Knowledge Graphs for Education

This workshop is part of an effort to launch a new Translational Institute on Knowledge Axiomatization. The intended outcome of the workshop is to develop a concrete conception of a knowledge resource to support education in the context of modern AI.

·         Why is a knowledge resource needed for education?

·         What information should it contain?

·         How should it be created?

We will answer the above questions by exploring a dual interplay between AI and Education: How can AI help improve education? How might education help improve AI?

How can AI help Education?

As pointed out at a recent AI/ED keynote, in some ways generative AI has changed everything in education, and in other ways it has changed nothing. It has changed what we prioritize in student learning and the tools available to support it. It, however, has no impact on the inherent quality of the teaching materials, complexity of the subject to be learned, number of concepts to be mastered, etc.

We will begin the workshop with a discussion on concrete experiences in the use of generative AI to improve student learning. We will do this through the lessons learned in the development and evaluation of Khanmigo, use of knowledge graphs for math education, and through an expert assessment of NotebookLM by a biology teacher. As innovators are exploring a wide range of techniques to supplement foundation LLMs in order to better serve education needs. Where might a knowledge resource have the most promise, and what would be a compelling step to demonstrate this promise?

The true value of knowledge resources in education extends far beyond intelligent content delivery. Consider these crucial applications:

Precision Knowledge Tracing. Traditional methods often struggle to pinpoint specific conceptual gaps. KGs, by modeling prerequisite relationships and fine-grained knowledge components, enable highly granular knowledge tracing. We can more accurately track a student's learning trajectory, understand precisely which foundational concepts are mastered or lacking, and provide truly adaptive support.

Grounding AI in Verifiable Knowledge. As AI, particularly LLMs, becomes more integrated into education (e.g., tutors, assistants, content generators), ensuring factual accuracy, pedagogical soundness, and explainability is paramount. KGs serve as the essential grounding layer, anchoring AI outputs to curated, verifiable knowledge. This combats hallucination, ensures alignment with curricula, and allows AI interactions to be traced back to reliable sources, building trust and efficacy.

The unique demands of the educational domain necessitate pushing the boundaries of traditional knowledge graph (KG) architectures that are used for creating a knowledge resource. Simple (entity, relationship, entity) triples are insufficient to capture the richness of pedagogical strategies, common misconceptions, learning pathways, or causal links between instructional activities and learning outcomes. This challenge opens exciting new research areas in developing more expressive KG frameworks tailored for education.

How can Education help AI?

Current approach to AI relies on training on large amounts of data. Human learning, as we have known for thousands of years, has relied on social education. Children start learning by looking at examples, but their learning takes off when they enter the formal education system where they are taught not just by examples, but by teaching them concepts across a range of topics. Much work exists in learning sciences that can inform us on concepts learned by children at each stage.

Could a new method for training AIs be devised that relies on our knowledge of the concepts that are to be learned at each grade level? Could we test the success of that learning not through imperfect exams, but through cognitive interviews? What would need to be done to operationalize such learning for Ais?

Present Workshop

Motivated by these opportunities, and as part of our continuing NSF OKN project, we are convening key stakeholders: researchers, educators, developers, policymakers, and industry partners, who can both benefit from and contribute to this vision. This workshop aims to:

●     Amplify the Impact. Deepen understanding and effectively communicate the significance of prior classroom efficacy trials using KG-based educational tools.

●     Harness Modern AI. Strategize the construction of large-scale educational KGs (like TOKN) leveraging cutting-edge LLMs and hybrid AI approaches.

●     Address Real-World Needs. Identify current requirements, bottlenecks, and challenges faced by educators and edtech developers where KGs can offer solutions.

●     Enable Knowledge Tracing & Grounding. Explore best practices and specific applications for using KGs to enhance knowledge tracing models and ground AI educational tools.

●     Chart a Scalable Path. Define practical, economically viable pathways to develop, deploy, and sustain KG-based educational infrastructure and tools.

●     Map the Research Frontier. Identify fundamental research questions and opportunities related to expressive KGs for education, potentially informing future large-scale initiatives like a new NSF AI Institute focused on this domain.

We believe that collaborative development and research centered on educational knowledge graphs will be foundational for the next wave of impactful AI in education. We invite you to join this critical conversation and help shape this exciting future.