Team UEDIN @ AutoMin 2021: Creating Minutes by Learning to Filter an Extracted Summary

Abstract

We describe the University of Edinburgh’s submission to the First Shared Task on Automatic Minuting. We developed an English-language minuting system for Task A that combines BERT-based extractive summarization with logistic regression-based filtering and rule-based pre- and post-processing steps. In the human evaluation, our system averaged scores of 2.1 on adequacy, 3.9 on grammatical correctness, and 3.3 on fluency.