主題模型強化之短文本檢索與生成品質提升

Abstract

檢索增強生成（RAG）的效能在很大程度上取決於檢索階段的品質。傳統的RAG 系統主要依賴向量相似度匹配，卻無法完美捕捉查詢與文檔之間的潛在主題結構，導致檢索結果不理想。為了解決這個問題，本研究在 RAG 的檢索階段引入主題建模（透過 BERTopic 流程 ），對 Amazon 產品評論進行主題聚類，並利用查詢重寫來判定查詢的主題 。透過將檢索範圍限定於相關主題，我們的方法能有效過濾無關文檔，提升檢索精度，並進一步改善下游文本生成的品質。我們在多個檢索基準數據集上進行實驗評估，結果顯示，相較於傳統的向量檢索方法（Naive RAG），BERTopic 檢索策略在檢索相關度與 RAG 的整體表現方面均有顯著提升。

The effectiveness of Retrieval-Augmented Generation (RAG) largely depends on the quality of the retrieval stage. Traditional RAG systems primarily rely on vector similarity matching, which cannot perfectly capture the underlying thematic structure between queries and documents, leading to suboptimal retrieval results. To address this issue, this study introduces topic modeling (via the BERTopic process) in the RAG retrieval phase, performs topic clustering on Amazon product reviews, and utilizes query rewriting to determine the topics of queries. By restricting the retrieval scope to relevant topics, our approach effectively filters out irrelevant documents, enhances retrieval precision, and further improves the quality of downstream text generation. We conducted experimental evaluations on multiple retrieval benchmark datasets, and the results indicate that, compared to traditional vector-based retrieval methods (Naive RAG), the BERTopic retrieval strategy significantly improves both retrieval relevance and the overall performance of RAG.