Claim Missing Document
Check
Articles

Found 1 Documents
Search

Analisis Komparatif Kinerja IndoBERT dan XLM-RoBERTa Dalam Analisis Sentimen Ulasan Google Maps Mitra10 Siliwangi Pamulang Rizki Reza; Sajarwo Anggai; Abu Khalid Rivai
Bulletin of Data Science Vol 6 No 1 (2026): October 2026
Publisher : Forum Kerjasama Pendidikan Tinggi (FKPT)

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47065/bulletinds.v6i1.11071

Abstract

Google Maps customer reviews provide rich information about shopping experiences, but the growing volume of unstructured text makes manual analysis inefficient. This study classifies customer-review sentiment for Mitra10 Siliwangi Pamulang and compares IndoBERT, an Indonesian monolingual language model, with XLM-RoBERTa, a multilingual model. A total of 1,305 reviews were collected through web scraping and categorized into positive, neutral, and negative sentiment. Data processing included selection, duplicate removal, case folding, text-format normalization, tokenization, a maximum input length of 128 tokens, and dataset splitting into 1,043 training samples, 131 validation samples, and 131 test samples. Class imbalance was handled using class weighting in weighted CrossEntropyLoss without oversampling. Both models were fine-tuned with identical settings: 5 epochs, batch size 8, learning rate 2×10⁻⁵, AdamW optimizer, and random seed 42. XLM-RoBERTa achieved 96.95% accuracy, 97.25% weighted precision, 96.95% weighted recall, and a 97.07% weighted F1-score. IndoBERT achieved 95.42% accuracy, 96.90% weighted precision, 95.42% weighted recall, and a 95.88% weighted F1-score. Although XLM-RoBERTa was numerically 1.53 percentage points higher in accuracy, the exact McNemar test produced a p-value of 0.625, indicating no statistically significant difference on the test set. Both models therefore performed strongly on this dataset; however, the negative class should be interpreted cautiously because only 40 of the 1,305 reviews were negative.