Pragmatically Appropriate Abstractive Summarization of JTAC Radio Conversations
Abstract
In this work, I explore the development of computational methods for automatically creating after-action reports of JTAC radio conversations. Prior research has investigated related issues of sentence compression, text summarization, and conversation summarization (Banerjee, Mitra, & Sugiyama, 2015; Clarke & Lapata, 2008; L. Wang & Cardie, 2012; Raffel et al., 2020). However, this work makes limiting assumptions about what features are relevant to a summary and what sources of information should be included. I propose methods that combine knowledge from linguistic, procedural, and domain sources to address these limitations. Results indicate that the proposed model performs better than some of our baseline models, but highlight several challenges of automatically scoring computational summaries, particularly in technical low-resource domains.
