Innovation and practice of "three-step" training mode of normal students' educational practice under the empowerment of artificial intelligence

In the context of the flourishing development of artificial intelligence in the education field, the improvement of normal students' educational practice ability through AI - based means has become an urgent issue. This study developed an AI - supported "three-step" training model for normal students. It uses the K - means clustering algorithm (k = 5) for learning style classification and personalized learning path creation. BiLSTM - attention networks (3 - layer, dropout = 0.3) predict teaching behavior (F1 - score = 0.87), and a BERT - based NLP model (fine - tuned with 12,000 examples) provides 91.3% accurate real - time feedback. A virtual lab (89.2% task completion rate) is integrated into the training. In the six-month controlled experiment (N=350), compared with the traditional method, the "three-step" model achieved a significant improvement in the teaching ability score, with an overall improvement of 28.4% (P < 0.001, η = 0.42), especially in the aspects of classroom management and teaching design, which increased by 37.1% and 29.6% respectively. In addition, in another group of experiments (n=200), the evaluation score of TPACK is verified by A/B test, which shows that the model can improve the teaching efficiency by 23.7%(p<0.05, Cohen's d=1.2). These results show that the AI-driven training model, which combines machine learning and natural language processing technology, has a significant effect on improving the comprehensive ability of teacher candidates.