Nolan M. Yumen, Angie C. Canillo
Teaching evaluations generate valuable unstructured feedback that institutions struggle to analyze systematically. Traditional manual analysis examines only 31 percent of available comments, leading to substantial information loss. This study develops a multi-modal deep learning framework integrating Long Short-Term Memory networks, Convolutional Neural Networks, and transformer architectures to analyze 4,410 student comments from University of Antique across four colleges and four teaching dimensions. The framework addresses dataset imbalance through stratified sampling, text augmentation, class-weighted loss functions, and ensemble prediction. Results demonstrate 92.7 percent accuracy with 0.94 F1-score, reducing manual analysis time by 83.8 percent. Ablation studies confirm multi-modal integration outperforms single-architecture approaches by 8.3 to 12.4 percentage points. Optimization through mixed-precision training and quantization enables deployment on consumer-grade hardware. User validation with 42 faculty and 18 administrators achieved 89.2 System Usability Scale score. The framework provides interpretable insights through attention visualization, enabling evidence-based teaching improvement while maintaining transparency essential for educational contexts. © 2025 IEEE.
College of Computer Studies, University of Antique, Tario-Lim Memorial Campus, Antique, Philippines; University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
