Social Media Adverse Drug Reaction Detection Based on Bi-LSTM with Multi-head Attention Mechanism

Social media text contains a large amount of adverse drug reaction (ADR) information, which is an important channel for ADR information extraction. It is difficult to extract the detection features from the social media text by the traditional ADR extraction methods. Convolutional neural network (CNN) and its variants have the disadvantages of low modeling efficiency and space insensitivity when constructing spatial information. An ADR detection method from social media text is proposed based on bidirectional short and long time memory network (Bi-LSTM) with multi-head attention mechanism (MHAM). After pretreatment on corpus to markers of adverse drug reactions, and constructs the distributed word vector features, part-of-speech tags, character vector and each sentence Chinese medicine things and emotional words as the characteristics of the model input, contrast experiments, to solve the lack of space relationship between features in the process of classification and building the model of the problem of low efficiency. Experimental results on SMM4H corpus validate that the proposed method is effective and has good performance in the detection of ADR events in social media. To improve the detection efficiency of ADRs, the multi-head attention mechanism is introduced bi-directional long short-term memory (Bi-LSTM). Experiment results on Social Media Mining for Health (SMM4H-2017) corpus dataset validate the proposed method can significantly improve the recognition and classification performance of ADRs.