Dealing with Class Imbalance in Sentiment Analysis using Deep Learning and SMOTE

Abstract

In textual data, sentiments or opinions expressing polarities (positive or negative) often form the basis for human decision-making. Therefore, sentiment analysis has always been an important area of research in the field of artificial in- telligence. Recently, deep learning models have been used for the sentiment analy- sis. However, the class imbalance of the dataset adversely affects the performance of these models. To address this issue, the paper presents a method to resample the dataset using Synthetic Minority Over-sampling Technique (SMOTE). The pro- posed method is applied to three different datasets of customer reviews, each of which exhibits different class imbalance ratios. To show the impact of the method, the modern Recurrent Neural Network (RNN) based architectures (LSTM, GRU, and Bi-directional) are trained with both the originally imbalanced datasets and the SMOTE balanced datasets, and comprehensively analysed the performance of these approaches. The obtained results show that the models trained with balanced datasets outperform other methods across most models with significant improve- ments over the original dataset.