Large-Scale Kannada Sentiment Classification Using TF–IDF and Attention-Guided CNN- BiLSTM Networks: An Empirical Study Review

Abstract

Sentiment classification is still very difficult in Kannada due to its morphological richness, inflectional variations, compound formations, spelling variations and lack of large Kannada labelled corpora. The present study introduces a large-scale empirical comparison between conventional machine-learning and attention-guided deep-learning models with a binary Kannada review dataset with 119,952 labelled samples. The collected data set consisted of 58,797 negative reviews and 61,155 positive reviews, giving an almost balanced experimental environment. Three classical classifiers: TF–IDF with logistic regression, linear support vector machine and multinomial naïve Bayes were compared with attention-guided bidirectional long short-term memory network and convolutional neural network–BiLSTM. A stratified partition (80:20) resulted in 95,961 training and 23,991 independent test observations. The accuracy, macro-F1 and ROC–AUC of logistic regression were 82.19%, 82.18%, and 0.9054, respectively, which were higher than the other classical methods. The attention guided BiLSTM model was able to obtain nearly 86% accuracy and macro-F1 while the CNN–BiLSTM model obtained nearly 87% accuracy and class-balanced F1-scores. The latter boosted test accuracy by ~4.8 percentage points compared to the best TF–IDF baseline. According to the results of the architectural analysis, the CNN–BiLSTM needed around 2.91 million parameters and 14.47 million multiply–accumulate operations per review. Although the temporal pooling model had a few more parameters than the standalone BiLSTM, it had less sequential computational complexity, and showed faster inference when executed in the common notebook setting. The results show that the combination of local convolutional feature extraction, bidirectional sequence modelling and attention pooling is effective in the case of large-scale Kannada sentiment classification without the need for a computationally-intensive pretrained transformer.