Studi Komparasi IndoBERT dan IndoBERTweet dengan Full Fine Tuning dan LoRA untuk Klasifikasi Emosi pada Tweet Bertopik GERD

Abstract

This study compares IndoBERT and IndoBERTweet for emotion classification on Indonesian tweets related to Gastroesophageal Reflux Disease (GERD). The term GERD-related tweets is used because the data were collected from social media X, and the clinical status of each user cannot be verified. The initial dataset contained 1,389 records. After validation, text cleaning, empty-text removal, label filtering, and deduplication, 1,343 tweets remained and were manually assigned into five emotion classes: angry, fear, joy, sad, and neutral. The experiments were conducted in Python on Google Colab with a Tesla T4 GPU using Pandas, PyTorch, Hugging Face Transformers, PEFT, scikit-learn, Matplotlib, and Seaborn. The models were evaluated using Stratified 5-Fold Cross Validation, lexical substitution-based data augmentation on the training data of each fold, early stopping, and Focal Loss to address class imbalance. The results show that full fine-tuning achieved the best performance. IndoBERTweet obtained the highest accuracy of 0.5703 and macro F1-score of 0.5629, while IndoBERT achieved an accuracy of 0.5145 and a macro F1-score of 0.5034. In contrast, IndoBERT + LoRA only achieved a macro F1-score of 0.2412, and IndoBERTweet + LoRA achieved 0.2227. These findings indicate that LoRA was not effective for this small and imbalanced health-emotion dataset, although it trained less than 1% of the model parameters. Parameter efficiency therefore did not translate into better classification performance in informal and semantically overlapping social media text.