基於個人化圖譜增強大型語言模型於主動式資訊召回之研究

Abstract

在日常生活中，人們常會碰到需要回憶過往生活的時刻。然而人們不可能記得每一個細節，因此提供一個可以主動地去提醒人們是否有細節被遺忘或錯置的系統便相當重要。在此研究中，我們著重於兩個目標：如何透過大型語言模型來進行主動式的資訊召回和如何將個人化的知識圖譜運用於提升原始模型的能力。 我們在架構中導入了一個用於修正原始答案的模組，此模組會分析來自於知識圖譜模組的資訊，藉此來改善原始模型預測的結果。在實驗結果中表明，我們的架構不僅能夠提升ChatGPT和Llama3等大型語言模型的性能，也能運用在不同的SOTA模型。 整體來說，我們的研究提出了一個新架構，不僅可以用於加強原始模型的準確率，且在彈性高的特性下，我們可以替換不同的原始模型及搜索知識圖譜的方式，讓模型的成效有更進一步的提升。這個架構不僅在主動式資訊召回這個任務下取得不錯的結果，同時也為其他領域的研究提供了如何利用知識圖譜來運用大型語言模型的方法。

Several studies have focused on developing information recall systems to help individuals remember their personal life experiences. This is due to the fact that people cannot recall every detail of their life events and may sometimes confuse different events. Providing a system that can proactively remind individuals of forgotten or confused details is therefore essential. Recent advancements in large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation. Many works are also exploring how to integrate LLMs into personalized applications. This thesis focuses on two objectives: (1) leveraging LLMs for proactive information recall and (2) introducing personal knowledge graphs to augment the capabilities of detecting information recall needs by refining the decision. We construct a framework that incorporates a refinement module to consult structure information from personal knowledge graph, thereby achieving precise recall support. This framework offers high flexibility, allowing for the replacement of different base models and modification of fact retrieval methods to further improve robustness. Experimental results show that our framework more effectively aids in detecting forgotten events, achieving the goal of helping users recall past experiences.