PatentALL: Multi-label Patent Classification Using Adaptive Label Learning

Patent classification codes, such as IPC, serve as labels for both manual and automatic patent classification. We argue that the semantic relationships between these labels can provide valuable additional information beyond the patent text for neural patent classification. In this paper, we propose PatentALL, a patent classification framework that integrates both classification codes and patent texts. Our approach includes an adaptive label learning mechanism based on Graph Convolutional Networks (GCNs) to capture long-tail and cross-category relationships among classification codes. This mechanism involves re-learning label representations to refine the relationships between different classification codes. We combine these refined label representations with patent text representations obtained using BERT, facilitated by an attention mechanism. Additionally, we introduce a SuperLoss function for supervised contrastive learning in our model. Experimental results demonstrate that our proposed patent classification model outperforms baseline models on the USPTO-2M and M-patent datasets. The adaptive learning mechanism excels particularly in handling long-tail labels and capturing intricate semantic relationships across different categories of labels. Furthermore, our findings highlight the effectiveness of contrastive learning in enhancing patent classification accuracy.