.png){kind=icon}
.png){kind=icon}
Return
Knowledge Graph Augmented Generation for Examination Questions
.png){kind=icon}
PRE
AI
DOI：10.1109/tlt.2026.3682325.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Examination question generation (EQG) is a critical educational task that helps educators and learners acquire new knowledge. Traditional EQG approaches rely heavily on manually crafted rules and expert intervention, constraining their scalability and adaptability. Although large-language-model-based methods can harness extensive general knowledge to generate more varied questions, they remain prone to hallucinations stemming from inadequate domain-specific expertise and limited reasoning capabilities. In this article, we propose a unified framework for EQG tasks with methods, domain-specific datasets, evaluation metrics, and an agent system. Specifically, we first propose a keyword-guided method to construct a structured 2-D knowledge graph (KG), where keywords drive the extraction of meaningful triples and reinforce entity relations. Second, with the support of KG, we present a KG augmented generation (KAG) method, which generates examination questions and answers with external knowledge and graph reasoning. Third, we construct two new datasets and introduce novel metrics from both subjective and objective perspectives. Extensive experiments demonstrate that KAG significantly outperforms state-of-the-art baselines. Finally, we design an agent system that can automatically generate high-quality well-structured examination papers, easily adaptable to other subjects due to its training-free nature. We hope our proposed methods, datasets, metrics, and agent will facilitate and accelerate the adoption of EQG tasks.
Keywords:
Examination question generation (EQG)
GraphRAG
knowledge graph
large language model (LLM)
Journal
IF:
4.9
Papers:
123
Citations:
3.0K