Research And Implementation Of End-to-End Intelligent Dialogue System Fusing Multi-Domain Knowledge Base

Task-oriented dialogue systems enable machines to interact with users using conversational communication to receive and understand user behaviour and ultimately complete tasks proposed by the user.The neural models currently used to build dialogue systems rely on a large amount of domain-specific data for training.This approach to model training allows the models to maintain good accuracy and task completion rates in the training data domain,however it also limits the models to a specific task domain.This makes it difficult for dialogue systems to adapt to multi-domain dialogue scenarios and presents a significant barrier to practical application.In order to solve the problem of indistinguishable dialogue domains in multi-domain dialogues,this paper proposed a cross-domain knowledge modeling model,which allows the dialogue system to better cope with complex real-world environments and thus achieve crossdomain communication more effectively.At the same time,this paper investigated the problem of the lack of consistency between sentiment and user behaviour in system responses,and proposed a multi-turn dialogue generation model that fusing act prediction and sentiment analysis,so that the model has a more human-like dialogue generation effect.Finally,an open domain task-based intelligent dialogue system was designed and implemented based on the proposed two dialogue models.The system can detect user user behaviour and automatically switch between the task-oriented dialogue system and the chat dialogue system,alleviating the formulaic and cut-throat feel of the system’s responses and greatly enhancing the user’s immersion when engaging in dialogue with the system.The specific research work in this paper is as follows:(1)a Cross-domain knowledge modeling model.In order to strengthen the domain recognition capability of the model,this paper constructed a model to model the domain information of knowledge,and introduced the implicit fusion of domain information and context into the dialogue generation process to give the model domain recognition capability.For the domain modelling module,this paper designed a single-domain pre-training process to migrate the model’s information capture capability in a single domain to a multi-domain scenario,further enhancing the domain information extraction accuracy of the domain modelling module and enabling the model to focus on knowledge that is highly relevant to the current dialogue domain.After practical validation with Multi WOZ 2.2,the F1 score of the model is 53.7% higher compared to the traditional baseline model,and its score increases by more than 12.5%.(2)A multi-turn dialogue generation model fusing act prediction and sentiment analysis.The model is built using pre-training model,which introduces a user behaviour prediction layer and a sentiment analysis layer to extract user act and sentiment representations of conversations,and dynamically fuses them with contextual representations for conversation generation.The loss functions of the modules are weighted and summed to jointly train all parameters of the model.The model was experimented on the casual conversation dataset Daily Dialog,where each BLEU score of the model exceeded that of the baseline model,and ablation experiments confirmed the enhancement of the model’s effectiveness by each module.(3)Open domain task-based intelligent dialogue system.An open-domain task-based intelligent dialogue system was implemented based on the research of the two dialogue system technologies mentioned above.This paper presented a detailed analysis of the technical requirements of the system,followed by an architectural design for the integration of the two models mentioned above,and finally a system test to troubleshoot operational problems.The system has more practical application than a restricted domain dialogue system.