Search papers, labs, and topics across Lattice.
This paper introduces ScholarSum, a novel hierarchical reflective graph-based framework designed to enhance the quality of abstractive summarization in scientific literature by balancing linguistic fluency with factual accuracy. The method organizes documents into a hierarchical knowledge graph, allowing a student model to generate an initial draft that is then refined through iterative teacher-like review and evidence retrieval. Experimental results show that ScholarSum significantly surpasses existing methods in both completeness and faithfulness, addressing critical limitations in current summarization approaches.
ScholarSum achieves a groundbreaking balance between fluency and factual accuracy in scientific summarization, outperforming existing methods by a significant margin.
Abstractive summarization plays a crucial role in enabling efficient understanding of scientific literature, yet it inherently demands both linguistic fluency and factual faithfulness. Existing approaches often fail to reconcile these two requirements. Extractive methods rely on rigid sentence splicing that disrupts macro-level logical coherence, while large language model (LLM)-based generative approaches, despite mastering linguistic fluency, exhibit limited factual consistency. In this work, we propose ScholarSum, a hierarchical reflective graph-based framework that emulates a student-teacher writing process for fluent and faithful scientific summarization. ScholarSum first organizes the document into a hierarchical knowledge graph by segmenting it into semantically coherent units, whose multi-layered community structure captures global logic and macro-level themes. Guided by this global structure, the student generates an initial draft, which is subsequently refined through fine-grained evidence retrieval. To ensure factual consistency, a teacher-like reviewer then iteratively examines the draft, identifies unsupported content, and prompts targeted re-retrieval and rewriting until the summary meets rigorous quality standards. Extensive experiments demonstrate that ScholarSum significantly outperforms previous baselines in terms of both completeness and faithfulness. Our code is available at https://github.com/Xiaoyu-Tao/ScholarSum.
