<![CDATA[Large Language Model (LLM) applications such as ChatGPT, Gemini, Copilot, Claude, and Perplexity have been massively adopted in Indonesia, yet user experience evaluation remains largely limited to global sentiment analysis. This study implements Aspect-Based Sentiment Analysis (ABSA) using a dual-Transformer approach: DeBERTa zero-shot for aspect extraction and IndoBERT for sentiment classification on 5,000 Indonesian-language reviews from the Google Play Store across four aspect categories. Manual validation by two annotators on 300 samples yielded Cohen’s Kappa of (aspect) and (sentiment), both Moderate. Evaluation against the gold standard showed aspect accuracy of 37.5% (weighted F1 = 0.42) and sentiment accuracy of 64.7% (weighted F1 = 0.61). Sensitivity analysis across five hypothesis templates revealed inter-template Kappa of 0.19–0.63, confirming template selection impact on predictions. Comparative analysis reveals Copilot achieves the highest satisfaction (mean score 4.67), while Claude records the most complaints (36.9% negative). This study contributes a validated comparative ABSA framework for Indonesian-language LLM applications]]>
Copyrights © 2026
