Drug-Drug Interaction Extraction using Pre-training Model of Enhanced Entity Information

It is a common phenomenon to use two or more drugs during the treatments. However, a safety concern, drug-drug interaction (DDI), might occur which will cause unexpected damage to patients. Therefore, how to avoid DDI has been a major consideration all along the drug development. This paper will present a new model for extracting DDI from biomedical literature called Bio-ER-BERT, which combines the characteristics of the pre-training model BioBERT for biomedical text mining and the R-BERT model for relationship extraction. To improve the initial performance of the model, BioBERT model is introduced as the pre-training model in this paper. Meanwhile, to enhance the accuracy of DDI extraction, the R-BERT model is improved by changing the approach it obtains embedded information about two drug entities from averaging operation to long short-term memory network (LSTM). According to the experiments carried out on DDIExtraction 2013 dataset, the F1-score of our model reaches 83.88%, which has an obvious better performance comparing to the existing models such as SCNN and AB-LSTM.