RETELL: RST-Centric Related Work Generation with LLM

ACL ARR 2025 May Submission4024 Authors

Abstract: Related work cherishes structural relationships. While existing automated methods, including those leveraging Large Language Models (LLMs), have advanced content summarization capabilities, they often struggle to replicate the crucial argumentative flow and explicit inter-paper relational structures found in human-written related work, despite many commendable efforts in recent years. To address this concern, we propose a novel approach centered on Rhetorical Structure Theory (RST). We introduce a structure-aware Related Work Generation (RWG) pipeline where LLM agents are guided by an RST-derived structural plan to generate related work with iteratively improved structural relationships. Finally, we craft two structure-specific metrics, Width Profile Similarity (WPS) and Edge Coverage Ratio (ECR), to evaluate the coherence of the generated related work. Through extensive experiments, we demonstrate that our RST-centric generation method significantly enhances the structural and overall quality of RWG.

Paper Type: Long

Research Area: Summarization

Research Area Keywords: multi-document summarization, evaluation and metrics

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 4024