This study introduces a novel approach for predicting student engagement levels in a language-based AI curriculum. The curriculum was integrated into English Language Arts classrooms, in which 106 students from five classes participated five web-based machine learning and text mining modules for 2 weeks. Sentiment and categorical analyses, performed by a hybrid model of Bidirectional Encoder Representations from Transformers (BERT) and Multilayer Perceptron (MLP), were employed to predict students' engagement levels. The input textual data and categorical data were extracted from the learning modules, resulting in a testing accuracy of 78.5%. This innovative engagement level identification approach provides an objective method for student engagement auto-prediction and paves the way for targeted interventions to optimize AI learning experiences.