Textbook2KG: Decoding Knowledge Graph from Very Long Textual Content via Prompt Engineering
Hengzhi Wu, Xiaoting Zhong, Junying Yuan · 2025
We introduce Textbook2KG for building knowledge graphs from lengthy textbooks using large language models (LLMs) and smart prompts. Our framework uses a step-by-step approach: splitting texts, extracting key facts, and connecting concepts through reasoning. Tested on three real textbooks, it achieved $85.5 \%$ accuracy in direct fact extraction but $65 \%$ for inferred links, showing where LLMs shine and where they need help. The textbook datasets we created with $\mathbf{5 2 8 k} \boldsymbol{+}$ words analyzed give researchers a solid base to improve educational KG tools. While current models make KG creation easier for teachers, smarter reasoning remains a challenge. This work shows how simple prompt tweaks can unlock LLMs’ hidden skills for organizing academic knowledge, blending AI power with classroom needs.