A multidimensional knowledge tracing hybrid model integrating semantics and dialogue behavior for cultural education

In intelligent dialogue systems for cultural teaching, traditional knowledge tracing models exhibit a limited capacity to interpret open-ended responses and struggle to characterize multidimensional cultural learning abilities or to effectively leverage the structural information embedded in instructional dialogues. To address these limitations, this paper proposes an improved SAINT+ (Separated Self-Attentive Neural Knowledge Tracing)–DKVMN (Dynamic Key-Value Memory Network) hybrid knowledge tracing model tailored for cultural teaching. The core idea is to introduce semantic–dialogue–ability structural constraints at both the sequence modeling and memory modeling levels simultaneously. Specifically, the model unifies the semantic information of students' open-ended responses—encoded via RoBERTa (Robustly Optimized BERT Pretraining Approach)—with dialogue-behavior features into structured learning events, thereby incorporating the dialogue process into sequence modeling. At the memory layer, the conventional single-dimensional mastery probability is extended into multidimensional concept mastery states, capturing the evolution of distinct ability dimensions in cultural learning, including factual understanding, explanatory ability, reflective depth, and confidence. The traceability of state transitions is further enhanced through an update mechanism constrained by behavioral conditions. Building on these components, the model achieves a balance between predictive performance and interpretability through joint multi-task training. Experimental results show that the proposed model attains an overall prediction accuracy of 0.862 (95% CI 0.856, 0.868) and an AUC (Area Under the Curve) of 0.905 (95% CI 0.899, 0.912). For the rubric classification of open-ended responses, the overall accuracy reaches approximately 0.877, with RMSE (Root Mean Square Error) values ranging from 0.09 to 0.11 across all ability dimensions. Moreover, the designed attention and memory-writing visualizations provide teachers with traceable concept-level evidence of student learning. Overall, the experiments demonstrate that the proposed model offers strong effectiveness and interpretability in intelligent dialogue systems for cultural teaching, providing reliable technical support for intelligent tutoring systems aimed at cultivating complex cultural competencies.