This study introduces a novel dual-branch Deep Matrix Factorization (DeepMF) framework enhanced by NLP techniques for predicting student performance in the Intelligent Tutoring Systems. Building on previous research, the proposed approach adopts a fundamentally different modeling strategy by transforming discrete educational features-such as learner ID, exercise, question, skill group, session start time, and repeated attempts-into structured sentences that capture both temporal and sequential information. These inputs are processed through two complementary branches. The first branch employs pre-trained GloVe embeddings, followed by a self-attention layer that captures intra-sequence dependencies before passing the representations into a DeepMF module. The second branch leverages a BERT-based model to extract contextualized language features. To address the issue of class imbalance, Focal Loss is applied during training on both the KDDCup 2010 and Assistment 2017 datasets. Experimental results demonstrate substantial improvements in prediction accuracy: RMSE is reduced from 0.418 to 0.167 (a 60.1% reduction) on KDDCup 2010, and from 0.472 to 0.186 on Assistment 2017 (representing a 60.6% relative improvement). These findings confirm the effectiveness of integrating contextual, sequential, and temporal modeling with DeepMF in educational data mining.