Integrating Natural Language Processing with Deep Knowledge Tracing in Intelligent Tutoring Systems for Low-Resource Languages:A Case Study on Burmese STEM Curricula

Intelligent Tutoring Systems (ITS) offer powerful mechanisms for personalized instruction. However, their pedagogical efficacy remains fundamentally bottlenecked in low-resource language environments due to data isolation, severe digital asset scarcity, and a lack of mature processing baselines. Traditional sequence-based Deep Knowledge Trac ing (DKT) architectures infer student cognitive profiles over time by processing primitive binary success states, completely discarding the rich diagnostic indicators embedded within open-ended natural language feedback. This paperpresents anovel hybridarchitecture that natively fuses tokenization-independent Natural Language Processing (NLP) text features with Recurrent Neural Network (RNN) based DKT pipelines. We engineer a sub-word character-level semantic parser tailored for the non-segmented Burmese script, mapping free-form textual explanations directly into continuous concept-correctness vectors. These vectors dynamically update a Long Short-Term Memory (LSTM) network core to construct latent multi-skill mastery profiles and predict subsequent learning performance. Empirical evaluations using an engineered interaction dataset demonstrate that incorporating seman tic text embeddings yields a substantial predictive lift, achieving an AUC of 0.824 and an absolute performance gain of +0.062 over text-blind baselines. This architecture establishes an access-optimized deployment pathway for sophisticated educational frameworks across low-resource infrastructural constraints.