DeepFusionSent: A novel feature fusion approach for deep learning-enhanced sentiment classification

Sentiment Analysis (SA) is the process to identify, extract, and quantify subjective information from text data, such as opinions, attitudes, emotions, and sentiments. Traditional approaches to sentiment classification primarily relied on utilizing raw data, which may not fully capture the detail insights within the dataset. Here, we propose an approach that seeks to enhance sentiment classification performance by leveraging feature fusion derived from underlying dataset characteristics rather than considering the entire raw feature subset. Our proposed approach incorporates feature fusion like average distance between positive and negative words, average number of positive words immediately followed by negative words, average number of negative words immediately followed by positive words, average number of positive words consecutively followed by negative words, and average number of negative words consecutively followed by positive words. We employ these features instead of using the entire raw feature subset. The study aimed to enhance sentiment classification by comparing its performance with existing methods. The existing methods include various Deep Learning (DL) techniques like Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Network (CNN)-LSTM (CNN-LSTM), and CNN-BiLSTM. Moreover, different Machine Learning (ML) techniques namely, N & auml;ive Bayes (NB), Logistic Regression (LR), Decision Tree (DT), K-Nearest Neighbor (KNN), and AdaBoost models are also implemented for comparative result analysis. This study also compares the performance of BERT and RoBERTa models with the proposed approach. Moreover, study also compare the results of the proposed approach with traditional word embedding techniques, where the complete feature subset is employed. Additionally, we assess the statistical significance of the results. The experimental findings indicate that the proposed fusion feature approach surpassed the performance of the existing methods.