條款圖譜優化檢索增強生成模型：以壽險產業為例

Abstract

檢索增強生成（RAG）技術雖能有效提升大型語言模型處理特定知識的能力，但在面對如保險條款等具有高度內部關聯性的專業領域文件時，傳統的扁平化檢索方式常導致上下文理解不足，因此生成零碎化的答案。為了解決此問題，本研究提出並實現了名為 Clause-RAG 的新型模型框架。此模型框架的核心在於解析保單條款的條文之間邏輯關係，並且建構出此條款的圖譜，此圖譜用以捕捉並表達條文間的結構化關係，從而讓模型在檢索階段時，可以整合更完整的上下文資訊，進而在生成答案階段時，能具備更深層次的上下文理解能力，並且生成出更全面且精準的答案。

本研究透過在真實壽險保單文件上進行的實驗，將 Clause-RAG 模型與標準 RAG 模型進行了嚴謹的比較。我們設計了兩種不同難度的問題集：「單點事實檢索型」與「多點邏輯推理型」，並由人類專家從全面性、多樣性與賦能性三個維度進行評估。

實驗結果清晰地表明，Clause-RAG 模型在整體表現上顯著優於標準 RAG 模型。此優勢在處理需要整合多個條文資訊的複雜推理問題時尤為突出，尤其在提升回答的「多樣性」方面，Clause-RAG 展現了明顯優於標準 RAG 模型的表現。研究亦發現，即便在基礎的單點事實查詢中，Clause-RAG 也能提供品質更佳的答案。

最後此研究也提出了 Clause-RAG 模型架構的限制以及未來研究方向，以期可以更近一步

本研究的主要貢獻在於驗證了以圖譜結構增強上下文理解，是解決專業領域複雜問答任務的有效途徑，為壽險產業抑或是相關領域的應用發展提供了具體的實證與方向。

While Retrieval-Augmented Generation (RAG) technology effectively enhances the ability of large language models to process specific knowledge, traditional flat retrieval methods often lead to insufficient contextual understanding and fragmented answers when dealing with documents from professional domains with high internal correlation, such as insurance policies. To address this issue, this study proposes and implements a novel model framework named Clause-RAG. The core of this framework lies in parsing the logical relationships between policy clauses and constructing a knowledge graph of these clauses. This graph is used to capture and express the structured relationships between clauses, enabling the model to integrate more complete contextual information during the retrieval phase. Consequently, during the answer generation phase, the model possesses a deeper level of contextual understanding and can produce more comprehensive and precise answers.

Through experiments conducted on authentic life insurance policy documents, this study performs a rigorous comparison between the Clause-RAG model and a standard RAG model. We designed two question sets of varying difficulty: "single-fact retrieval" and "multi-hop logical reasoning." The models were evaluated by human experts across three dimensions: comprehensiveness, diversity, and empowerment.

The experimental results clearly indicate that the Clause-RAG model significantly outperforms the standard RAG model in overall performance. This advantage is particularly prominent when processing complex reasoning problems that require the integration of information from multiple clauses. Especially in enhancing the "diversity" of the answers, Clause-RAG demonstrated a markedly superior performance compared to the standard RAG model. The study also found that even for basic single-fact queries, Clause-RAG can provide higher-quality answers.

The main contribution of this study is the validation that enhancing contextual understanding through a graph structure is an effective approach for solving complex question-answering tasks in professional domains. This provides concrete empirical evidence and a clear direction for the development of next-generation intelligent question-answering systems.