ScholarSum: Student-Teacher Abstractive Summarization via Knowledge Graph Reasoning and Reflective Refinement

Bohou Zhang, Mingyue Cheng, Huijie Liu, Shulan Ruan, Qi Liu

Keywords: Scientific Summarization, Knowledge Graph

Abstract:

Abstractive summarization of scientific papers is essential for efficient knowledge access.
Although numerous approaches have been proposed, they often fail to capture the logical structure of the scientific paper, omit key factual information, and may produce hallucinated content.
In this work, we propose ScholarSum, a Student–Teacher framework inspired by the human writing process, including drafting, reviewing, and revising.
First, to capture paper structure, the student module constructs a knowledge graph based on the paper, divides it into semantic subgraphs, and performs graph-based reasoning to produce drafts aligned with the paper structure.
Second, to improve coverage in long contexts, the student module retrieves key fact triplets from the global graph and integrates them into the draft, minimizing the loss of key factual information.
Third, to strengthen factual fidelity, the teacher module conducts quality assessment via prompting and reference-guided reflection. Based on the assessment outcome, the module selects acceptance, minor revision, or regeneration.
The collaborative design enables dynamic quality control, improving structural coherence and ensuring both factual completeness and accuracy.
Experimental results on scientific summarization benchmarks demonstrate that ScholarSum consistently outperforms strong baselines, producing summaries that are structurally coherent, factually comprehensive, and well aligned with human-written reference summaries. Our code is available at https://anonymous.4open.science/r/ScholarSum-Anonymous.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 12069