[ODP-074] Deep learning-based relation extraction on biomedical text concerning antimicrobial plant extracts
Plant extracts contain various bioactive metabolites, and have been explored for their antimicrobial activities. The literature data on bioactivity often gives a hint to select the next candidate from a wide variety of the plants. In this research, relation extraction techniques based on deep learning were applied to biomedical text in order to extract information on antimicrobial plant extracts automatically. 600 sentences containing words related to plant extracts and microorganisms were obtained from MEDLINE database, and were manually labeled with/without antimicrobial relation. Then, the data was inputted to PCNN-ATT model (Lin Y. et al., 2016) and BERT model (Devlin J. et al., 2019) to classify the presence/absence of relations. Both models showed good classification performance (micro-F1: 0.8-0.9) in three-fold cross-validation. These results suggest the models are effective to extract antimicrobial relationships between plant extracts and microorganisms from biomedical text with speed and accuracy.