The article presents a new model for relation classification (RC), a task crucial in text mining applications such as knowledge graph construction and entity interaction discovery in biomedical text. Current RC models often overlook the pattern of imbalanced predictions, which arise from having few valid relations needing positive labeling in a large set of predefined relations. The proposed model tackles these issues with a customized output architecture and additional input features. The study suggests that addressing the imbalanced predictions leads to significant performance improvements, even with modest training design. The model outperforms others on benchmark datasets commonly used in relation classification tasks.
Publication date: 26 Sep 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2309.13718
