What question did this study set out to answer?

This research focuses on improving educational question-answering systems by utilizing curriculum-aware hybrid retrieval and teacher-aligned prompting to enhance response quality.

June 10, 2026Open Access

EduChatQA: curriculum-aware hybrid retrieval and teacher-aligned prompting for textbook-grounded educational QA

Key Points

This research focuses on improving educational question-answering systems by utilizing curriculum-aware hybrid retrieval and teacher-aligned prompting to enhance response quality.
Developed EduChatQA framework integrating hybrid retrieval, knowledge-graph binding, and dual response heads.
Evaluated using classical QA metrics and RAGAS indicators on the QuAC-Edu benchmark across multiple subjects.
Conducted ablation studies to determine optimal knowledge graph performance and retrieval-augmented classification tasks.
EduChatQA achieved an average ROUGE-1 score of 0.69 and a Response Relevancy-5 score of 0.85, outperforming strong baselines by several points.
Human judgment resulted in Context Precision@5 of 0.52, Context Recall of 0.67, and a Faithfulness score of 0.71, indicating better grounding in textbook content.
The hybrid retriever had a precision of 0.75, recall of 0.76, and F1 score of 0.75, surpassing results from vector-only and graph-only strategies.

Abstract

The emergence of large language models (LLMs) has opened new possibilities for education-oriented question-answering (EQA) systems. Nonetheless, limitations in textbook-grounded retrieval, multi-turn discourse control, and response-quality assessment remain pressing challenges. This study presents EduChatQA, a fine-tuned LLM framework that integrates curriculum-aware hybrid retrieval, combining learnable fusion of dense and BM25 signals, with knowledge-graph binding, a role-aligned teacher-student prompt, a dual no-answer head, and a topic-shift detector. A comprehensive evaluation protocol combining classical QA metrics and RAGAS indicators is employed to assess answer quality. Experimental results on the QuAC-Edu benchmark, covering Physics, Chemistry, Biology, and World History, show that EduChatQA achieves an average ROUGE-1 score of 0.69 and a Response Relevancy-5 score of 0.85, outperforming strong baselines by 3–6 points in ROUGE-1 and 6–10 points in relevancy. Human judgment further reports Context Precision@5 of 0.52, Context Recall of 0.67, and Faithfulness score of 0.71, indicating enhanced grounding in textbook content. Ablation studies on the knowledge graph module show optimal performance with 75–100% triple retention, where Precision@5 reaches 51.6 at full retention. In retrieval-augmented classification tasks, the hybrid retriever achieves a precision of 0.75, recall of 0.76, and F1 score of 0.75, outperforming vector-only and graph-only retrieval strategies. This enhancement is attributable to the symmetric fusion of lexical (BM25) and semantic (dense) signals—an approach that effectively integrates the advantages of each retrieval mode while compensating for their individual limitations. These findings demonstrate that curriculum-aware hybrid retrieval combined with role-aligned prompting provides a reliable, textbook-traceable, and pedagogically coherent solution for classroom-ready EQA systems.

Perguntar à IA

Bookmark

View Full Paper