Research And Implementation Of Knowledge-driven Generative Dialogue Model

In recent years,the advancement of deep learning technology and the emergence of massive dialogue data have made it possible to build data-driven generative dialogue systems.Traditional generative dialogue systems tend to generate generic and boring responses due to the lack of background knowledge.To this end,researchers try to introduce external knowledge into the dialogue generation model to improve the informativeness of the generated responses.This thesis mainly studies the knowledge-driven dialogue generation model and enhances the quality of the response by improving the model from the perspective of knowledge selection and the response generation process.The main work of this thesis is as follows:(1)This thesis proposes a dialogue generation model incorporating external knowledge.Global and local knowledge selection methods are proposed to improve the model's ability to utilize external knowledge.The model first selects knowledge from a global perspective based on the relevance between the dialogue context and candidate knowledge,which provides the decoder with initial knowledge information.During response generation,local knowledge selection is performed to capture relevant external knowledge based on an attention mechanism.In addition,to alleviate the impact of Out-Of-Vocabulary words on the consistency between responses and external knowledge,this thesis introduces a copy machine to the decoder by improving the pointer-generator network.This enables the model to reference words in external knowledge directly.This thesis conducted comparative experiments on the public dataset Duconv and the Wizard-of-Wikipedia,respectively.the results showed that the proposed model could utilize external knowledge to generate more coherent and informative responses than the baseline models.(2)This thesis proposes a multi-turn dialogue generation model based on the reinforcement learning framework.This thesis models and trains response generation and knowledge selection,respectively.First,the model is pre-trained with parallel corpus to improve the ability to generate responses given the knowledge and dialogue context.Then reinforcement learning methods are used to optimize the model's knowledge selection strategy during multi-turn dialogue through direct and indirect rewards.The direct reward considers the influence of knowledge selection on the amount of information,and the indirect reward considers the impact of knowledge selection on context relevance and knowledge consistency.This thesis conducted experiments on the Persona Chat dataset,proving that the proposed model can make proper knowledge selection in multi-turn dialogue and improve the coherence of the dialogue.(3)Based on the proposed models,this thesis designs and implements a human-machine dialogue system.The system can use the knowledge in the movie field to conduct multi-turn dialogue with users.