Team Symantlytical @ AutoMin 2021: Generating Readable Minutes with GPT-2 and BERT-based Automatic Minuting Approach

Abstract

This paper describes our participation system run to Automatic Minuting @ Interspeech 2021 1 . The task was motivated towards generating automatic minutes. We make a initial step towards, namely Main Task A , Task B and Task C . The main task A, was to automatically create minutes from multiparty meeting transcripts, while task B to identify whether the minute belongs to the transcript and task C. GPT-2[1]. The shared task, consist-ing of three subtasks, required to produce, contrast and scruti-nize the meeting minutes. The process of automating minuting is considered to be one of the most challenging tasks in natural language processing and sequence-to-sequence transforma-tion. It involves testing the semantic meaningfulness, readability and reasonable adequacy of the Minutes produced in the system. In the proposed work, we have developed a system using pre-trained language models in order to generate dialogue summaries or minutes. The designed methodology considers cov-erage, adequacy and readability to produce the best utilizable summary of a meeting transcript with any length. Our evaluation results in subtask A achieve a score of 11% R-L which by far is the most challenging than subtask as it required systems to generate the rational minutes of the given meeting transcripts.