Dynamic Slot-Making-and-Filling Method for Improving Long-term Dialogue Consistency

Abstract

Maintaining coherence in utterances over an extended period is challenging in dialogue systems, even with current technology. In scenarios where tasks can be predefined, an approach known as ``Slot (Dialogue State)'' has been proposed to enhance long-term coherence. This involves predefining slots for necessary task-related items and updating these slots' values as the conversation progresses. The objective of this study is to propose a method for maintaining long-term coherence in situations where it is challenging to align with changes in the dialogue context, such as role-playing in vocational training or tabletop role-playing games (TRPG), which conventional methods find difficult to address. In this research, we focus on TRPG game masters and create a system utilizing GPT's output as external dynamic slots. We demonstrate the effectiveness of this approach. Results indicate the proposed method's superiority in terms of agreement with human evaluation.