Sentiment Analysis with RNNs: Insights from LIME Interpretability

This research article explores different facets of sentiment analysis like data pre-processing, extracting useful features, model designs & evaluation criteria. We explore recurrent neural network (RNN), having two layers of Bidirectional LSTM, to classify user language into positive, and negative classes. Our experiment involves training these models on large-scale sentiments-based datasets and evaluating their performance by extending more features and techniques of deep learning. We compare and visualize accuracy with respect to accuracy score, confusion matrix, precision-recall curve as well as ROC curve. The results demonstrate that the BiLSTM model achieves competitive sentiment classification accuracy. The classification trust concern is addressed using LIME, an interpretable machine learning model. The major contribution of this research is towards the field of sentiment analysis by evaluating the model and emphasizing importance of explainability.