Google Play Review Quality Scoring for Digital Engagement and App Development Using Transformer Models

High-quality Google Play reviews act as trust signals that shape digital engagement-driving discoverability, installs, and retention-yet many reviews are vague or noisy. We present a transformer-based framework that scores review quality on a 1−5 scale to surface informative feedback for engineering and engagement workflows. Our Thai-language corpus includes 27,768 reviews from 201 apps across eight categories. Ground truth is built via a hybrid pipeline: 5,968 human labels (inter-rater reliability Fleiss' κ=0.37), 20,000 multi-persona LLM labels (developer and user perspectives), and 1,800 pseudo-labels. We fine-tune PhayaThaiBERT for text regression (sequence length 256; L1 loss) and evaluate with Mean Absolute Error (MAE) across five variants. Balanced training distributions matter more than dataset size: the optimized model (v5), which applies a 30 % reduction to dominant labels (2 and 4), achieves a test MAE of 0.7734, outperforming a GPT-4o-mini baseline (1.2580). Confusion-matrix analysis of discretized predictions shows residual errors concentrated between adjacent quality levels (3↔4,4↔5), mirroring human disagreement in borderline cases. The system enables developers to prioritize actionable bug reports and feature requests, while marketers integrate quality-filtered signals into App Store Optimization, brand monitoring, and consumer-insight dashboards. Overall, domain-specific fine-tuning combined with principled class balancing offers a practical, scalable path to engagement-aware review analytics in mobile marketplaces.