Multimodal Fusion of BERT and ResGCN for Chemical Patent Classification

The chemical industry’s massive scale makes patent classification crucial for technology landscape analysis. With rapid global patent growth, traditional manual IPC classification is inadequate. Chemical patents are highly specialized and suffer from class imbalance, leading to insufficient classification accuracy. This study proposes a multimodal framework integrating BERT and Residual Graph Convolutional Network (ResGCN), which models chemical entity associations via heterogeneous graphs and employs residual connections to mitigate over-smoothing. The methodology includes: (1) using ChemDataExtractor to extract 5 entity types and 6 semantic relationships; (2) designing a ResGCN–BERT dual-channel architecture with differentiated learning rates for stable training; (3) evaluation on the 2775-sample, 7-class PETROCHEMICAL dataset. Experiments demonstrate 82.34% accuracy and 0.7782 macro-average F1, 3.29 percentage points higher than the BERT baseline. Ablation studies show residual connections boost minority-class F1 by 4.2%, providing a scalable framework for intelligent chemical patent examination.