EduKGs: Educational Knowledge Graphs

• Educational Knowledge Graphs (EduKGs) are structured representations linking subjects, concepts, resources, and learner interactions via semantically typed relations.
• They are constructed using methods like NER, relation extraction, and human-in-the-loop validation to transform heterogeneous educational materials into coherent graphs.
• EduKGs enable applications such as question answering, personalized learning recommendations, and curriculum sequencing while enhancing explainability and retrieval efficiency.

Educational Knowledge Graphs (EduKGs) are structured representations of educational knowledge in which entities such as subjects, concepts, facts, resources, courses, slides, assessments, and learners are connected by semantically typed relations and, in several systems, by machine-readable triples or typed attributed graphs (Zhao et al., 2022, Abdelmagied et al., 15 May 2025). Unlike general-purpose KGs, EduKGs are explicitly organized around instructional structure, curricular context, learning resources, and, in some systems, learner state and recommendation context; published instances range from heterogeneous K–12 graphs to MOOC graphs, curriculum-scale higher-education graphs, contextual learning-object graphs, and resource-focused teaching graphs (Zhao et al., 2022, Abu-Rasheed et al., 2024, Abu-Rasheed et al., 21 Jan 2025, Ilkou et al., 2024).

1. Conceptual scope and formal models

EduKGs differ substantially in scope, but the literature converges on a common design premise: educational content is modeled as a graph whose nodes denote pedagogically relevant entities and whose edges denote relations useful for teaching, retrieval, sequencing, or explanation. In CourseMapper’s MOOC setting, the graph is formalized as $G = (V, E, R, A)$, where $V$ are nodes, $E$ are edges with types $R$, and $A$ are node and edge attributes; triples $(h, r, t) \in V \times R \times V$ encode directed relations (Abdelmagied et al., 15 May 2025). In the HCMUT case study, the KG is formalized as a labeled directed multigraph $G = (V, E, \phi)$, with an RDF-style triple set $T = \{(s,p,o)\}$ and an ontology tuple $O = (C, R, A)$ (Bui et al., 2024). In the contextual learning-object work, the starting point is an expert-curated five-level hierarchy, Journey $\rightarrow$ Course $V$0 Topic $V$1 Educational Package $V$2 Educational Content, which is then converted into a graph by retaining hierarchical links and adding semantic relations (Abu-Rasheed et al., 2024).

The schema of an EduKG depends on educational level and use case. EDUKG, the heterogeneous K–12 graph, defines an interdisciplinary and fine-grained ontology with 635 classes, 445 object properties, and 1314 datatype properties, and it explicitly models rhetorical roles as entities, including roles such as “content” and “definition” (Zhao et al., 2022). CourseMapper’s production EduKG instead centers on instructional delivery structure and retrieval nodes: Course, Learning Material, Slide, Main Concept, Related Concept, Wikipedia Article, Wikipedia Paragraph, and Learner, with relations such as contains, has_main_concept, linked_to_article, related_concept, has_paragraph, and did_not_understand (Abdelmagied et al., 15 May 2025). In higher-education curriculum modeling, the ontology can be even more explicitly split into curriculum, domain, and user models, with entities such as Module, Lecture, Session, Topic, Sub-Topic, Domain, Sub-Domain, and Student/Learner, linked by relations such as hasTopic, hasSubTopic, alignsWithDomainConcept, hasPrerequisite, and semanticallySimilarTo (Abu-Rasheed et al., 21 Jan 2025).

A recurrent distinction in the literature is between encyclopedic relations and pedagogically meaningful relations. EDUKG argues that textbook-style questions often require relations that conventional KGs do not encode, which is why it treats rhetorical roles as first-class entities and supports questions such as “What is the content of Newton’s first law of motion?” through KG queries rather than full-text search (Zhao et al., 2022). Graphusion, by contrast, fixes a compact pedagogical relation inventory for NLP education—Is-a-Prerequisite-of, Used-for, Part-of, Hyponym-of, Evaluate-for, Compare, and Conjunction—to support curriculum reasoning and tutoring tasks (Yang et al., 2024). The contextual LO graph emphasizes has_semantic_relation_to as a cross-hierarchy contextual edge type, while the HCMUT graph highlights institutionally grounded entities such as Service, Policy, Intent, Program, Course, and Class (Abu-Rasheed et al., 2024, Bui et al., 2024). This suggests that EduKGs are best understood not as a single ontology family but as task-specific educational graph models that share a common goal: making pedagogically relevant structure explicit.

2. Construction pipelines and curation strategies

A major line of work treats EduKG construction as an information extraction and curation problem over heterogeneous educational resources. EDUKG proposes a flexible and interactive pipeline for K–12 textbooks that separates entity labeling from relation and property labeling in order to reduce error propagation. Its preprocessing combines Named Entity Recognition for candidate entities, entity linking for disambiguation against EDUKG and Xlore, and Open Information Extraction for candidate triplets; annotators then assign ontology classes and Xlore equivalences in an entity-recognition interface and verify candidate triplets in a triplet-labeling interface, while OpenIE predicates are canonicalized to EDUKG properties by Sentence-BERT embeddings and cosine similarity (Zhao et al., 2022). In a user study, traditional relation extraction without candidate generation yielded an average of 4.98 sections/hour per annotator, whereas the interactive tool yielded 12–15 sections/hour (Zhao et al., 2022).

In MOOC environments, PDF- and slide-centric pipelines are prevalent. CourseMapper compares a Top-down and a Bottom-up automatic construction strategy over PDF learning materials. Both begin with layout-aware text extraction using PDFMiner, apply SIFRank_SqueezeBERT for keyphrase extraction, use DBpedia Spotlight with support=5 and confidence=0.35 for concept identification, and weight concept salience via SBERT cosine similarity between learning-material or slide text and Wikipedia article text (Ain et al., 15 May 2025). The Top-down approach extracts $V$3 the number of slides keyphrases at learning-material level and discards slide-level concepts not already present in the learning-material graph; the Bottom-up approach extracts 15 keyphrases per slide, computes a dual weight $V$4, and aggregates slide-level graphs into a learning-material graph without discarding slide-specific concepts (Ain et al., 15 May 2025). The paper adds a Human-in-the-Loop moderation workflow in which course moderators preview extracted main concepts, edit or remove irrelevant ones, add missing concepts, link them to slides, and then publish the verified EduKG (Ain et al., 15 May 2025).

A later CourseMapper pipeline refines this slide-first strategy into an end-to-end worker-based system. It introduces font-size analysis with a separation threshold of $V$5 pt, text distance analysis with segmentation at $V$6 line height or width, location-aware noise filtering for repeated headers and footers, bullet-point analysis, DBpedia Spotlight linking, SBERT all-mpnet-base-v2 weighting, disambiguation-page replacement, top-20 related-concept expansion, top-5 category expansion, and pruning with $V$7 (Ain et al., 5 Sep 2025). The architecture uses Redis as a message broker, Neo4j for graph persistence, and PostgreSQL for precomputed Wikipedia-derived embeddings and metadata (Ain et al., 5 Sep 2025).

Higher-education curricula introduce additional heterogeneity. The multi-source college-curricula framework constructs separate graphs from textbooks, slides, and syllabi, using rule-based regular expressions and CRF-based Chinese NER to produce hierarchical structures such as Course $V$8 KnowledgeUnit $V$9 KnowledgeChapter $E$0 KnowledgeBlock $E$1 KnowledgePoint, then cleans extracted text with a deep learning-based language error correction model before importing triples into Neo4j (Li et al., 2023). The LLM-assisted higher-education recommendation work instead uses Whisper to transcribe lecture videos and GPT‑4o to extract Topics and Sub-Topics from syllabi, slides, manuscripts, and transcripts, with instructor validation after extraction and semantic linking across modules via embedding similarity and LLM analysis (Abu-Rasheed et al., 21 Jan 2025). Resource-focused teaching KGs extend the construction problem further by coupling a domain teaching graph to lecturer Personal KGs and by adding topic modeling over lecture notes as an enrichment step (Ilkou et al., 2024).

Across these systems, human oversight is not treated as incidental. EDUKG uses interactive annotation and backend weight adjustment; CourseMapper adds moderator review before publication; higher-education LLM pipelines use expert validation of extracted topics and descriptions (Zhao et al., 2022, Ain et al., 15 May 2025, Abu-Rasheed et al., 21 Jan 2025). A plausible implication is that sustainability in EduKG construction is not merely continuous ingestion, but continuous alignment between automated extraction, ontology control, and domain-expert correction.

3. Fusion, completion, optimization, and representation learning

Once graph fragments are constructed, a second technical problem arises: how to merge, complete, or optimize them. The college-curricula fusion framework performs entity name matching by string, literal, and semantic similarity; among equivalent entities $E$2 with levels $E$3, the fused node is placed at the highest, coarsest level $E$4, while attributes are combined by union,

$E$5

and original hierarchical and contextual edges are retained and reconnected to the fused node (Li et al., 2023). This fused graph then supports cross-course correlation analysis through

$E$6

with representative values such as $E$7 between Communication Principles and Signals and Systems (Li et al., 2023).

Graphusion frames fusion differently: as a global consistency problem over LLM-extracted triplets. Its pipeline first generates 688 seed concepts by BERTopic over 4,605 ACL abstracts, then performs zero-shot triplet extraction, and finally applies an LLM fusion prompt to merge aliases, enforce a single relation per concept pair, resolve relation conflicts with background text, and infer novel cross-graph triplets (Yang et al., 2024). The resulting fused EduKG is evaluated both intrinsically and downstream through TutorQA, an expert-verified benchmark of six tasks and 1,200 QA pairs (Yang et al., 2024). The emphasis here is not ontology-guided schema control but global adjudication over locally extracted graph fragments.

Completion and refinement have also been formalized as graph inference or optimization. The prerequisite-inference work for CourseMapper defines ten unsupervised criteria—document-based, Wikipedia hyperlink-based, graph-based, hierarchy-based, and text-based—and aggregates directional evidence through a voting scheme. For an ordered concept pair, two one-hot arrays capture evidence for each direction, yielding a raw difference $E$8 and a normalized score

$E$9

with $R$0 criteria and a decision threshold $R$1 (Alatrash et al., 5 Sep 2025). The method explicitly enforces asymmetry, removes self-loops, and performs cycle checks to preserve a DAG structure for prerequisite edges (Alatrash et al., 5 Sep 2025).

A more explicitly information-theoretic approach appears in rate-distortion-guided graph construction from lecture notes. There, lecture content is modeled as a metric-measure space $R$2 and a candidate EduKG as $R$3; Fused Gromov–Wasserstein alignment measures distortion, while the graph size contributes a rate term

$R$4

and the overall objective is

$R$5

Refinement operators—add, split, merge, remove, and rewire—are greedily accepted when they decrease the objective (An et al., 18 Nov 2025). The same paper reports that refined KGs produce interpretable rate-distortion curves and improved MCQ generation quality over raw notes (An et al., 18 Nov 2025).

Representation learning adds yet another layer. In “Joint Embedding Learning of Educational Knowledge Graphs,” an educational KG is formalized as $R$6, where literals are central rather than peripheral. Structural embeddings are learned with a TransE score,

$R$7

literal embeddings are computed by BERT over textual literals, and three GRUs fuse structural and literal vectors into joint embeddings scored again by a TransE-style objective (Yao et al., 2019). The paper argues that in EduKGs, rich literals such as definitions and properties carry more instructional value than topology alone (Yao et al., 2019). This claim aligns with later systems that rely heavily on descriptions, rhetorical roles, and linked textual contexts rather than on sparse prerequisite edges alone.

4. Educational applications

Question answering is one of the most developed EduKG applications. EDUKG supports both knowledge-based QA and text-based QA by mapping natural-language questions to SPARQL templates. A datatype property such as “starting time” can answer “What is the starting time of the French Revolution?”, while rhetorical roles modeled as entities let the graph answer “What is the content of Newton’s first law of motion?” without falling back to full-text search (Zhao et al., 2022). The HCMUT system applies a different QA architecture: an institutional LLM translates Vietnamese questions into Cypher over a Neo4j-backed cross-data KG, retrieves matching subgraphs about intents, policies, and courses, linearizes retrieved triples, and uses them as factual context for answer generation (Bui et al., 2024).

Graph retrieval-augmented generation extends QA into learner support. CourseMapper’s Graph RAG pipeline combines an EduKG with a Personal Knowledge Graph. When a learner marks a Main Concept as did_not_understand, the PKG-based Question Generation module retrieves the current slide text and slide concepts, prompts GPT-3.5-turbo to generate diverse questions, and re-ranks them by cosine similarity to slide text (Abdelmagied et al., 15 May 2025). The EduKG-based Question Answering module retrieves paragraph nodes from Wikipedia articles linked to main concepts, uses extractive prompting to answer from retrieved text, and falls back to graph-guided traversal of related concepts when primary retrieval fails (Abdelmagied et al., 15 May 2025). Evaluated by three expert instructors on three MOOCs, question generation achieved weighted averages of Fluency 2.981, Clarity 2.835, Conciseness 2.973, Rel_slide 2.853, Rel_dnuconcept 2.668, and overall Avg $R$8, while QA accuracy was 56.67% on Learning Analytics, 45.00% on HCI, and 33.33% on Web Technologies, with a weighted average of 45.00% (Abdelmagied et al., 15 May 2025).

Explainability is another major application. In the learning-recommendation setting, a KG can be used as a bounded factual context for LLM prompts. The recommendation-explanation work retrieves for each recommended learning object its hierarchical placement, semantic neighbors, community subgraph, and metadata, inserts these facts into an expert-designed prompt template, and asks GPT4-1106-preview to fill explanation slots such as “Why this LO,” “How it supports the goal,” and “Connections to related LOs” (Abu-Rasheed et al., 2024). On 52 reference explanations drawn from 10 recommended learning paths, KG-contextualized explanations consistently improved recall, precision, and F1 across ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-Lsum, and learners and experts rated explanation quality highly, with an acceptance score of 4.7/5 (Abu-Rasheed et al., 2024).

Personalization and adaptive sequencing often rely on explicit learner state. FOKE models domain knowledge as a hierarchical knowledge forest

$R$9

with tree retrieval by

$A$0

and personalized next-tree recommendation by

$A$1

where $A$2 is a forest-level mastery vector (Hu et al., 2024). In a different tutoring setting, KG-EIR combines a multi-dimensional exercise graph with skill-importance weights and a Neural Attentive cognitive diagnosis model to recommend exercises; on ASSISTments 2009–2010, NACD achieved AUC/ACC of 0.772/0.735, and on Eedi 2020 it achieved 0.751/0.687 (Li et al., 2023). The KGs-as-subquestion-generators line applies graphs at question level: for approximately 8,000 NCERT-style physics items, an LLM first constructs a knowledge graph of the question’s internal logic and then uses it to generate sub-questions. On a human evaluation of 100 questions, decomposition with KG achieved success rates of 84.45 on numerical questions and 74.58 on conceptual questions, compared with 81.33 and 71.30 for decomposition without KG, and 77.62 and 68.53 for standard prompting (Addala et al., 2024).

These application profiles show that EduKGs are used not only as static repositories but also as operational control structures for retrieval, explanation, decomposition, diagnosis, and sequence-aware recommendation.

5. Evaluation regimes, empirical profiles, and scale

Empirical evaluation of EduKGs spans graph construction, linking, structure, downstream reasoning, and educational utility. EDUKG evaluates entity linking against human-labeled ground truth on physics, biology, and geography textbooks, reporting Recall/Precision/F1 of 81.01/87.88/85.14 for physics, 86.48/77.01/81.47 for biology, and 66.38/90.8/76.79 for geography (Zhao et al., 2022). The same project publishes a graph with more than 252 million entities and 3.86 billion triplets (Zhao et al., 2022). CourseMapper’s comparison of Top-down and Bottom-up construction reports P@15/MRR/MAP of 0.807/0.941/0.807 for Top-down and 0.812/0.941/0.812 for Bottom-up, while Simple Random Sampling accuracy estimates are $A$3 with $A$4 for Top-down and $A$5 with $A$6 for Bottom-up (Ain et al., 15 May 2025).

The optimized CourseMapper pipeline quantifies engineering gains more explicitly. Its SRS-based EduKG accuracy increases from $A$7 to $A$8, which the paper states as a 17.5% relative improvement, while generation time decreases from 24.35 s to 2.3 s per slide and expansion time from 222 s to 1.89 s per concept, corresponding to speedups of approximately 10.59 and 117.46 (Ain et al., 5 Sep 2025). The same study reports Precision/Recall/F1 improvements in concept annotation from 0.13/0.17/0.14 to 0.18/0.29/0.23 after introducing weighting, disambiguation, and pruning (Ain et al., 5 Sep 2025).

Structural graph metrics are used when recommendation or contextualization is the target. The contextual LO graph reports Average Degree Centrality increasing from 1.079 in the hierarchy baseline to 2.262 in the completed KG; the number of communities rises from 253 to 541; average modularity falls from 0.779 to 0.636; Weakly Connected Components decrease from 63 to 35; and average Betweenness Centrality rises from 1.57 to 15.1 (Abu-Rasheed et al., 2024). The multi-source Electronic Information EduKG spans over 60,000 entities and 80,000 triples across hundreds of courses and uses equivalence-driven correlation analysis plus frequency-based ranking to identify “hot” concepts and 16 cross-source intersection concepts in Communication Principles (Li et al., 2023). The LLM-assisted higher-education recommendation graph evaluates 1,197 extraction samples and reports, for Embedded Systems, Topic Precision/Recall/F1 of 0.99/0.94/0.96 and Sub-Topic Precision/Recall/F1 of 1.00/0.97/0.98; for FPGA, Topic Precision/Recall/F1 of 0.97/1.00/0.98 and Sub-Topic Precision/Recall/F1 of 0.89/0.99/0.94, while Average Degree Centrality rises from 0.90 to 1.03 and modularity falls from 0.769 to 0.767 after semantic linking (Abu-Rasheed et al., 21 Jan 2025).

Representation-learning and reasoning benchmarks offer a different picture. On the CS-Wiki educational graph, the joint literal-aware embedding model reports Mean Rank Raw/Filtered of 68.8/68.7 and Hits@10 Raw/Filtered of 32.0/32.0, versus substantially worse structural baselines such as TuckER with Filtered MR 919.3 and Hits@10 15.8 (Yao et al., 2019). Graphusion’s zero-shot prerequisite link prediction on LectureBankCD–NLP reaches Accuracy 0.8117 and F1 0.8181 with GPT-4o plus RAG, while TutorQA shows gains over zero-shot on all six tasks, including Task 1 accuracy 92.00 versus 69.20 and Task 2 similarity 80.29 versus 64.42 (Yang et al., 2024). By contrast, the prerequisite-inference voting framework is explicitly precision-focused: on AL-CPL Data Mining it reports Precision 1.0, Recall 0.27, and F1 0.42, and on Biology Precision 0.48, Recall 0.35, and F1 0.41 (Alatrash et al., 5 Sep 2025).

The evaluation literature therefore does not present a single dominant metric family. Entity linking uses Recall, Precision, and F1; construction pipelines use SRS accuracy and processing time; contextualization uses degree, modularity, and centrality; recommenders and QA systems use ROUGE, AUC, accuracy, or task-specific similarity; graph completion and embedding work use Hits@10, Mean Rank, and link-prediction metrics. This suggests that EduKG evaluation is inseparable from intended function.

6. Limitations, disputed assumptions, and research directions

Several recurring limitations are explicit across the literature. EDUKG reports weaker geography recall and attributes it to numerous location entities and ElasticSearch segmentation at the Chinese character level; it also notes that QA semantic parsing currently relies on template matching and identifies sequence-to-sequence semantic parsing as future work (Zhao et al., 2022). CourseMapper’s Graph RAG reports that Wikipedia-sourced answers can be abstract and that disambiguation failures remain problematic, as in “Emergency Exit” in UX versus building evacuation (Abdelmagied et al., 15 May 2025). The HCMUT pilot notes that its Neo4j-based triples retriever lacks ranking and redundancy elimination, while its unsupervised intent discovery still produces noise and duplicates (Bui et al., 2024).

Automatic construction from PDF and slide sources remains brittle. The optimized CourseMapper pipeline still lists complex tables, formulas, ligatures, hyphenation, and richer layout analysis as open challenges, and it explicitly states that the current graph focuses on content relations rather than pedagogical relations such as prerequisite chains and difficulty levels (Ain et al., 5 Sep 2025). The Top-down versus Bottom-up comparison likewise emphasizes that automatic quality before moderation is relatively low and that human review is necessary before publication (Ain et al., 15 May 2025). In higher education, the LLM-assisted ontology work reports high extraction accuracy on two modules but only modest structural changes after cross-linking, and it identifies broader scaling across faculties and institutions as future work (Abu-Rasheed et al., 21 Jan 2025).

A more substantive controversy concerns what kinds of relations should be privileged in an EduKG. Some systems emphasize instructional discourse and rhetorical roles, others prioritize prerequisites and curriculum sequencing, and still others focus on semantic similarity or linked encyclopedic content (Zhao et al., 2022, Yang et al., 2024, Alatrash et al., 5 Sep 2025). The prerequisite-inference study explicitly sacrifices recall to maximize precision, whereas Graphusion allows LLM-mediated conflict resolution and novel edge discovery without explicit numeric confidence aggregation (Alatrash et al., 5 Sep 2025, Yang et al., 2024). This suggests a trade-off between conservative graph quality and broader semantic coverage.

Privacy and provenance also remain unresolved in many deployments. PKG overlays in MOOC systems store did_not_understand annotations, mastery estimates, and activity logs, so access controls, anonymization, and policy compliance become part of graph maintenance rather than merely an external application concern (Abdelmagied et al., 15 May 2025). Resource-focused teaching KGs emphasize PROV-O, licensing, and FAIR-aligned practices, especially when materials are linked from lecturer Personal KGs or when access is partly restricted (Ilkou et al., 2024). FOKE similarly highlights fairness, transparency, and privacy of user profiles as central design constraints in intelligent educational services (Hu et al., 2024).

The main research trajectory is therefore not simply toward larger graphs. It is toward EduKGs that are more sustainable, more provenance-aware, more pedagogically typed, and more tightly coupled to reasoning and learner modeling. Explicit future directions already named in the literature include better candidate generation for educational entity linking, richer learning resources beyond Wikipedia, sequence-to-sequence semantic parsing for QA, domain-adaptive prerequisite inference, integration of prerequisite and competency relations, coreference resolution, multilingual and multimodal support, human-in-the-loop validation, and standardized evaluation protocols across intelligent education systems (Zhao et al., 2022, Abdelmagied et al., 15 May 2025, Ain et al., 5 Sep 2025, Abu-Rasheed et al., 21 Jan 2025).