基於自傳式記憶進行對話之個人化社交機器人

Abstract

在社會逐漸進入高齡化老人照護變成一個很重要的議題，然而我們的照護人力往往無法符合社會的需求，因此將機器人導入老人照護領域變成一個有效的做法，老人可以在與機器人交流中得到心靈上的陪伴。 在以往機器人與人類互動的過程中對話是一個很重要的元素，然而現有的人機對話未臻完善，其中一個很大的原因是機器人沒有記憶的能力，導致在長期的互動中機器人不能根據使用者過去所提過的資訊進行個人化的對話。本論文主要提出一個基於自傳式記憶之個人化社交機器人，機器人與使用者對話會以基於自傳式記憶的架構存進我們的記憶庫裡面，在機器人生成回覆前會結合基於變形器編碼的特徵以及階層式自傳式記憶去從記憶庫裡擷取出當前對話上下文最相關的記憶，最後把記憶和對話上下文送進生成模組，在生成模組中我們基於大型語言模型提出了個人化的提示策略以及重排序機制，最後使用者可以得到一個適當的個人化回覆。 在實驗的部分主要分成三個部分，第一個是記憶擷取的實驗，第二個是記憶對話生成的實驗，最後一個是人機互動的實驗，實驗結果顯本系統可以有效擷取過去對話的相關記憶並且可以自然地將記憶融入日常對話中，以此可以讓機器人提供長期個人化的陪伴。

In an aging society, elderly care has become an important issue. However, our caregiver workforce often cannot meet the demands of society. Therefore, introducing robots into the field of elderly care has become an effective approach, as the elderly can receive emotional companionship through interactions with robots. In the process of interaction between robots and humans, conversation is a crucial element. However, existing human-robot conversations are not yet perfect, and a significant reason for this is the lack of memory capabilities in robots. This limitation prevents robots from engaging in personalized conversations based on the past user-provided information during long-term conversations. This thesis proposes a personalized social robot based on autobiographical memory. Conversations between robot and user are stored in our memory bank using an autobiographical memory framework. Before generating a response, the system combines Transformer encoding and hierarchical autobiographical memory to retrieve the most relevant memories based on the current dialogue context. Finally, the memories and conversation context are input into the generative module. In the generative module, we propose a personalized prompting strategy and ranking mechanism based on a large language model. As a result, the user receives an appropriate personalized response. The experimental part of this thesis is divided into three sections: memory retrieval experiments, memory-grounded dialogue generation experiments, and human-robot interaction experiments. Experimental results show that the system can effectively retrieve memories of past conversations and naturally integrate them into everyday conversations, thus allowing the robot to provide long-term personalized companionship.