problems to the clear extraction of topic categories belongs to the topic clues in the process of dialogue,and further clarify the topic continuation and transition topic categories.A topic clustering method based on the improved Single Pass model is proposed.This method is very suitable for topic clustering of conversation flow.Combined with the above semantic similarity analysis method based on Siamese-Bigru-Attention,the effect is particularly good.The above research results have a certain role in promoting the analysis and research on the conversation structure of open dialogue technology based on high-quality multi-round and unrestricted conversation data,and have important significance.Furthermore,the quality of dialogue generation in the field of open dialogue under the chat robot is further improved,so that the elderly people can have a better chat experience.