This paper integrates LSTM networks and attention mechanisms to construct a deep knowledge tracking model based on feature embedding and attention mechanisms, and evaluates the predictive performance of this model. A university English intelligent adaptive learning system is designed, and the characteristics of teacherstudent speech behavior and teacher-student interaction in English classrooms are analyzed, with corresponding optimization strategies proposed. DKT-FA achieved prediction accuracies of 0.8402, 0.8821, 0.7506, and 0.7976 on the ASSISTment2009, ASSISTment2017, EdNet, and English datasets, respectively, achieving the best performance among all tested models. In Lesson Example 1 (an English teaching classroom based on an adaptive learning system), the teacher speech ratio, student speech ratio, teacher direct influence ratio, teacher indirect influence ratio, student active response rate, and student passive response ratio were 49.5%, 34.6%, 38.3%, 11.2%, 4.4%, and 24.2%, respectively. Case 2 (traditional teaching method) had the following ratios: 58.4%, 17.3%, 45.5%, 12.9%, 3.5%, and 7.6%. In Case 2, teacher speech behavior was concentrated in the first half and exceeded student speech behavior. In Case 1, teacher speech behavior was more balanced, and teacher-student interaction frequency was more stable.
