Semi-automatic Summarization of Spoken Discourse for Recording Ideas using GPT-3

Abstract

Although sticky notes are generally used to record and structure ideas stated in a face-to-face workshop, participants sometimes forget to write down their stated ideas due to the inconvenience. This study aims to apply speech recognition to record the ideas stated in a face-to-face workshop using a model created by fine-tuning the GPT-3 to choose the utterances to be recorded and then paraphrase spoken utterances. In the evaluation experiment, we compared the effects of including preceding and following utterances in addition to the summarizing and paraphrasing target utterance, and demonstrated that including only preceding utterances in addition to the summarizing and paraphrasing target utterance in the model input resulted in an F1 value over 0.8 for the selection of utterances to be recorded and ROUGE-1 up to 0.48 for paraphrasing utterance content.