ロールプレイを用いたチャットグループの集合場所の推薦

抄録

In this paper, we consider a service in which AI can infer the situation and requests of members from conversation (hereafter referred to as chat) history and recommend a meeting place. The problem of deciding where multiple entities will meet has been widely studied as a facility placement problem or a rendezvous problem, but most of these studies have been conducted after formulating constraints and costs. For this reason, it has not been easy to decide on a meeting place for entities whose requests when meeting are difficult to incorporate into the format. Based on this, in this paper, we propose a method to apply a large-scale language model (LLM) to directly recommend a meeting place from the entities' chat history without going through such a formulation. However, although it is expected that the requests and values of each entity can be read using LLM, there is no way to consolidate them. In this paper, we achieve this through role-playing. Specifically, role-playing is performed based on the persona estimated from the chat, and a meeting place that reflects the values that match them is recommended. Through computer experiments, we were able to recommend meeting places according to the types of people involved and the reason for their meeting, and the reasons for the recommendations, which have traditionally tended to be a black box, became clear. Furthermore, human surveys verified that the recommendations were appropriate to a certain extent, clarifying current issues.