AI-Driven Textual Feedback Analysis in E-Training Using Enhanced RoBERTa

Abstract

In corporate e-training environments, traditional metrics like course completion and quiz scores often fail to reflect actual job performance. Rich insights are embedded in unstructured textual feedback, yet they remain underutilized due to limitations in existing analytical models. This study proposes E-RoBERTa, an enhanced transformer-based model designed to predict employee job performance by analyzing open-ended feedback from digital training platforms. The model aims to improve accuracy, domain adaptability, and interpretability. E-RoBERTa integrates Domain-Adaptive Pretraining (DAPT) to fine-tune RoBERTa on corporate-specific language and introduces Dynamic Attention Scaling (DAS) to highlight semantically critical tokens. A real-world, GDPR-compliant dataset containing 16,000 feedback entries from 3,500 employees across multiple departments was used. Preprocessing included tokenization, sentiment tagging, and feature extraction. The model achieved superior performance with a macro F1-score of 0.875, outperforming standard RoBERTa, LSTM, and SVM baselines. Attention visualizations revealed alignment between influential tokens and human-interpretable performance indicators. E-RoBERTa provides a transparent and accurate framework for evaluating job performance through textual feedback. Its use of domain adaptation and dynamic attention mechanisms supports scalable, ethical, and explainable AI in corporate learning analytics, offering actionable insights for personalized interventions and strategic HR decision-making.