Reproduce “CitationSum: Citation-aware Graph Contrastive Learning for Scientific Paper Summarization” on OpenResearch

About this paper

Zheheng Luo, Qianqian Xie, Sophia Ananiadou

CitationSum introduces a self-supervised framework for scientific paper summarization that leverages hierarchical graph contrastive learning to integrate salient full-text content and semantic correlations from references. The model achieved state-of-the-art ROUGE scores on a new biomedical dataset, PubMedCite, and existing benchmarks, demonstrating improved summary generation by effectively fusing citation graph information.

Key insights

• Hierarchical graph contrastive learning, comprising Document Representation Alignment (DRA) and Token Representation Alignment (TRA), is fundamental for deeply fusing multi-granular information from citation graphs into source document representations.
• The proposed graph contrastive learning implicitly performs matrix factorization, theoretically capturing both the semantic content of documents and the topological structure of their citation relationships.
• Leveraging high-quality salient content from full-text references and their quantified semantic correlations is significantly more effective than indiscriminately using only reference abstracts or simple content concatenation.

Summary by alphaXiv.