Key Technology Study Of Sociable Conversational Recommendation System

In recent years,with the significant progress in the natural language processing field and the increase of applications such as voice assistants and chatbots,the Conversational Recommender System(CRS)have also received extensive attention for supporting natural language interactive recommendations.How to use its natural language interaction advantages to achieve a higher recommendation success rate is one of the key challenges in the field of conversational recommendation.Compared with task-oriented conversational recommender systems,sociable conversational recommender systems focus on better natural language interaction capabilities and more diverse interaction methods.Therefore,we investigates the key issues of sociable conversational recommender systems.Currently,conversational recommender systems usually model user preference and item representation and make recommendations accordingly,so high-quality user preference representation and item representation are the key to improving the success rate of recommendation.In addition,most dialog recommender systems consist of a recommendation module and a dialogue generation module.Under this non-unified conversational recommendation framework,the inconsistency between the two modules makes the recommendation performance of the recommendation module not well reflected in the final response.Aiming at the above problems,we carried out the following research work:(1)Aiming at the problem of modeling user preference representation in conversational recommender systems under multiple rounds of social conversational recommendation tasks in multi-recommendation fields,we propose the concept of topic-related user preference graph and a Conversational Recommender Systems Based on Topicrelated User Preference Graph(CRTPG)based on topic-related user preference graph.CRTPG avoids the interference of entity information irrelevant to the recent topic or the user is not interested in by modeling the user’s preference related to the recent topic,and predicts the key entities of the subsequent dialog based on the user preference graph to guide the dialog generation module to generate a response.We conduct experiment on the DuRecDial dataset.Compared with the public literature we know,CRTPG achieved the best results in both recommendation indicators and dialog indicators.(2)Aiming at the problem of difficulty in learning low-frequency item representations in conversational recommendation datasets,we propose a template-based data augmentation method by constructing new samples of low-frequency items in the recommendation scene.This method helps the model learn better low-frequency item representation,thereby improving recommendation performance.In addition,we propose a baseline method based on oversampling,which is also compared with template-based data augmentation methods from a data processing perspective.Experiment results on the ReDial and INSPIRED datasets demonstrate the positive effect of the method on low-frequency item representation learning.(3)Aiming at the consistency problem of conversational recommender systems under the non-unified conversational recommendation framework,this thesis proposes a recommendation-aware dialog generation module,consisting of a pre-trained language model adapted to the recommended domain and a recommended scene discriminator.The recommendation-aware dialog generation module improves the generation accuracy of recommendation slots in reply templates with only a small number of parameters added.On this basis,we built a Conversational Recommender System with HighConsistency(CRS-HiCon),including a semantic fusion module,a domain-adaptive dialog generation module,and a recommendation module.The experiment results on the ReDial dataset show that the domain-adaptive dialog generation module and the recommendation discriminator narrow the recommendation performance gap between the recommendation module and the final response,and improve the recommendation consistency between the two modules.In general,the research of this thesis solves some problems in the existing dialog recommendation field from three aspects,and improves the recommendation performance of the social dialog recommendation system.