具有多步推理能力的個人記憶問答系統

Abstract

具備記憶功能的機器人能夠幫助我們回憶過去的個人事件，例如特定的人、事、物等。對於那些記憶力較弱或容易忘事的人來說，這類機器人可以作為一種認知支援工具，幫助他們記住重要資訊。此外，亦有研究指出，社區長者對這種具備記憶輔助功能的機器人有著較高的接受度和偏好，顯示其在提升長者生活自主與關懷方面的潛力。

現實世界使用者的問題可能複雜或表述不清，因而需要多步驟推理或整合多個資訊來解答，這也是當前智慧助理所面臨的挑戰之一。本論文提出了一個具多步推理能力的記憶輔助問答系統。首先，我們從與使用者的對話中提取出事件導向的記憶內容，分析事件中的命名實體及其關係，並儲存在記憶庫中。當使用者提出問題時，系統能迭代地檢索相關資訊並評估檢索結果的信心程度。我們的檢索系統不僅使用文字檢索，還運用圖形檢索以更好地捕捉和利用記憶之間的關係來增強檢索效果。根據信心評估結果，系統可以將問題進一步拆解成子問題或進行改寫，以便更精確地檢索答案。在這個過程中，系統會構建一個樹狀結構來組織解答，最終提供一個合適的答案給使用者。

實驗結果顯示，我們設計的動態樹查詢分解模組與圖搜尋技術相結合，在多跳問答資料集上能夠更有效地針對複雜問題檢索所需的參考資訊並生成答案。此外，這一方法在使用大型語言模型時，所需的token花費顯著低於相關的研究工作。

Robots equipped with memory capabilities can assist individuals in recalling past personal events, including specific people, events, and objects. They can serve as cognitive support tools for those who have difficulty remembering things, helping them to recall important information. Additionally, studies have shown that community-dwelling elders show higher acceptance and preference for such memory assistance features, demonstrating their potential to enhance the autonomy and care of elderly individuals.

In real-world scenarios, users' queries can be complex or poorly expressed, requiring multi-step reasoning or aggregating several pieces of information to provide an answer. This can be challenging for existing intelligent assistants. This thesis proposes a memory-assisted question-answering system with multi-step reasoning capabilities. Firstly, the system can extract event-oriented memory content from the user's conversations, identify the named entities and their relationships within those events, and store them in a memory bank. Given a user's question, the system iteratively retrieves relevant information and evaluates the confidence level of the retrieval results. Our retrieval system uses not only text-based retrieval but also graph-based one to better capture and utilize the relationships between memories to enhance retrieval effectiveness. Based on the confidence assessment, the system can further decompose the question into sub-questions or rewrite it to allow for more precise retrieval of answers. During this process, the system constructs a tree structure to organize the answers, ultimately providing the user with an appropriate response.

The experimental results show that our designed dynamic query decomposition tree paradigm, combined with graph search techniques, can more effectively retrieve the necessary reference information and generate answers for complex questions in multi-hop question-answering datasets. Furthermore, this method incurs significantly lower token costs when using large language models compared to related research works.