A Knowledge-enhanced model with syntactic-aware attentive graph convolutional network for biomedical entity and relation extraction

Biomedical entity and relation extraction has recently gained increasing interest. Nevertheless, it continues to pose a significant challenge due to the unique domain-specific characteristics of the biomedical corpus. Biomedical documents often include many highly specialized terms, acronyms, and abbreviations. This presents a significant challenge for those entity and relation extraction models that neglect biomedical domain-specific knowledge, as they may show biases in interpreting biomedical texts and may fail to accurately associate these specialized terms with the appropriate biomedical entities. Moreover, most existing models struggle to process lengthy sentences in the biomedical corpus, which hinders their performance. To address these limitations, this paper proposes a Knowledge-Enhanced Model with Syntactic-Aware Attentive Graph Convolutional Network (KESAAGCN) for biomedical entity and relation extraction. Given an input text, KESAAGCN first constructs an initial span graph representing its initial understanding of the text. Specially, a syntactic-aware attentive graph convolutional network based on the dependency trees of the sentences is utilized to capture relations between long-distance entities in lengthy sentences. Then, a domain knowledge graph is constructed based on an external knowledge base to incorporate useful domain-specific information into the model. Finally, the initial span graph and the domain knowledge graph are fused to obtain a more refined graph for final prediction. Extensive experimental results on three benchmark datasets demonstrate the effectiveness and superiority of our methods.