Drug-Drug relationship extraction based on entity information and graph neural networks

Abstract:

Drug-Drug interaction refers to the mutual promotion or inhibition between drugs. For the existing drug relationship extraction methods, the use of external background knowledge and natural language processing tools leads to the problem of error propagation and accumulation, and most existing studies blind drug entities at the data preprocessing stage, ignoring the target drug entity information that is helpful to identify the relationship category. In this paper, a drug interaction extraction model based on pretrained biomedical language model and word map neural network is proposed. In this model, the original feature representation of sentences is obtained by pretrained language model, and the global feature information representation of sentences is obtained by convolution operation on the word map constructed based on data set. Finally, the feature representation of drug interaction relationship extraction task was constructed by stitching the feature with drug target entities, which can not only obtain rich global feature information but also avoid using natural language processing tools and external background knowledge, and improve the accuracy of the model. The F1 value of the model on the DDIExtraction 2013 dataset achieved 83.25%, which outperforms the current latest methods by 2.35%.