Abstract:
Adverse drugs reaction is potentially risky issue which may lead to illness or death in patients. Adverse Drug Event (ADE) extraction is important problem in biomedical research. Existing work mostly use neural network models or feature based pipeline methods that use just word embedding as an input features. When feature-based models are used, many feature engineering attempts must be produced. We proposed a model for entity and adverse drug event extraction labels words with the association tags of ADRs in the input order. We also embraced FastText embedding in relation to word level embedding which produced better result than word embedding and also use character level embeddings. We concatenate both FastText and character embeddings through embedding level attention mechanism. Through this mechanism we evaluates how much information come from FastText and character level embeddings. We use two classifier for output which are auxiliary classifier and main classifier. Auxiliary classifier check intermediate input of the model and combine it with main classifier output which improve the model performance. The proposed model get estimated match F1 scores 0.844 and 0.910 on twitter and PubMed dataset, respectively for ADRs identification. It presents the best results on the both dataset.
