Construction of Educational Knowledge Graphs and Their Application in Question Answering
Kazakov, Roman
Kazakov, Roman
Author
Supervisor
Department
Natural Language Processing
Embargo End Date
30/05/2025
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Knowledge graphs (KGs) can be a valuable tool for learners and educational AI innovators. Our work discusses their potential in application to question answering. We propose a pipeline for KG construction, including concepts and relations extraction with an LLM, rulebased concept deduplication, and embedding-based relation labeling. We incorporate the KG into a RAG system, proposing two methods for retrieval: 1) Naive, which identifies concepts in the question and extrapolates the contextual subgraph; 2) Embedding, which utilises vectors to find relevant parts of the KG and performs radius expansion to extract broader context. Finally, we conduct our experiments on two datasets NLPBench and a new subset of NLP-related questions from various sources. The results suggest moderate yet statistically significant improvement compared to baselines. This work contributes to the research on LLM-driven KG construction, graph-based RAG, and applying NLP techniques to educational materials.
Citation
Roman Kazakov, “Construction of Educational Knowledge Graphs and Their Application in Question Answering,” Master of Science thesis, Natural Language Processing, MBZUAI, 2025.
Source
Conference
Keywords
Knowledge Graph, Retrieval-Augmented Generation (RAG), Question Answering, AI for Education, Large Language Model (LLM), Concept Extraction