LLM-Based Structuring of Oral Discussion in Workshop to Support Collaboration Among Local Government and Simulated Citizens

In oral discussions, such as workshops, notes and sticky notes are frequently used to record content of the utterances. However, the content of the discussion may not be accurately captured or be forgotten through manual writing. Although the results from speech recognition could be used, they are difficult to understand because they are redundant and consist of fillers. This study addresses this problem by recording the speech-recognition results into text, similar to writing on a sticky note. The text should be well-structured for participants to understand. The objective of this research is to develop a discussion structuring system for use in real time and demonstrate its utility as a substitute for notes. This system will enable users to easily reflect on the content and process of discussions. The large language model GPT-4 is used for paraphrasing speech-recognition results into simple sentences and structuring them. It will also be used for opinion generation to incorporate novel viewpoints to a target discussion. A discussion experiment using a prototype of this system was conducted for a workshop involving local government staff. It was shown that the system could serve as an alternative to sticky notes when people reflect on the discussions of the group and the artificial-intelligence-generated opinions from the system are effective in supporting discussion.