A knowledge graph represents information as identifiable things and meaningful relationships among them. A simple educational graph might say that an article is about Fractions, that the article is intended for Secondary 1 revision, and that the article cites a particular source. Instead of storing these descriptions as unconnected words, a graph makes the links explicit: article → covers concept, article → has intended level, and article → references source. A reader or software system can then ask questions involving several connected facts.
This is vocabulary functioning as technology rather than merely as a dictionary. The names of concepts and relationships become a shared language for finding resources, comparing explanations, checking claims and navigating from a familiar idea to a more specialised one. A graph can connect people to institutions, texts to topics, methods to prerequisites, and findings to their evidence. But a graph full of incorrect or ambiguous relationships is not a trustworthy brain: it is a structured collection of assertions whose identity, meaning and provenance must still be examined.
This guide explains nodes, edges, RDF triples, identifiers, vocabularies, queries, inference, linked data and source tracking through original classroom cases. It uses the World Wide Web Consortium’s RDF and SPARQL materials for the technical foundation, while translating them into accessible teaching examples. The goal is to show how carefully managed words can multiply the questions a learner is able to ask, without pretending that a diagram or database automatically understands the world.
Editorial note: All students, articles, graph records and query results used as examples are invented. The standards support the technical concepts but do not attest to the accuracy of the fictional data.
Contents: nodes, relationships, queries and knowledge reliability
Open all 55 chapters and worked graph examples
- Begin with three simple relationships
- What the graph adds to a list
- A node is something that can be referred to
- An edge names the relationship
- RDF triples: subject, predicate and object
- A graph has meaning only with interpretation
- The same resource can appear in many statements
- Identifiers are more stable than labels
- Why strings are poor stand-ins for identity
- One concept, many labels
- Typed relationships limit confusion
- Topic hierarchy is not always a prerequisite map
- Part-of and kind-of are different relations
- Related and identical are not synonyms
- Knowledge graph versus concept map
- Knowledge graph versus controlled vocabulary
- Knowledge graph versus a spreadsheet
- Knowledge graph versus an ordinary website link graph
- The first query: which resources cover this concept?
- A two-hop query finds a wider connection
- A three-hop query can answer a richer question
- SPARQL asks questions of RDF graphs
- A pattern query is not human understanding by itself
- Graph inference can extend explicit statements
- The open-world assumption changes absence claims
- Unknown is not false
- Contradictory edges are a data-quality problem
- Named graphs can help separate contexts
- Provenance should accompany consequential facts
- A citation edge is not an endorsement
- A revision edge preserves history
- Time-sensitive information requires dates
- Location fields need clear semantics
- Graph identity across languages
- Do not confuse close matches with identity
- Entity resolution requires evidence
- A knowledge graph can support question answering
- Graph search and language-model answers are different
- A knowledge graph can help locate missing coverage
- A knowledge graph can also magnify errors
- The smallest useful graph in a classroom
- Worked case: finding an algebra foundation
- Worked case: literature vocabulary and texts
- Worked case: a Science claim and its source
- Worked case: two records for one article
- Worked case: a false school-status claim
- Worked case: a multilingual learning concept
- What to ask before creating a new edge
- A simple data-quality checklist
- Privacy and sensitive learner records
- When a graph is unnecessary
- A basic lesson plan: from sentences to a graph
- Ten questions with concise answers
- Frequently asked questions
- Primary sources and your next practical step
1. Begin with three simple relationships
Imagine a fictional library containing an article named ‘Understanding Equivalent Fractions’. The article covers the concept Equivalent Fractions. The concept is associated with a broader Fractions topic. The article has a resource type, Worked Example.
Draw boxes for the article, concept and resource type, with labelled arrows connecting them. Now ask ‘Which worked examples cover ideas under Fractions?’ The links give a possible route to relevant material.
This small drawing is a conceptual illustration. A deployed graph also needs stable identities, well-defined relationships and a search mechanism that knows how to follow them.
2. What the graph adds to a list
A list can name Fractions, Algebra, Equivalent Fractions and Worked Example. It does not say which words refer to topics, which to resource types or how one article connects to another.
A graph adds the type of relation. Equivalent Fractions can be narrower than Fractions; an article can cover Equivalent Fractions; and a worksheet can be an example of a resource type.
The extra meaning is in the links and their labels. More nodes without meaningful edges can leave a graph no more useful than an unsorted word list.
3. A node is something that can be referred to
In a general graph representation, nodes stand for items such as entities, concepts, documents, events or other resources. Depending on the model, some values may be represented as literals rather than resource nodes.
A label such as ‘Algebra’ may appear on a concept node, but the label itself need not be its complete identity. Distinct resources can share names, and one resource can have several labels.
Always ask what the node represents and how its identity is established. Without that step, a graph can accidentally merge unrelated subjects.
4. An edge names the relationship
An edge connects two graph elements and can carry a relationship name such as covers, authored by, part of or cites.
The direction and semantics matter. ‘Article A cites Source B’ is not the same as ‘Source B cites Article A’. Neither is the same as saying the two texts are identical.
Use precise relation words rather than a generic arrow labelled ‘connected’. The label tells software and humans what a path means.
5. RDF triples: subject, predicate and object
The W3C RDF 1.1 Primer describes graph assertions as triples containing a subject, predicate and object. For example, in an invented graph: Article-17 — coversConcept — Fractions.
The subject identifies what the statement is about, the predicate identifies the relationship, and the object supplies the related resource or value. Several triples can share the same subject or object, forming a larger network.
This example is informal notation to teach the pattern. A machine-readable RDF file requires valid identifiers and syntax, not merely three words separated by dashes.
6. A graph has meaning only with interpretation
A syntactically correct triple can still be factually wrong. Article-17 — approvedBy — Teacher-A should not be recorded unless that approval actually occurred and the relationship is defined appropriately.
RDF defines how to represent and reason over graph statements under a formal semantics. It does not guarantee that asserted statements were verified against the real world.
A learner should separate a valid data structure from a justified claim. Reliability depends on source, scope and review as well as technical format.
7. The same resource can appear in many statements
An article can be related to a topic, author, language, publication date and source. Reusing the same identified article across statements lets the graph connect those properties.
If the system creates a new unrelated article identity for every statement, joining information becomes difficult or misleading.
Identity management is therefore a foundation of graph usefulness. Many relationships only help when they really refer to the same entity.
8. Identifiers are more stable than labels
A school may rename a resource while keeping the underlying article and its history. A stable identifier allows connections to remain attached to that resource despite the human-readable title change.
RDF commonly uses IRIs to identify resources and properties. The W3C primer explains how globally scoped identifiers can support linking across datasets.
A local educational worksheet can use simple IDs to teach the idea. A public cross-system graph needs an identifier scheme that remains consistent and correctly resolves identity.
9. Why strings are poor stand-ins for identity
Two pupils in a fictional story may both be called Alex. A graph that uses the string Alex as the only identifier could incorrectly combine their assignments and actions.
Conversely, one pupil may appear as Alexander on one school record and Alex on a local nickname list. Different strings do not automatically mean different people.
Responsible systems distinguish the entity’s identity from its display name. Sensitive personal data also demand appropriate safeguards and must not be casually assembled into a shared graph.
10. One concept, many labels
The concept ‘Equivalent Fractions’ can be displayed using a preferred label and associated with suitable alternative expressions. A graph can attach those labels to the concept’s identifier.
The W3C SKOS Reference describes properties for preferred labels, alternative labels and broader or related concept links.
The label scheme improves discovery; it does not replace a full definition or prove an educational claim. Consult the Controlled Vocabulary guide for label governance.
11. Typed relationships limit confusion
A relation such as isPrerequisiteFor suggests a particular learning dependency; isRelatedTo conveys a weaker connection. Treating them as interchangeable could make a student’s study plan overly rigid.
A graph’s predicates must be defined by the relevant application, with their intended interpretation and constraints made explicit.
Before adding an edge, ask what a reader could legitimately conclude from it and whether the source supports that conclusion.
12. Topic hierarchy is not always a prerequisite map
A concept may sit below a broader topic in a library hierarchy. For example, Equivalent Fractions may be indexed under Fractions.
That organisation helps navigation but does not establish that every article under Fractions is a prerequisite for every Equivalent Fractions task.
A graph should distinguish navigation relations from teaching-order claims, which depend on the intended learning activity and method.
13. Part-of and kind-of are different relations
A wheel may be part of a bicycle. A bicycle is a kind of vehicle. These statements support different kinds of reasoning.
Treating part-of as subclass-of would lead to absurd classifications, such as claiming that every wheel is a bicycle or vehicle in the same sense.
Label the relationship according to its actual meaning, and test it with a simple counterexample before propagating it through a graph.
14. Related and identical are not synonyms
Fractions and Percentages are related topics, but they are not identical concepts. An article about converting between them might appropriately connect both.
A graph that treats every related link as exact equivalence can combine distinct instructional content and mislead search results.
Use appropriately scoped relations. SKOS includes related and mapping properties that should not be collapsed into one generic match.
15. Knowledge graph versus concept map
A concept map is commonly an external learning aid in which a person diagrams important ideas and labelled connections. A knowledge graph is a structured representation that may be designed for machine processing and queries.
Both can show nodes and edges, but a drawn diagram does not automatically provide formal identifiers, query semantics or validated data.
The Cognitive Artefact guide explores how a diagram can support human reasoning. A production knowledge graph adds engineering and governance requirements.
16. Knowledge graph versus controlled vocabulary
A controlled vocabulary manages the labels and relationships of concepts used to describe information. A knowledge graph can contain those concepts while also connecting specific documents, events, people and findings.
The controlled vocabulary is often part of the graph’s semantic infrastructure, but the two are not identical. A scheme of educational topics can exist without resource-level facts.
Build clear terms first where necessary, then connect them to verified information. An uncontrolled graph may amplify confusion across a larger network.
17. Knowledge graph versus a spreadsheet
A spreadsheet stores rows and columns with explicitly named fields. Many useful questions can be answered through it, and a graph should not be chosen merely because the diagram looks impressive.
A graph can become attractive when the same entities have varied relationships and queries traverse several kinds of links. An article may reference a study that was revised from an earlier study and concerns a concept related to another topic.
Choose the representation based on the required questions, the data and maintenance cost. A small school library may be better served initially by well-designed tables.
18. Knowledge graph versus an ordinary website link graph
Web pages often link to one another using hyperlinks. Those links enable navigation but may not describe the precise relationship between their contents.
A structured graph can represent ‘teaches topic’, ‘requires concept’, ‘cites source’ or ‘is revision of’ as distinct predicates rather than treating every hyperlink as an identical edge.
Both kinds of linking can coexist. A useful page link can lead the reader to an explanation while a structured relation tells software why the connection exists.
19. The first query: which resources cover this concept?
Assume the fictional graph contains Article-A covers Equivalent Fractions and Article-B covers Percentages. A simple concept query can find Article-A when the user asks for resources covering Equivalent Fractions.
If a document covers several topics, it can have several appropriate edges. A result can also include a title and accessible link from other stored properties.
The query is only as trustworthy as the source tagging. Incorrect topic edges can return inaccurate recommendations.
20. A two-hop query finds a wider connection
Suppose Equivalent Fractions is indexed as narrower than Fractions. A query might first identify narrower concepts under Fractions, then find resources that cover those concepts.
This allows a reader to navigate from a broad learning topic to a specialised resource without guessing its exact title.
Not every query engine automatically expands hierarchies. The data model, relation semantics and chosen query logic determine whether this behaviour occurs.
21. A three-hop query can answer a richer question
An invented collection connects resources to topics, topics to broader subjects, and resources to resource types. A user might ask for worked examples about topics under Number.
A graph query can join these relationships if the required edges and values exist. The result still needs clear descriptions so the learner can decide which item to read.
More hops do not mean more truth. A single mistaken relationship can produce an invalid answer that looks impressively connected.
22. SPARQL asks questions of RDF graphs
The W3C SPARQL Query Language defines query patterns for RDF data. A basic graph pattern matches combinations of subject–predicate–object relationships.
In ordinary words, a query can request resources that both cover a given concept and have a specified resource type, or ask which sources an article references.
Students do not need to learn SPARQL syntax to grasp the educational value of typed relations. Programmers implementing a live system should follow the relevant specification and query engine behaviour.
23. A pattern query is not human understanding by itself
A graph returns every article marked teaches Algebra. That does not mean the system has read each article and independently verified its instructional quality.
It has matched represented relationships under the query rules. A reviewer may still need to inspect the content and the intended learner’s prerequisites.
Use graph search to narrow the investigation, then apply subject knowledge and evidence checks before presenting a high-confidence learning recommendation.
24. Graph inference can extend explicit statements
Some graph systems support inference from declared relationships and logical rules. For example, a formal property hierarchy may allow a more specific relation to imply a broader one under appropriate semantics.
Other structures, such as many SKOS hierarchies, express navigation relationships whose implications are not identical to strong ontology-class reasoning.
Before promising automatic deductions, inspect the exact vocabulary, reasoning regime and application logic. An arrow on a diagram is not automatically a valid theorem.
25. The open-world assumption changes absence claims
In many RDF/OWL contexts, an absent statement does not mean the negation is true. If the graph contains no recorded author for one article, that alone does not establish that the article has no author.
A database interface might choose closed-world assumptions for a specific checked inventory, but that is a separate decision about its semantics and completeness.
Teach learners to distinguish ‘not recorded in this graph’ from ‘does not exist in the world’. This prevents a dangerous leap from a missing edge to a universal claim.
26. Unknown is not false
An educational graph records that a worksheet has a subject and stage but no verified answer-key status. The honest value may be unknown.
A user searching for resources with answers should not be told that the worksheet definitely lacks answers merely because the field was not filled.
Represent uncertainty and data completeness explicitly where they affect retrieval. Good graph design protects information gaps rather than concealing them.
27. Contradictory edges are a data-quality problem
A fictional record is tagged both as Approved and Awaiting Review under a policy that makes those states incompatible. The graph may store both assertions without automatically resolving the contradiction.
One may be outdated, refer to a different version or simply be wrong. The editor must inspect time, source and scope.
A model with timestamps and provenance can help diagnose the conflict. Syntactic graph validity is not sufficient to establish the current truth.
28. Named graphs can help separate contexts
RDF datasets can include named graphs as well as a default graph. These allow an application to group statements by graph context in a structured way, though the semantics depend on how they are used.
One application could keep approved editorial statements separate from imported provisional research claims. That separation does not by itself verify either graph’s accuracy.
The W3C RDF Primer and SPARQL specifications provide formal background. A learning guide can treat named graphs as one optional technique rather than a universal implementation rule.
29. Provenance should accompany consequential facts
If an article is recorded as approved, a useful system may also preserve who authorised the state, under what policy, and which version was checked.
Without that history, another editor may be unable to distinguish a legitimate current approval from an obsolete label. This is a provenance problem rather than a graph-shape problem.
The companion guide to provenance trails explores how entities, activities, agents and revisions can make the route from source to result inspectable.
30. A citation edge is not an endorsement
An article may cite a research paper as background, as a contrary example or as a source of a particular fact. A generic cites edge records a relationship but may not explain how the source supports a particular claim.
An editor should retain the specific claim and relevant evidence location when the relationship matters for fact checking.
Do not treat ‘has a citation’ as equivalent to ‘is scientifically verified’. A graph can carry a reference without proving the inference drawn from it.
31. A revision edge preserves history
An article may be revised after an error is corrected. Connecting the new version to the previous one can support comparisons and an audit of changes.
A revision relation should not imply that the earlier content was always identical to the new version. The change may alter definitions, numbers or relevant conditions.
Preserve enough temporal and editorial context for readers who encounter a historical citation or link.
32. Time-sensitive information requires dates
A statement that a school session begins at four may be accurate for one timetable and wrong after an update. A graph should not treat the time value as timeless if it has an effective period.
Temporal properties, event versions or validity intervals can help represent that scope. The exact data model depends on the domain and whether the source supplies trustworthy dates.
Do not invent an expiry date or current status where no authoritative update exists. A system should distinguish known dates from assumptions.
33. Location fields need clear semantics
A lesson article about a city and a tuition centre physically located in that city do not have the same relationship to the place.
One resource is about a location; the other has a service address. If the graph uses a single vague location edge, a user can get misleading results.
Choose predicates such as discussesPlace or serviceLocatedAt only when the distinctions match the domain and verified records.
34. Graph identity across languages
A concept may have labels in English, Mandarin or another language. Keeping one well-scoped concept identifier can help users search across language labels.
But translation can change nuance and a single-word match may not preserve the concept’s full meaning. A source term might require a phrase or scope note.
Check mappings with relevant linguistic and subject expertise, especially for official terms and specialised concepts.
35. Do not confuse close matches with identity
Two course titles may cover similar material but have different eligibility, academic stages and assessment conditions. An unsafe graph merge can imply formal equivalence that does not exist.
A close topical relationship is useful for discovery but weaker than an official qualification equivalence.
Do not use same as as an attractive replacement for a more accurate related or mapped concept link.
36. Entity resolution requires evidence
An imported article title resembles an existing one. Are they two versions of the same page, an excerpt and its source, or separate publications about the same topic?
The graph should not merge them solely because the strings match. Inspect identifiers, authorship, dates, URLs and other relevant attributes.
A mistaken identity merge can spread false author, citation and approval information across many connected records.
37. A knowledge graph can support question answering
A properly maintained graph can help answer structured questions such as ‘Which published resources on fractions are intended for this learning stage and have a verified source link?’
The question still relies on each part being recorded with the intended meaning. If a source field is incomplete, the result may be partial.
The system should disclose important uncertainty rather than invent missing relationships to produce a confident-sounding answer.
38. Graph search and language-model answers are different
A language model may summarise a topic from patterns in text, while a graph query can match specified stored relationships. They offer different kinds of evidence and failure modes.
Using a graph as a retrieval layer may improve traceability when answers cite the returned entities and sources. It does not remove the possibility of misunderstanding a question, selecting a wrong record or producing a misleading summary.
Keep the answer linked to the actual evidence and distinguish retrieved facts from generated explanation.
39. A knowledge graph can help locate missing coverage
An editorial system may record a concept such as Equivalent Fractions but find no reviewed worked examples attached to it for a particular study purpose.
That could indicate a genuine content gap, a classification omission or an incomplete import. A graph can help frame the question but cannot identify the cause without checking the underlying data.
Use a review queue instead of automatically generating duplicates whenever a query returns no matching resources.
40. A knowledge graph can also magnify errors
Suppose the graph incorrectly says Algebra is a subtype of every Secondary subject. Automated broader-topic queries could then retrieve large amounts of irrelevant content.
The error can spread through traversal and inference, making the result set appear comprehensively connected while being wrong.
Graph quality depends on disciplined relation definitions, validation and correction—not simply on adding more edges.
41. The smallest useful graph in a classroom
Begin with five concept cards and three document cards. Give each item a unique identifier. Add labelled arrows for coversConcept, narrowerTopic and citesSource where appropriate.
Ask another learner to answer a two-hop question using only the graph. Then include one deliberately wrong link and observe how it changes the apparent answer.
The activity demonstrates why explicit relations and data checks matter. It does not require deploying a server or publishing personal learner records.
42. Worked case: finding an algebra foundation
A fictional catalogue stores an article about signed numbers, another about simple linear equations and a third about evaluating expressions. A tutor asks which resource addresses a student’s demonstrated difficulty with signed-number operations.
If the graph accurately connects the learner’s task demand to relevant concepts, it can help locate candidate materials. But the actual learning recommendation still depends on diagnosing the learner’s work.
Do not infer that every signed-number article is an automatic prerequisite for every algebra problem. The Prerequisite guide explains this important limit.
43. Worked case: literature vocabulary and texts
A school library identifies a particular story, its author, a theme concept and two vocabulary guides discussing that theme.
A graph can connect the story to the author and the guides to the theme without claiming the guides themselves were written by the story’s author.
Incorrect edge directions or shared-name identity mistakes would produce misleading results. Ask learners to read each labelled link aloud before following the path.
44. Worked case: a Science claim and its source
An article describes evaporation and cites an appropriate scientific explanation. The graph stores an article-to-source relationship and an article-to-topic relationship.
That helps a researcher find the cited item and related topic resources. It does not establish that every sentence of the article is supported by the source.
Record claim-level support or editorial verification where a specific statement matters. The graph should preserve evidence rather than imply it from proximity.
45. Worked case: two records for one article
An education library imports a page under two slightly different titles and dates. The entries might be the same article, a revised version or two related articles.
An editor checks the stable URL, original publication identity and version history before merging or connecting them.
Only after the evidence supports a shared identity should the graph reuse one node or record an appropriate revision link.
46. Worked case: a false school-status claim
A fictional graph records that a workbook was approved for use. The underlying document shows that only the draft was sent for review.
The approval edge is unsupported. Remove or correct it through a documented editorial workflow and preserve the legitimate received-for-review event if it occurred.
That repair illustrates why provenance and typed state relations can be as important as subject indexing.
47. Worked case: a multilingual learning concept
One concept for basic fraction comparison has carefully reviewed labels in two languages. The graph connects both labels to a stable concept ID and links a suitable translated explanation to that concept.
A user can potentially find relevant resources through either language, provided the search application supports the labels. The translated article’s language and review status remain separate metadata.
Do not declare every bilingual phrase exact merely because both appear under the same topic. Review the actual content and terminology.
48. What to ask before creating a new edge
Which two entities are being connected? What precisely does the predicate assert? Which evidence supports it? What scope or date limits it? Who can correct it if new information appears?
Writing those five answers may reveal that the proposed relationship is only an assumption or a broad topical association.
A reliable small graph often helps more than a huge collection of poorly labelled links.
49. A simple data-quality checklist
Check for duplicate identities, ambiguous labels, false synonyms, reversed links, missing source references, old statuses and relations whose meaning has not been defined.
Add a few negative tests: ask a question that should not return a given article, and check whether the graph incorrectly includes it. Positive-only tests can miss false mergers.
Record observed errors and revisions instead of presenting one tidy visualisation as proof the data are correct.
50. Privacy and sensitive learner records
A graph can connect many pieces of information about a person, potentially revealing more than each record alone. Student performance, contact details and private notes deserve particular care.
Build public concept and content graphs without incorporating personal learner data unless there is a legitimate, authorised need and appropriate safeguards. The fact that a relationship can be represented does not mean it should be published.
Data-minimisation and access controls are governance requirements beyond basic RDF syntax.
51. When a graph is unnecessary
A small collection with five neatly described articles may not need a graph engine. Clear category pages, metadata tables and links may meet the users’ needs.
A knowledge graph becomes useful when questions genuinely traverse multiple relations and the team can maintain the identities and definitions.
The technology should follow the purpose. The attractive name knowledge graph is not a sufficient reason for adopting complexity.
52. A basic lesson plan: from sentences to a graph
Give learners six fictional factual sentences about three educational resources. Ask them to underline entities and circle relationship expressions.
Convert each appropriate sentence to a subject–predicate–object statement. Then ask two questions that require joining statements. Finally, insert an unsupported sentence and let students explain why the graph’s answer could become wrong.
The activity shows how vocabulary can organise inquiry. It does not claim the students have built a production RDF service or validated a full ontology.
53. Ten questions with concise answers
1. Is a knowledge graph just a bag of words? No. 2. What makes an edge useful? Its defined relation. 3. Can the same resource participate in many triples? Yes. 4. Does a valid triple guarantee a true claim? No.
5. Can graphs support multi-hop queries? Yes. 6. Is every missing edge proof of a negative? No. 7. Are related and identical the same? No. 8. Can controlled labels improve identity and search? Yes, when used correctly.
9. Do RDF and SPARQL replace human fact checking? No. 10. What is the practical advantage? Explicit concepts and relations let people and tools ask structured questions across connected information.
54. Frequently asked questions
Are all knowledge graphs RDF? No. RDF is a widely used standards-based graph data model, but knowledge graphs can use other representations. Is a knowledge graph the same as AI? No; it may support AI workflows but is a distinct information structure.
Can I start with paper cards? Yes, to understand the concept. A live queryable implementation needs structured data and software. Will a graph make an article accurate? No. Accuracy depends on evidence, interpretation and maintenance.
Why does vocabulary matter? Names and relationships must carry stable meanings before connected records can be reliably queried or explained.
55. Primary sources and your next practical step
The W3C RDF 1.1 Primer explains triples, identifiers and graph structure. The SPARQL Query Language specifies pattern queries against RDF data. The SKOS Reference describes controlled concepts and relationships that can support an information graph. The PROV Ontology offers ways to connect entities to activities, agents and derivations.
All educational entities, queries, diagrams and results in this article are invented examples. The standards define representations and operations, not a guarantee that any given graph has been reviewed or its statements are true.
Choose three learning resources and two concepts. Give them stable local identifiers, connect them with clearly defined relations, then ask a question that requires following two links. Add a source or uncertainty note for every claim that matters. This is vocabulary becoming a usable graph of knowledge.