Automatic Rephrasing of Transcripts-based Action Items

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2862–2873 August 1–6, 2021. ©2021 Association for Computational Linguistics 2862 Automatic Rephrasing of Transcripts-based Action Items Amir DN Cohen1,2, Amir Kantor1, Sagi Hilleli1 and Eyal Kolman1 Microsoft1 Department of Computer Science, Bar-Ilan University2 amirdnc@gmail.com,{amir.kantor,sagih,eyalko}@microsoft.com Abstract The automated transcription of spoken lan- guage, and meetings, in particular, is be- coming more widespread as automatic speech recognition systems are becoming more accu- rate. This trend has significantly accelerated since the outbreak of the COVID-19 pandemic, which led to a major increase in the number of online meetings. However, the transcrip- tion of spoken language has not received much attention from the NLP community compared to documents and other forms of written lan- guage. In this paper, we study a variation of the summarization problem over the transcrip- tion of spoken language: given a transcribed meeting, and an action item (i.e., a commit- ment or request to perform a task), our goal is to generate a coherent and self-contained rephrasing of the action item. To this end, we compiled a novel dataset of annotated meet- ing transcripts, including human rephrasing of action items. We use state-of-the-art super- vised text generation techniques and establish a strong baseline based on BART and UniLM (two pretrained transformer models). Due to the nature of natural speech, language is of- ten broken and incomplete and the task is shown to be harder than an analogous task over email data. Particularly, we show that the baseline models can be greatly improved once models are provided with additional informa- tion. We compare two approaches: one in- corporating features extracted by coreference- resolution. Additional annotations are used to train an auxiliary model to detect the relevant context in the text. Based on the systematic hu- man evaluation, our best models exhibit near- human-level rephrasing capability on a con- strained subset of the problem. 1 Introduction Most of natural language processing (NLP) re- search focuses on written language, such as emails or Web pages, and less on an increasingly large body of spoken language converted to text via au- tomatic speech recognition (ASR). Particularly, to- day, more and more meetings are conducted online, especially since the COVID-19 social distancing constraints. Online meetings may be transcribed upon request, generating a huge amount of spoken language text.1 Spoken language has different characteristics from written language, and, from our experience, it is typically vaguer and harder to understand: sen- tences tend to be broken, less orderly, incomplete (relying on subtext), and prone to speech-to-text transformation errors (see Section 5). Action items (AIs) are a common and particularly important part of workplace meetings. An AI is a commitment or a request to perform a certain task by any of the parties involved. For example, “I will send you the file later today.” AIs occur naturally during conversations, but are not always clear without relevant context. For ex- ample, the AI “I will do it” contains a commit- ment to act, but the nature of this action …

Content truncated to 500 words