关键词: 科学文献摘要, 抽取式摘要, 图神经网络, 超图注意网络, 图对比学习

Abstract: Summarizing long documents has been identified as a challenging task in natural language processing due to difficulties in capturing inter-sentence relationships, handling long-range dependencies, and efficiently encoding and extracting document information. Scientific documents, characterized by multiple chapters and paragraphs with complex hierarchical structures, further increased the difficulty of the summarization task. To address these challenges, a scientific document summarization model based on multi-graph neural networks and graph contrastive learning, named MGCSum, was proposed. First, for a given input document, intra-sentence and inter-sentence relationships were modeled using homogeneous and heterogeneous graph neural networks, respectively, to generate initial sentence representations. These representations were subsequently fed into a multi-head hypergraph attention network, where self-attention mechanisms were leveraged to capture relationships between nodes and edges, enabling the updating and learning of cross-sentence representations. Then, a graph contrastive learning module was introduced to enhance global topic awareness, improving the semantic consistency and discriminability of sentence representations. Finally, a multi-layer perceptron and a normalization layer were applied to calculate scores for determining whether sentences should be selected for summarization. The experimental results on the PubMed and ArXiv datasets indicate that the MGCSum model outperformed most baseline models. On the PubMed dataset, MGCSum achieved ROUGE-1, ROUGE-2, and ROUGE-L scores of 48.97%, 23.15%, and 44.09%, respectively, with improvements of 0.2, 0.71, and 0.26 percentage points over the state-of-the-art model HAESum. By integrating multi-graph neural networks and graph contrastive learning, MGCSum effectively captures hierarchical structural information and inter-sentence relationships, enhancing the accuracy and semantic consistency of summarization, and demonstrating its advantages in scientific document summarization tasks.

Key words: scientific document summarization, extractive summarization, Graph Neural Network (GNN), HyperGraph ATtention network (HGAT), Graph Contrastive Learning (GCL)