Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author(s),

I believe that your paper brings an important and current topic and although the study is primarily aimed at experts in the NLP and AI fields, I would like to suggest just a few changes so that it can be beneficial to those from the field of education as well. These remarks are not a condition for acceptance of the work, but represent suggestions that I believe would further improve the clarity and educational value of the work:

1. Consider including at least one specific example of questions and answers your system generates to understand the usability of the study in an educational context.

2. In the discussion, you could raise the question of how MDKAG could be integrated into existing educational platforms and how it can be interpreted in the teaching process.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Review Report

MDKAG: Retrieval-Augmented Educational QA Powered by a Multimodal Disciplinary Knowledge Graph

 

Summary:

 

The paper presents MDKAG, a novel framework for enhancing educational question-answering (QA) systems by integrating retrieval-augmented generation (RAG) with a multimodal disciplinary knowledge graph (MDKG). The approach leverages diverse educational resources—textbooks, slides, and classroom videos—and employs ERNIE 3.0 and other LLMs to extract high-precision entities, construct a structured knowledge graph, and enable accurate, context-aware response generation. A key innovation is the introduction of an answer-verification module that checks semantic overlap and entity coverage to reduce hallucinations, coupled with a dynamic graph update mechanism. The system is evaluated across three university courses and shows significant improvements in accuracy and reduced hallucination rates compared to baseline RAG and KAG models.

 

Strengths:

 

1. Innovative Integration of Multimodal Data and Knowledge Graphs: The framework’s ability to unify heterogeneous educational materials (text, images, audio/video) into a coherent, queryable knowledge graph is a significant advancement. The preprocessing pipeline for different modalities is well-designed and practical for real-world deployment.

 

2. Effective Answer Verification and Dynamic Update Mechanism: Unlike most RAG systems that stop at retrieval and generation, MDKAG introduces a dual-criteria verification system (semantic similarity and entity coverage) and triggers incremental knowledge graph updates when gaps are detected. This closed-loop design enhances reliability and supports long-term adaptability—critical for evolving educational content.

 

 

 

 

Questions:

1. Scalability and Computational Overhead:

The paper mentions the use of LLMs for ontology design and entity extraction, as well as graph traversal for retrieval. Could the authors provide more details on the computational cost and latency of the full pipeline—especially during graph construction and retrieval? How would MDKAG scale to hundreds of courses or large institutions, and what strategies are in place to manage indexing and update bottlenecks?

 

2. User Interaction and Feedback Integration:

While the dynamic update mechanism includes "user feedback integration" the specifics are limited. Can the authors elaborate on how user corrections (e.g., from students or instructors) are validated and incorporated into the KG? Is there a risk of introducing noise or bias through unverified feedback, and how is this mitigated?

 

3. Generalization Across Disciplines and Languages:

The evaluation is limited to three Chinese university courses. To what extent do the authors believe MDKAG can generalize to other domains (e.g., medicine, law) or languages? Were any domain-specific prompts or entity types manually tuned, and if so, how much effort would be required to adapt the system to a new discipline?

 

Overall Assessment:

This is a strong, well-structured, and impactful contribution to the field of AI in education. The methodology is sound, the experiments are relevant, and the results convincingly demonstrate improvements over existing baselines. With minor clarifications on scalability and generalization, this paper is suitable for publication after revision.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx