Exploiting Graph Embeddings from Knowledge Bases for Neural Biomedical Relation Extraction

Integrating external knowledge into neural models has been extensively studied to improve the performance of pre-trained language models, especially in the biomedical domain. In this paper, we explore the contribution of graph embeddings to relation extraction (RE) tasks. Given a pair of candidate entity mentions in a text, we hypothesize that the relations between them in an external knowledge base (KB) help predict whether a relation exists in the text, even if the KB relations are different from those of the RE task. Our approach consists of computing KB graph embeddings and estimating the plausibility that a KB relation exists between the candidate entities to better predict the target relation in the text. Experiments conducted on three biomedical RE tasks show that our method outperforms the baseline model PubMedBERT and achieves comparable performance to state-of-the-art methods. Our code is available at https://github.com/Bibliome/KBPubMedBERT .