Research on paragraph-level functional structure recognition in scientific literature: a data augmentation method based on LLMs and lexical function

Abstract The automatic recognition of functional structures within scientific literature enhances fine-grained information retrieval and mitigates issues associated with imbalanced text classification. Although multilevel functional structure research is relatively advanced, achieving high accuracy in overall label prediction and recognition, precise functional structure identification at the paragraph level remains challenging. To address this issue, we propose an innovative method called SLSG , which stands for ( S ynonym replacement + L exical function based LLM Auto-labeling + S ciBERT- G CN). This method serves as a data augmentation (DA) strategy for the identification of paragraph-level functional structure. Specifically, SLSG integrates several mechanisms, including synonym replacement and lexical function-based auto-annotation using Large Language Models (LLMs) for DA. It combines augmented data with a SciBERT-GCN model to effectively extract features by leveraging contextual sequence information between paragraphs. Applied to the ScienceDirect dataset, SLSG achieves an F1 score of 86% for paragraph-level functional structure recognition, marking an 18% improvement over the baseline models and demonstrating a significant enhancement in classification performance. Moreover, SLSG employs graph neural networks to capture both dependency relationships and topological structures among word nodes. This approach not only augments the representation of scientific literature, but establishes a solid research paradigm to address the challenges related to unbalanced text classification.