A hybrid approach to service performance measurement in the era of artificial intelligence: Integrating BERT-based aspect-based sentiment analysis and synthetic data

This study advances aspect-based sentiment analysis (ABSA) by examining how data quality and human–AI collaboration enhance the reliability of service performance measurement from online reviews. A hybrid-validated dataset and transformer-based models are developed to assess whether ABSA can generate attribute-level performance signals from restaurant reviews. A four-study framework is employed: Study 1 constructs a hybrid-validated dataset by integrating rule-based, large language model-assisted, and human-coded labels; Study 2 benchmarks model variants; Study 3 applies the optimal model to TripAdvisor reviews to derive aspect-level sentiment scores; and Study 4 evaluates their complementarity with structured ratings. Results show that hybrid validation improves model performance. ABSA-derived features provide moderate explanatory power and, when combined with structured ratings, achieve comparable performance to ratings alone. These findings suggest that ABSA complements rather than replaces structured evaluations, offering scalable text-based signals for service performance assessment.