Team Matus and Francesco @ AutoMin 2021: Towards Neural Summarization of Meetings

As online meetings are becoming increasingly ubiquitous, there is an increasing demand to record the main outcomes of these meetings for future reference. Automatic summarization of meetings is a challenging, yet relatively unex-plored natural language processing task with a wide range of potential applications. This paper describes our submission to the First Shared Task on Automatic Minuting at Interspeech 2021. In contrast to previous research focused on the summarization of narrated documents, we examine the specifics of bullet-point spoken language summarization on the AutoMin dataset of online meetings in English. Furthermore, we investigate whether existing abstractive summarization systems can be transferred to this new domain. In this regard, we develop a minuting pipeline based on the state-of-the-art PEGASUS summarization model. This includes pre-processing of conversational data, few-shot transfer learning using reference minutes generated by human annotators, filtering and post-processing of the resulting candidate summaries into a suitable bullet-point minutes format. We conclude by evaluating the completeness and shortening aspects of our system, and discuss its limitations and potential future research directions.