Background: The cause of the disease is one of the main contents of biomedical research. Extracting effective relational information from a large number of biomedical texts has important applications for biomedical research. At present, most of the work of biomedicine is to use manual screening or use rule-based or feature-based pipeline network models to obtain screening characteristics. These methods require a lot of time to design specific rules or features to complete specific tasks, resulting in some features that non-compliant features cannot be filtered out.
Results: The model gets micro-F1 scores of 0.802 and 0.876 on the Chemprot data set and DDI data set, respectively. The resources that can be used in this project can be found in https://github.com/HunterHeidy/DDICPI-.
Conclusions: Experiments have proved that without Bert, you can get good results by learning from Bert<s core ideas.