Relation Extraction in Biomedical Texts: A Cross-Sentence Approach

Zhijing Li,Liwei Tian,Yiping Jiang,Yucheng Huang

DOI: https://doi.org/10.1109/TCBB.2024.3451348

2024-09-06

Abstract:Relation extraction, a crucial task in understanding the intricate relationships between entities in biomedical domains, has predominantly focused on binary relations within single sentences. However, in practical biomedical scenarios, relationships often extend across multiple sentences, leading to extraction errors with potential impacts on clinical decision-making and medical diagnosis. To overcome this limitation, we present a novel cross-sentence relation extraction framework that integrates and enhances coreference resolution and relation extraction models. Coreference resolution serves as the foundation, breaking sentence boundaries and linking entities across sentences. Our framework incorporates pre-trained deep language representations and leverages graph LSTMs to effectively model cross-sentence entity mentions. The use of a self-attentive Transformer architecture and external semantic information further enhances the modeling of intricate relationships. Comprehensive experiments conducted on two standard datasets, namely the BioNLP dataset and THYME dataset, demonstrate the state-of-the-art performance of our proposed approach.