AI-Powered Adaptive Assessment Using NLP-Based Scoring and Sentiment-Driven Item Selection for Personalized Learning

This research proposes and evaluates an AI-powered adaptive assessment framework that integrates automatic scoring and sentiment analysis to personalize item sequencing in digital learning environments. The study utilizes a dataset of 14,960 open-ended student responses sourced from 420 learners across 48 assessment items, each labeled with a six-level cognitive score and three-level sentiment polarity. A transformer-based scoring model achieved an exact-match accuracy of 0.64, a macro F1-score of 0.71, and a quadratic weighted kappa of 0.82, while correlating strongly with expert ratings (Spearman = 0.87). The sentiment classifier reached an overall accuracy of 0.79, with a macro F1-score of 0.76, and demonstrated recall of 0.83 for neutral responses, 0.72 for negative responses, and 0.69 for positive responses. Simulation of 50 adaptive learners over 12 steps showed an average difficulty progression of +1.8 bands, a 96 percent overload-prevention rate, and sentiment-triggered difficulty downgrades in 14 percent of selections, stabilizing difficulty variation in 72 percent of learners by step eight. These results confirm that NLP-based scoring reliably models expert judgment, sentiment inference identifies affective modulation, and the joint use of both signals enables real-time, emotion-aware adaptation. The findings demonstrate that open-ended assessment can be automated with cognitive-affective sensitivity, providing scalable personalization for online learning ecosystems.