Drugs Reviews Sentiment Analysis using Weakly Supervised Model

The enormous amounts of vital data available in social media and health forums containing the expression of customers comments have recently provided resources for the research of Adverse Drug Reactions (ADRs) identification. However, in the existing methods, the neglect of noise data and the need for manual labeled data reduce the accuracy of prediction results and greatly increase human labor when applying social comment data. Based on the drug reviews of patients’ feedback, our objective is to recognize the sentiment inclination (positive/negative) of drug reviews by sentiment analysis for ADRs identification. Firstly, this paper proposes a weakly supervised mechanism (WSM) that applies the weakly labeled data to pre-train the parameters of the model then use labeled data to fine-tune initialed parameters, which reduces the effect of noise data on the consequences. Secondly, we present a novel architecture applys the WSM and combines the strength of Convolutional Neural Network (CNN) and Bi-directional Long Short-term Memory (Bi-LSTM) named as WSM-CNN-LSTM to complete the task for ADR reviews sentiment classification. Finally, experimental results demonstrate that our model marginally out-performes the comparison models in ADR identification and a small quantity of ADR-labeled examples are needed to attain optimal performance, which declines the influence of noise and reduces the manual labeled data.