Team AutoMinuters @ AutoMin 2021: Leveraging state-of-the-art Text Summarization model to Generate Minutes using Transfer Learning

Abstract

This paper presents our submission for the first shared task of automatic minuting (AutoMin@Interspeech 2021). The shared task consists of one main task generate minutes from the given meeting transcript. For this challenge, we leveraged state-of-art text summarization models to generate minutes using the transfer learning approach. We also provide an empirical analysis of our proposed method with other text summarization approaches. We evaluate our system submission quantitatively with 33% BERTscore and 11.6 % ROUGE L, which is rela-tively higher than the average submission in the shared task. Along with the qualitative evaluation, we also vouch for quantitative assessment, where we achieve (2.32, 2.64, 2.52) scores out of five for adequacy, grammatical correctness, and fluency. For the other two tasks, we use Jaccard and cosine text similarity metrics for a given transcript-minute pair corresponding to the same meeting (Task B) and if a given pair of meeting minutes belong to the same meeting (Task C). However, our simple approach yielded 94.8 % (task B) and 92.3% (task C), clearly outperforming most submissions in the challenge. We make our codebase release here https://github. com/mahajanparth19/Automin_Submission .