Abstract: The automatic recognition of structure function helps improve the efficiency of tasks such as fine-grained information retrieval, keyword extraction, and citation analysis. In response to the current challenges faced by structure function recognition research, including weak expression of internal textual dependencies and insufficient model generalization and transferability, this paper utilizes graph convolution neural networks to capture inherent dependency information and topological structures among word nodes, enhancing the modeling and representation capabilities of scientific publications. Additionally, adversarial learning is introduced to improve the generalization ability of the structure-function recognition model. The ScienceDirect dataset is selected to examine the recognition effectiveness of various model approaches for structure function at three different granularities: Header, Section, and Paragraph. Furthermore, we tested the transferability of multiple models across domains on PubMED-20k, a medical abstract structure function recognition dataset. Experimental results demonstrate that BERT+GCN get the best performance at the Header level, with an value of 88%, which is a 3% improvement over baseline models. At the Section level, the combination of BERT and GAN achieves the best performance, which is also a 3% improvement over baseline models. At the section paragraph level, the score reaches 68%. BERT+GCN exhibits superior cross-domain transferability compared to other models, achieving an score of 90% on cross-domain data.

Key words: Functional Structure, Graph convolution network, Generative adversarial networks, Scientific literature, Information recognitio