Biomedical Relation Extraction via Syntax-Enhanced Contrastive Networks

Extracting biomedical relations from biomedical literature automatically is essential for discovering new biomedical knowledge. However, in the biomedical domain, some texts with different types have semantic similarities, which makes the differences between these types not obvious. Furthermore, lengthy and complex sentences in biomedical literature can impact the model’s ability to comprehend the long-range grammatical structure of the text. We propose a contrastive network for extracting biomedical relations that are syntax-enhanced. The model successfully highlights the distinctions between types that are semantically similar by drawing point clusters of the same kind together in the embedding space and pushing clusters of different types farther apart. Meanwhile, this model can enhance the correlation between biomedical entities while increasing the number of positive pairs and making the classification effect between different types more obvious through syntactic enhancement. Compared with other methods, the experimental findings obtained on two publicly accessible biomedical datasets demonstrate that the approach we proposed performs the state-of-the-art.