UniMS-RAG: A Unified Multi-source Retrieval-Augmented Generation for Personalized Dialogue Systems This paper proposes a novel unified multi-source retrieval-augmented generation (UniMS-RAG) framework for personalized dialogue systems. The framework addresses the challenge of selecting and retrieving knowledge sources, and generating personalized responses in dialogue systems. The system decomposes the task into three sub-tasks: knowledge source selection, knowledge retrieval, and response generation. It introduces two types of tokens: acting tokens to decide the next action (e.g., which source to use) and evaluation tokens to evaluate the relevance score between dialogue context and retrieved evidence. The framework is trained in a sequence-to-sequence paradigm, allowing the model to adaptively retrieve evidences and evaluate relevance using these tokens. A self-refinement mechanism is also introduced to iteratively refine the generated response during inference, considering the consistency scores between the generated response and retrieved evidence, as well as the relevance scores between the dialogue context and retrieved evidence. The framework is evaluated on two personalized datasets, DuLeMon and KBP, and achieves state-of-the-art performance on the knowledge source selection and response generation tasks. The results show that UniMS-RAG outperforms previous methods in terms of performance and effectiveness. The framework is also evaluated through extensive experiments, including automatic evaluation and human evaluation. The proposed framework not only improves the training and inference of SAFARI but also offers more extensibility and flexibility for personalized dialogue systems. The paper is partially supported by grants from the RGC General Research Funding Scheme.UniMS-RAG: A Unified Multi-source Retrieval-Augmented Generation for Personalized Dialogue Systems This paper proposes a novel unified multi-source retrieval-augmented generation (UniMS-RAG) framework for personalized dialogue systems. The framework addresses the challenge of selecting and retrieving knowledge sources, and generating personalized responses in dialogue systems. The system decomposes the task into three sub-tasks: knowledge source selection, knowledge retrieval, and response generation. It introduces two types of tokens: acting tokens to decide the next action (e.g., which source to use) and evaluation tokens to evaluate the relevance score between dialogue context and retrieved evidence. The framework is trained in a sequence-to-sequence paradigm, allowing the model to adaptively retrieve evidences and evaluate relevance using these tokens. A self-refinement mechanism is also introduced to iteratively refine the generated response during inference, considering the consistency scores between the generated response and retrieved evidence, as well as the relevance scores between the dialogue context and retrieved evidence. The framework is evaluated on two personalized datasets, DuLeMon and KBP, and achieves state-of-the-art performance on the knowledge source selection and response generation tasks. The results show that UniMS-RAG outperforms previous methods in terms of performance and effectiveness. The framework is also evaluated through extensive experiments, including automatic evaluation and human evaluation. The proposed framework not only improves the training and inference of SAFARI but also offers more extensibility and flexibility for personalized dialogue systems. The paper is partially supported by grants from the RGC General Research Funding Scheme.