KE-LPG: Toward Semantic Refinement for Lesson Plan Generation

Hanghui Guo, Jia Zhu, Changfan Pan, Qing Wang, Cong Zhou, Chaojun Meng

Abstract: Lesson plan design can optimize the teaching process, making teaching more targeted and scientific, thereby improving teaching quality and student learning outcomes. But teachers often struggle to complete high-quality lesson plan design due to limited teaching time and insufficient design skills. To solve this problem, the emergence of large language models (LLMs) provides teachers with a more efficient way to plan teaching content, design teaching activities, and optimize teaching strategies, thus enhancing the quality and effectiveness of lesson plan design. However, existing LLMs have limitations. Their static nature leads to a lack of professional education knowledge and an inability to continuously learn new knowledge. As a result, they perform poorly when dealing with unfamiliar content and struggle to generate detailed and high-quality lesson plans. To address these issues, we propose a Knowledge-Enhanced Lesson Plan Generation method (KE-LPG). Specifically, this method extracts subgraphs related to lesson plan content from a subject-specific knowledge graph. By combining techniques like graph Laplacian learning and semantic relevance calculation, it accurately constructs a keyword graph to enhance the LLM’s knowledge expression and understanding in the education field. Furthermore, during the pre-training stage, the constructed keyword graph is integrated with lesson plan content for fine-tuning, improving the LLM’s performance in lesson plan generation. To our knowledge, this is the first time that a semantic refinement approach has been used to generate lesson plans. Experimental results show that our method not only provides high-quality curriculum plans but also ensures the reliability of the generated knowledge points. The resource of the paper is available at https://github.com/ghh1125/data.

External IDs:doi:10.1007/s13369-025-10808-4