Abstract
Automatic extracting inter-sentential relations between adverse drug event (ADE) entities from unstructured documents is one of the crucial tasks for pharmacovigilance. Significant advances have been made for handling this task with the rapid development of deep neural models in recent years. However, several challenges remain to be addressed. Among these, the state-of-the-art systems do not leverage information about relations from the entities, sentences, and documents simultaneously so as to learn entity-pair representation more efficiently. Besides, there can be a very large number of entity pairs with no relations, and only a few ones have predefined relations, also known as the sparsity problem. To alleviate the aforementioned problems, we propose, in this paper, a Gated Multi-hop Attention Fusion Network for extracting inter-sentential ADE relations, called GMAFNet. Firstly, the GMAFNet system extracts features from three levels: (1) entity-level feature with an adaptive localized context pooling, (2) sentence-level feature with an attentive hierarchical context pooling, and (3) document-level feature with an attentive global-local aggregation context pooling. Then, the system fuses the obtained features with our gated multi-hop attention fusion strategy. Finally, the system uses an asymmetric focal loss to overcome the sparsity issue. Experimental results conducted on the benchmark dataset from the n2c2 2018 challenge show that GMAFNet performs better than the cutting-edge systems, achieving an F1-score of 97.27%. The results further demonstrate the generalizability of GMAFNet on two well-known datasets, reaching an F1 of 78.19% and 87.10% on the CDR and GDA datasets, respectively. These findings highlight the effectiveness of our system for extracting complex inter-sentential ADE relations. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s41666-025-00214-8.