Research On Multilingual Knowledge-Grounded Conversational System

Dialogue systems aim to enable machines to have natural conversations with humans,which is an important and challenging task in the field of artificial intelligence.As an important category of dialogue systems,open-domain dialogue systems have received increasing attention and research from the academic and industrial communities.Open-domain dialogue systems primarily focus on chitchat and can engage in communication with users on any topic or in any domain.Currently,open-domain dialogue systems have been widely applied in people’s lives,with chatbots on various platforms being typical examples.However,open-domain dialogue systems struggle to engage in in-depth conversations on specific topics and often generate overly generic responses.To further improve the performance of dialogue systems in conversations,knowledge-grounded conversation systems have emerged.Knowledge-grounded conversation systems aim to enhance the quality of communication between dialogue systems and users by incorporating relevant background knowledge.Previous work mainly used monolingual knowledge to enhance dialogue systems,which suffers from a lack of background knowledge.Utilizing multilingual background knowledge helps cover a wider range of conversation topics,thus alleviating the problem of insufficient background knowledge.Furthermore,to serve users from different language and cultural backgrounds,dialogue systems also need to have the ability to handle multiple languages.Therefore,this thesis proposes the use of multilingual knowledge to enhance dialogue systems,which constitutes the task of multilingual knowledge-grounded conversation.The use of multilingual knowledge to enhance dialogue systems faces three challenges:(1)Cultural biases exist among different language knowledge,making it difficult to select appropriate multilingual knowledge.(2)Translation errors may occur in this multilingual task.(3)Currently,there is a lack of multilingual datasets and true labels for multilingual knowledge selection.To address these challenges,this thesis proposes a multilingual knowledge-grounded conversation model,which is constructed based on multilingual pre-trained models and consists of a multilingual knowledge selection module and a response generation module.Additionally,the thesis decomposes the multilingual knowledge selection task into two stages to achieve more fine-grained knowledge selection:first,selecting the most suitable knowledge from the knowledge set of each language to form the selected knowledge set,and then selecting the most appropriate knowledge from the selected knowledge set.Finally,the thesis designs a two-stage model optimization strategy combining contrastive learning and reinforcement learning to fully leverage existing data to optimize the proposed model and enhance the multilingual representation ability of the model.Experimental results on two datasets demonstrate that the multilingual knowledge-grounded conversation model outperforms baseline models,indicating its ability to select relevant knowledge from multilingual background knowledge set and generate more informative responses,thus enhancing the user’s dialogue experience.Ablation experiments demonstrate the effectiveness of each component of the multilingual knowledge-grounded conversation model and the model optimization strategy combining contrastive learning and reinforcement learning,as well as the effectiveness of the multilingual knowledge base in improving the performance of dialogue systems.