A Sentiment-Based Evaluation of Museum Display Design through Hybrid IndoBERT and Rule-Based Lexicon

Abstract

Traditional museums remain vital cultural institutions but face persistent challenges in engaging contemporary audiences. The Wayang Museum in Indonesia, despite partial digital renovation, continues to rely heavily on static, text-based displays that often hinder visual comfort and informational comprehension. Conventional sentiment analysis techniques have demonstrated limitations in accurately capturing the subtle feedback from visitors, particularly within the culturally nuanced contexts of Indonesia. To address this problem, this study proposes a hybrid Aspect-Based Sentiment Analysis (ABSA) model that integrates a fine-tuned IndoBERT transformer with a rule-based sentiment lexicon. The hybrid architecture combines probabilistic embeddings with domain-specific lexicon rules to enhance classification accuracy, calibration, and interpretability. Experimental results demonstrate that while a baseline IndoBERT model achieved high accuracy (98.3%) and macro F1-score (0.975), the proposed IndoBERT-Lexicon model achieved perfect classification accuracy and near-perfect calibration (Negative Log-Likelihood = 0.002, Expected Calibration Error = 0.002). In comparison, a classical SVM with TF-IDF achieved similar accuracy but exhibited significantly weaker calibration despite its superior computational efficiency. The originality of this work lies in demonstrating that integrating a culturally grounded rule-based lexicon with transformer embeddings can simultaneously improve accuracy, calibration reliability, and interpretability in aspect-based sentiment analysis for cultural heritage applications. The key contribution of this study lies in demonstrating the methodological effectiveness of combining transformer embeddings with a culturally grounded lexicon to enhance both accuracy and calibration.