Comparative Analysis of Textual Sentiment Analysis Models: ML, DL & TL

Sentiment Analysis is a notable NLP task for identifying and classifying opinions and attitudes present in text. With the advent of machine learning, deep learning and transfer-learning based models, several different kinds of models were developed with distinct capability of capturing lexical patterns, sequential context and contextual meaning. In this experiment, a controlled comparative study of 7 different sentiment classification models: Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), DistilBERT and RoBERTa was performed. The experiment was executed with the similar dataset and experimental protocol as implemented in the given notebook. The data after pre-processing has a total of 33,113 text samples divided into 5 classes: Negative, Neutral, Positive, Very Negative, Very Positive. The input text is a concatenation of Generated Caption and Sentiment Description. A stratified train-test split of 80:20 was applied while maintaining the class distributions. TF-IDF feature were used for classical ML models and token sequences were input for DL recurrent networks and transfer-learning based Transformer tokenization. From experiments, accuracies obtained for each model are RF:87.06, LR:82.03, SVM:84.70, LSTM:88.77, BiLSTM:87.06, DistilBERT:96.35 and RoBERTa:83.03, resulting in DistilBERT having the highest accuracy.