Enhancing Sentiment Analysis Models through Multi-Technique Data Augmentation: A Study with IndoBERT

In deep learning-based models, such as those used for sentiment analysis, one critical challenge is ensuring robust and generalized performance. While data augmentation has been identified as a key strategy for enriching datasets, its effectiveness in terms of improving model generalization, reducing overfitting, handling data imbalances, capturing contextual variations, and decreasing sensitivity to label noise has yet to be comprehensively studied. This study conducted several experiments involving three distinct data augmentation techniques. The primary goal is to prepare training data for sentiment analysis models using the fine-tuning method with the IndoBERT model. The techniques include Sequence Generative Adversarial Network (SeqGAN), Easy Data Augmentation (EDA), and An Easier Data Augmentation (AEDA). We have conducted various experiments by combining augmentation techniques. The highest macro-average F1-score was obtained from the combination of augmentation experiments using the EDA, AEDA, and SeqGAN methods, reaching 93.61%. This result represented an improvement compared to the baseline model without data augmentation, which achieved an F1-score of 92.45%. These figures show that our approach gives a promising result and can be implemented using another dataset.