Comparative Evaluation of IndoBERT and XLM-RoBERTa for Educational Sentiment Classification

Oleh :
Abdoulbassit Abbagana Oumar - K3522087 - Fak. KIP

Abdoulbassit Abbagana Oumar K3522087 COMPARATIVE EVALUATIONOF INDOBERT AND XLM-ROBERTA FOR EDUCATIONALSENTIMENT CLASSIFICATION.This study compares two transformer-based models, IndoBERT and XLM-RoBERTa, for three-class sentiment classification (positive, neutral, and negative)of Indonesian student feedback collected from the SIAKAD UNS evaluationsystem in the Informatics and Computer Engineering Education Study Program,Faculty of Teacher Training and Education, Sebelas Maret University. A total of322 feedback entries were cleaned, manually annotated by three experts usingmajority voting, and split into training and testing sets via stratified sampling. Toaddress class imbalance, a weighted cross-entropy loss was applied, and traditionalpreprocessing techniques were omitted in favor of each model’s native tokenizer.The results show that IndoBERT significantly outperformed XLM-RoBERTa.IndoBERT achieved 86.15% accuracy and a macro F1-score of 82.41%, whereasXLM-RoBERTa achieved 85.94% accuracy and a macro F1-score of 79.47%. Theneutral class proved the most challenging for both models due to limited data andsemantic overlap with mildly positive expressions. IndoBERT’s superiorperformance is attributed to its pretraining on approximately four billionIndonesian-language tokens, which enables better representation of formalIndonesian contrast, XLM-RoBERTa’s multilingual nature reduced itseffectiveness for this task. These findings demonstrate that language-specificmodels are more suitable for Indonesian educational sentiment analysis and otherdomain-specific natural language processing applications.