Dialogue Recommendation Based On Dialogue State Tracking

With the development of Dialogue System,Dialogue Technology provides new ideas and methods to solve the problem that traditional recommendation is unidirectional and difficult to fit users’ online behavior.Conversational Recommendation.Conversational Recommendation System(CRS)is a kind of recommendation system that is guided by the recommendation task to conduct multiple rounds of user interaction and gradually discover users’ interests and preferences.In the past few years,the field of dialogue recommendation has welcomed a new research boom,and relevant research has made significant progress in many aspects.However,it still faces the following challenges: 1)insufficient semantic extraction ability of dialogue,which does not make full use of dialogue text information.The existing dialogue recommendation algorithm simply integrates the dialogue information into the traditional recommendation model through the encoder.The dialogue processing is rough and noisy,and the user preferences cannot be accurately extracted,resulting in limited recommendation effect and insufficient dialogue generation ability.2)Loss of user history information and inaccurate construction of user portraits.The dialogue recommendation system takes the dialogue information as the research object,and the same user is processed separately and generates a new user representation each time,which does not make full use of the user’s historical information and leads to the "cold start" problem.To solve the above challenges,this paper proposes a Dialogue recommendation method based on Dialogue State Tracking(DST).Recommendation is regarded as a task.This paper attempts to apply the dialogue state tracking technology in task-based dialogue system to the semantic extraction of dialogue information in dialogue recommendation system,so as to improve the recommendation effect and dialogue generation ability of the model.However,directly applying the existing dialogue state tracking model to the dialogue recommendation domain is not ideal.In the scenario of dialogue recommendation,the existing dialogue state tracking methods have the following problems: 1)the joint accuracy of dialogue state tracking is low,and the dialogue state extraction is not ideal.Due to the existence of subjective or unexpected labeling results in the production and labeling process of current task-based dialogue datasets,such noise will have a relatively large negative impact on the training of the model,especially in the field of few-sample classification,resulting in low joint accuracy of the model and little improvement in the performance of the dialogue recommendation system.2)The dialogue state tracking efficiency is not high and the reasoning time is long.The dialogue state tracking method takes the entire dialogue history information as the model input.This input form will continue to increase the length of the input data with the progress of the dialogue,thereby increasing the computational load of model training and inference,and there is a great efficiency problem.To solve the above problems,this paper proposes a dialogue state tracking method based on contrast learning,which can effectively promote the extraction of dialogue text information in the scene of dialogue recommendation.In summary,the research work of this paper mainly includes the following three contributions:(1)A dialogue state tracking method based on contrast learning is proposed.Firstly,in order to improve the joint accuracy of the existing dialogue state tracking methods,this method adds contrast learning to the dialogue state tracking training.By aggregating the common features of the same dialogue state and alienating the characteristics of different dialogue states,the model’s dialogue state type classification ability is improved,and data enhancement is carried out for the sample categories with few classification data.The labeled categories with small amount of data were integrated to improve the learning ability of the model on small sample classification data.Secondly,aiming at the problem of low efficiency of dialogue state tracking in the scenario of dialogue recommendation,this method decomposes dialogue state tracking into two sub-tasks: 1)dialogue state type classification;2)Dialog state generation.For each dialogue state,{NONE,DONTCARE,GEN} state type is first classified,and then dialogue state generation is performed for the dialogue state whose state result is GEN.Since only part of the dialogue state is generated in each round of dialogue,compared with the existing dialogue state tracking model,the efficiency of dialogue state tracking is greatly improved.(2)A dialogue recommendation method based on dialogue state tracking is proposed.Firstly,in order to solve the problem of insufficient semantic extraction ability of dialogue recommendation and insufficient use of dialogue text information,this method regards recommendation as a task of task-based dialogue system,and introduces dialogue state tracking technology to enhance the semantic extraction ability of dialogue text information in dialogue recommendation.Secondly,in view of the loss of User history information and the imprecision of user portrait construction,this method constructs a User Embedding matrix to store user historical preferences,and dynamically adjusts its characteristics in each conversation to integrate the long and short term interests of users.The specific steps are: The embeddings of all items and states in the database are obtained through GCN and RGCN training.At the beginning of training,the User Embedding matrix is initialized and the item embedding and state embedding appearing in the dialogue are extracted.user embedding is updated through the semantic unification module.(3)In this paper,the method of conversation state tracking based on contrast learning is comprehensively tested on the open data sets MultiWOZ2.0 and WultiWOZ2.1 which are widely used in the academic circle.Compared with other advanced methods,The combined accuracy of the model is improved by 2.15%-31.18%and 3.28%-51.86% respectively,and the reasoning efficiency is improved by 34 times.In this paper,the dialogue status tracking based dialogue recommendation method has been extensively tested on the open dialogue data set ReDial and compared with the current mainstream dialogue recommendation methods in the recommendation and dialogue modules.The feasibility of the method has been demonstrated by experiments.In the dialogue recommendation scenario,the Recall@10 of the model has increased by 7.2%.The diversity of Di D-2 generated by dialogue increased by 35.9%.