Klasifikasi Sentimen Lintas Domain Pada Komentar Media Sosial Menggunakan Indobert Dengan Finetuning Dan Data Augmentation

Abstract

Sentiment analysis on social media comments in Indonesia faces significant challenges due to the complexity of non-standard language and the imbalance of sentiment class distribution. Pre-trained language models such as IndoBERT, trained on formal corpora, are proven to experience generalization failure (domain mismatch) when applied directly to this domain. This study aims to optimize IndoBERT’s performance through a fine-tuning strategy on a multi-domain secondary dataset, as well as conducting a comparative study on Synonym Replacement (SR) and Back-Translation (BT) augmentation techniques to address class imbalance. Evaluation was rigorously conducted using a combined held-out domain test dataset. Experimental results indicate that fine-tuning is a fundamental step, where the baseline model failed to recognize sentiment (F1- Score 0.3523), whereas the adapted model achieved an F1-Score of 0.5994. The study’s main findings reveal that Synonym Replacement (SR) augmentation is the most optimal strategy, recording an accuracy of 69.67% and a Weighted F1-Score of 0.6714. Conversely, the Back-Translation (BT) technique, despite being superior in minority class balance, was identified to trigger overfitting due to semantic noise. This study concludes and recommends the combination of fine-tuning with SR augmentation as the most robust and stable approach for handling crossdomain sentiment on imbalanced Indonesian text data.