SKG-KT: A New Paradigm for Semantic Knowledge Graph Construction and Knowledge Tracing Driven by Large Language Models

The SKG-KT project combines the semantic understanding capabilities of large language models (LLMs) with the structured representation of knowledge graphs, pioneering a brand-new knowledge tracing method. By automatically constructing semantic knowledge graphs and performing reasoning, the system breaks through the limitations of traditional knowledge tracing methods in semantic understanding and associative reasoning, enabling more accurate modeling of learners' knowledge states.

Traditional knowledge tracing methods have obvious limitations:

• Bayesian Knowledge Tracing (BKT)：Assumes knowledge concepts are in either mastered or unmastered states, which is overly simplified and fails to capture complex relationships;
• Deep Knowledge Tracing (DKT)：A black-box model that ignores semantic associations between knowledge concepts;
• Graph Neural Network Knowledge Tracing (GKT)：High cost of manual graph construction, difficulty in scaling, and lack of personalization. SKG-KT proposes an automated, semantic-aware solution to address these pain points.

SKG-KT uses LLMs to automatically construct semantic knowledge graphs, with core steps including:

1. Concept Extraction and Standardization：Identify knowledge concepts in exercises and unify the same concept with different expressions (e.g., "solving quadratic equations" and "solution of ax²+bx+c=0" are mapped to the same node);
2. Relationship Reasoning and Graph Construction：Infer semantic relationships between concepts such as prerequisite, relevance, and extended application;
3. Dynamic Graph Update：Optimize the graph structure with new exercises and learning data to adapt to different disciplines and difficulty scenarios.

Based on the constructed semantic knowledge graph, SKG-KT implements reasoning-based knowledge tracing:

• Graph Neural Network Encoding：Fuse semantic embeddings of knowledge concepts (generated by LLMs) with statistical data of learners' historical performance;
• Message Passing and State Propagation：Propagate changes in concept states to related nodes according to the relationship weights inferred by LLMs;
• Interpretable Diagnosis：Trace the reasoning path to clearly identify learning difficulties caused by insufficient mastery of prerequisite concepts.

LLMs play four key roles in SKG-KT:

1. Semantic Parser：Extract structured knowledge information from unstructured exercise texts;
2. Reasoning Engine：Judge semantic relationships between knowledge concepts in line with educational rules;
3. Generation Enhancer：Generate synthetic practice samples or knowledge explanations to expand training data;
4. Explanation Generator：Produce natural language diagnostic reports, converting model reasoning results into understandable feedback.

SKG-KT adopts a modular architecture with core components including:

• Data Preprocessing Module：Convert raw exercise data into a format processable by LLMs;
• Knowledge Graph Construction Module：Implement concept extraction and relationship reasoning through prompt engineering and few-shot learning;
• Knowledge Tracing Model Module：Knowledge state tracing algorithm based on graph neural networks;
• Evaluation and Visualization Module：Model performance evaluation tools and knowledge graph visualization functions. The project code is open-sourced, lowering the threshold for cross-domain research in educational AI.

Application scenarios of SKG-KT include:

• Personalized Learning Path Planning：Recommend optimal learning sequences based on graph dependency relationships;
• Precise Weak Point Diagnosis：Locate the root cause of knowledge gaps;
• Intelligent Content Generation：Automatically generate targeted exercises and explanatory materials;
• Cross-Course Knowledge Transfer：Identify knowledge overlaps between different courses to support interdisciplinary learning. Its educational significance lies in building an intelligent system that understands the essence of knowledge to assist human educators.

SKG-KT breaks through the limitation of traditional knowledge tracing relying on behavioral data, introducing new dimensions of semantic understanding and reasoning. This "LLM + Knowledge Graph" paradigm is not only applicable to the education field but also provides references for other scenarios requiring deep semantic understanding and structured reasoning. With the improvement of LLM capabilities and the maturity of knowledge graph technology, more cross-modal fusion innovations are expected to emerge.