Research On Adverse Drug Reactions Text Classification And Labeling Based On RNN

Adverse drug reactions are one of the main causes of increased morbidity and mortality in patients.Therefore,the rapid and accurate selection of samples containing adverse drug reactions from the text and labeling of the adverse drug reactions contained therein are of great practical significance for drug regulatory and biomedical research.Because the recurrent neural network is very good at processing sequence data,it has become a very important method in the classification and labeling of adverse drug reactions text.For the text classification of adverse drug reactions,we propose a hierarchical attention mechanism model based on RNN,which converges faster and has higher classification accuracy than the previous text classification models.For the labeling of adverse drug reactions,we propose an RNN-based end-to-end sequence labeling model,in addition to word-level embedding features,we also introduce the character-level embedding features,and through an embedded-level attention mechanism,the model can dynamically learn which level of features are more important to the results during the training process.Finally,we also use the intermediate output of the model as an auxiliary classifier and combine it with the final output of the model to further improve the labeling performance of the unknown words in the adverse drug reactions text.On the PubMed adverse drug reactions text classification dataset,our model obtained the accuracy of 0.851,which had a great performance improvement compared with the previous text classification models.We evaluated our model on two adverse drug reactions datasets,one is collected from tweets which contain many informal vocabulary and irregular grammar.Another dataset adopts abstracts extracted from PubMed biomedical papers which have many technical terms and descriptions.Our model has obtained the F1 score of 0.844 and 0.906 respectively on these two datasets,which are the state-of-the-art performance of the existing models.Since our model is end-to-end and does not require extra feature engineering,it can be easily generalized to any text classification and sequential labeling tasks.