Domain-Driven and Discourse-Guided Scientific Summarisation.

Scientific articles tend to follow a standardised discourse that enables a reader to quickly identify and extract useful or important information. We hypothesise that such structural conventions are strongly influenced by the scientific domain (e.g., Computer Science, Chemistry, etc.) and explore this through a novel extractive algorithm that utilises domain-specific discourse information for the task of abstract generation. In addition to being both simple and lightweight, the proposed algorithm constructs summaries in a structured and interpretable manner. In spite of these factors, we show that our approach outperforms strong baselines on the arXiv scientific summarisation dataset in both automatic and human evaluations, confirming that a scientific article's domain strongly influences its discourse structure and can be leveraged to effectively improve its summarisation. Our code can be found at: https:// github.com/TGoldsack1/DodoRank.