[2D1-GS-9] Human interface:

[2D1-GS-9-01] Proposal of a Topic Selection RAG for Developing Proactive Chatbots

〇Kazuya Matsuo¹, Atsushi Otsuka¹, Narichika Nomoto¹, Makoto Nakatsuji¹ (1. NTT Human Informatics Laboratories)

Keywords:dialogue system, RAG

In the domain of chatbots, maintaining conversational willingness is a significant challenge. Existing chatbots are passive and respond reactively to user input. This approach places the burden of generating conversation topics on the users, which reduces their willingness. To address this issue, we propose a chatbot capable of proactively providing a new topic that aligns with the user's interests, referencing external information. While RAG is commonly utilized for generating texts with external references, it struggles to handle ambiguous queries such as "retrieving new information of interest to the user." Consequently, we also propose a novel RAG tailored to this specific query type. Our RAG divides the query into two parts: acquiring themes consistent with the user's interests and retrieving information unknown to the user. To achieve these objectives, our RAG manages external information and users' interests and knowledge, which are estimated from previous interactions, based on themes. It then retrieves a theme that matches the user's interests and retrieves information from external sources related to this theme that the user does not already know. We evaluate the performance of our RAG, demonstrating its ability to more accurately obtain new information of interest to users compared to existing RAG systems.