Skip to content

Knowen white paper · November 2026

From knowledge graphs to explanation graphs

How ontologies, knowledge graphs and explanation graphs organize different layers of knowledge, and why cumulative understanding needs all three.

Three structures, three jobs

Ontology — the vocabulary and rules
subClassOfMetalclassMaterialclasshasPropertyproperty
Knowledge graph — statements about the world
has resistivityisACopper1.68×10⁻⁸ Ω·mMetal
Explanation graph — structured explanation of the world
requiresexplained throughElectricalresistivityElectric currentCharge-carrierscattering→ requires Phonons

Ontologies formalize a domain’s vocabulary and constraints: classes, properties, individuals, and rules about how they can relate. They are designed for semantic precision and machine reasoning.

Knowledge graphs represent facts about entities and their relationships. Their edges often carry much of the meaning: Copper — has crystal structure — FCC; Scientist — authored — Paper.

Knowen explanation graphs organize understanding. Each node is a self-contained explanatory unit, and directed edges indicate concepts or explanations needed to understand, justify or develop another node. A topic decomposes recursively, from a high-level explanation into reusable constituent concepts.

The key architectural difference is where the knowledge lives. In a knowledge graph, meaning is heavily encoded in typed edges and triples. In an explanation graph, most of the knowledge lives inside rich nodes — text, equations, code, evidence, diagrams, decisions — while edges organize how those explanations depend on one another.

A note on terminology. “Explanation graph” is not a new phrase: it appears in explainable-AI and structured-reasoning research (for example ExplaGraphs [7]). Knowen uses it in a broader knowledge-organization sense: a persistent, document-rich graph whose nodes are reusable explanations and whose edges encode explanatory dependency.

Comparison at a glance

DimensionOntologyKnowledge graphKnowen explanation graph
Primary jobDefine semantic vocabulary and constraintsRepresent entities, facts and relationsOrganize reusable explanations and explanatory dependencies
Typical nodeClass, property, individualEntity, concept, valueSelf-contained explanation, document, decision or evidence unit
Typical edgesubClassOf, domain, range, property restrictionTyped relation such as authored, locatedIn, hasPropertyRequires, explained through, supports, motivates, implements, verifies
Where meaning livesFormal axioms and schemaNodes plus labelled relations and triplesMostly in rich node content; edges organize explanatory structure
Typical questionWhat kinds of things and relations are valid?What is related to what?What do I need to understand this, and why should I trust it?
Human readingOften secondary to formal semanticsUseful for navigation and querying, but often tersePrimary design goal; nodes should read on their own
Machine reasoningStrong when formalized in OWL/RDFStrong for traversal, retrieval, joinsStrong for context selection and provenance-aware retrieval
EvolutionSchema and version managementAdd and update facts and relationsContinuous revision, feedback, verification, forks, merges, refactors
Natural structureTaxonomy plus formal relationsGeneral labelled graphOften a directed acyclic graph; cross-links allowed
Best atSemantic consistencyStructured factual retrievalCumulative understanding

Worked example 1: electrical resistivity

The same scientific subject looks very different in the three representations.

Ontology view

An ontology might define Material, Metal, TransportProperty and ElectricalResistivity, with properties such as hasTransportProperty. It can state that Metal is a subclass of Material and constrain which values a resistivity property may take. The emphasis is on semantic type and valid structure.

Knowledge-graph view

A knowledge graph might assert: Copper isA Metal; Copper hasResistivity 1.68×10⁻⁸ Ω·m; Copper hasCrystalStructure FCC. These are queryable facts: “Which FCC metals have resistivity below X?”

Explanation-graph view

An explanation graph starts with a readable node, Electrical resistivity, that explains the quantity, its physical meaning, units and role in transport. It points to the smaller set of concepts needed to understand it: electric current, electric field, conductivity and charge-carrier scattering. Scattering decomposes further into phonons, impurities and mean free path. If phonons are already explained elsewhere, the graph links to that node rather than duplicating it.

The ontology tells us how to classify resistivity. The knowledge graph tells us facts about resistivity. The explanation graph tells us how to understand resistivity.

Why the DAG matters

Scientific explanations reuse concepts. Phonons may be needed to explain electrical resistivity, thermal conductivity, superconductivity and lattice dynamics. A tree would duplicate the concept four times. An explanation DAG lets every explanation reuse the same maintained node, while keeping multiple paths into it.

TreeExplanation graphResistivityPhononsThermalPhononsSupercond.Phonons3 copies to keep in syncResistivityThermalSupercond.Phonons1 node, 3 paths
One maintained node instead of copies that drift apart.

Worked example 2: a software architecture decision

motivatesexplainsrecords discussionSecurity requirementimmediate revocationSession modelexplanationIssue #428Decisionserver-side sessionsImplementationauth/session.tsVerification18 tests pass
A project explanation graph connects requirements and explanatory context to decisions, implementation and verification.

A conventional project wiki might record “we use server-side sessions” in one document and leave the rationale scattered across issues, code, tests and chat history. A knowledge graph can connect these artifacts as entities. An explanation graph goes one step further: it makes the decision itself readable, connects it to the concepts and requirements that explain it, and keeps verification nearby.

  • Requirement: immediate session revocation and reduced token exposure.
  • Explanation: what server-side session state means and what trade-offs it introduces.
  • Decision: use server-side sessions for the initial architecture.
  • Provenance: issue, discussion, author or agent, timestamp, and prior alternatives.
  • Implementation: the source files and commits that realize the decision.
  • Verification: tests, benchmarks, reviews or observations that support the current status.

Explanation graphs are living knowledge

Knowen does not assume that the graph is complete, perfectly organized or permanently correct. That would be a peculiar requirement for a system meant to represent research and development. The graph should instead make imperfection legible and improvement cheap.

  1. Draft explanation
  2. Use and feedback
  3. Evidence and review
  4. Verified or disputed
  5. Refactor, fork or merge
  6. Improved graph

Progressive verification In development

A node can carry an epistemic or project status without pretending that all knowledge has the same authority: Draft, Supported, Reviewed and Verified, with independent flags such as Disputed, Stale or Superseded. Agents can then retrieve not only relevant content but the evidence and confidence needed to decide how far to rely on it.

Feedback, forks and refactors

  • Feedback: people or agents suggest corrections, missing dependencies, stronger evidence or clearer explanations.
  • Forks: competing explanations or alternatives coexist without forcing premature consensus. In development
  • Merges: when a synthesis emerges, branches are reconciled while preserving their lineage. In development
  • Refactors: duplicate or misplaced concepts are merged, split or re-parented, keeping provenance and history.

Provenance changes what agents can do

For an agent, retrieval is only part of the problem. It also needs to know where a statement came from, whether it was reviewed, what evidence supports it, and whether later work superseded it. A provenance-aware result answers four questions at once: what is the current explanation, why is it relevant here, what supports it, and what is its verification state?

How the three layers coexist

The strongest architecture does not force a choice. A Knowen explanation node can contain or link to structured assertions that participate in a knowledge graph, and those assertions can use an ontology for shared semantics and validation.

  1. Explanation graphReadable explanations, dependencies, provenance, evidence, review, evolution
  2. Knowledge graphEntities, assertions, structured relationships, cross-domain linking
  3. OntologyClasses, properties, semantic constraints, shared vocabulary

Where the idea comes from

The idea that knowledge has a graph structure has deep roots in epistemology and philosophy of science. Debate over how justification is structured has long set foundationalist hierarchies against coherentist webs (Sosa 1980). Thagard’s theory of explanatory coherence (1989) made this idea computational: propositions become nodes, explanatory relations become links, and which beliefs to accept is decided by the network as a whole.

Accounts of what makes an explanation good also lend themselves to graphs. On the unificationist view, explanatory power comes from deriving many results from a few shared patterns (Kitcher 1981). Other accounts tie explanation to causal or constitutive dependence, which can be modeled with directed graphs and structural equations (Pearl 2009; Woodward 2003; Halpern & Pearl 2005; Schaffer 2016). Still others ask for an explanation that keeps only the factors that make a difference (Strevens 2008) and is hard to vary (Deutsch 2011).

Learning theory, for its part, has shown that understanding is built along prerequisite hierarchies (Gagné 1968), and it has formalized such structures combinatorially (Doignon & Falmagne 1999) and in concept maps used for teaching (Novak & Gowin 1984).

Taken together, this work suggests that a body of scientific knowledge could be organized as a directed acyclic graph of self-contained explanations. Each node would be understandable from its ancestors, shared foundations would appear once and be reused, and new results would attach at the right depth. Knowen was conceived as exactly this kind of structure, but building and maintaining such graphs by hand is very labor-intensive. Large language models make it practical, for the first time, to test the idea at scale: we can generate explanation DAGs from papers and whole bodies of literature, and then measure how well they serve both learners and researchers.

What is new in Knowen’s approach

The novelty is not the bare use of a graph to represent explanations. Prior work used explanation graphs for structured commonsense reasoning and explainable-AI outputs [7], and education research has studied concept-dependency graphs [8]. Knowen differs in scope and persistence: the explanation graph is the primary, evolving substrate for ongoing work by people and agents.

  1. Document-rich nodes. Nodes hold compact articles, equations, code, figures, evidence, decisions and open questions, not short labels.
  2. Recursive explanatory decomposition. A node is analyzed into a small set of constituent concepts, built the same way, giving navigable depth.
  3. Reuse without duplication. Equivalent concepts are shared across branches, so one maintained explanation serves many contexts.
  4. Provenance and verification as first-class state. Who or what contributed a node, from which sources, and how it was reviewed are part of the knowledge object.
  5. Designed evolution. Feedback, forks, merges and refactors are expected operations, not cleanup after the knowledge base breaks.
  6. Symmetry between people and agents. Both read, contribute to, review and verify the same structure.

Conclusion

Ontologies, knowledge graphs and explanation graphs should not be collapsed into one category just because all can be drawn with boxes and arrows. Ontologies define meaning and constraints. Knowledge graphs encode structured facts and relations. Knowen explanation graphs organize understanding into reusable explanatory units that evolve through evidence, feedback, verification, forks and refactoring.

Knowledge is a process. It needs structure to improve.

Selected references

  1. W3C, “Resource Description Framework (RDF): Semantic Web Standards.” w3.org/RDF
  2. W3C, “RDF 1.0 Concepts and Abstract Syntax,” §3.1, RDF graphs as subject–predicate–object triples. w3.org/TR/rdf10-concepts
  3. W3C, “OWL 2 Web Ontology Language Primer (Second Edition).” w3.org/TR/owl2-primer
  4. W3C, “OWL 2 Structural Specification,” §5, entities and ontology vocabulary. w3.org/TR/owl-syntax
  5. Google, “Knowledge Graph Search API”: real-world entities as graph nodes. developers.google.com/knowledge-graph
  6. IBM, “What Is a Knowledge Graph?” Definitions vary across the literature and industry. ibm.com/think/topics/knowledge-graph
  7. Saha, S. et al. (2021), “ExplaGraphs: An Explanation Graph Generation Task for Structured Commonsense Reasoning,” EMNLP 2021. aclanthology.org/2021.emnlp-main.609
  8. Alsaad, F. et al. (2018), “Mining MOOC Lecture Transcripts to Construct Concept Dependency Graphs.” sundaram.cs.illinois.edu

Where the idea comes from

  • Deutsch, D. (2011). The Beginning of Infinity: Explanations That Transform the World. Viking, New York.
  • Doignon, J.-P., & Falmagne, J.-C. (1999). Knowledge Spaces. Springer, Berlin.
  • Gagné, R. M. (1968). Learning hierarchies. Educational Psychologist, 6(1), 1–9.
  • Halpern, J. Y., & Pearl, J. (2005). Causes and explanations: A structural-model approach. Part II: Explanations. British Journal for the Philosophy of Science, 56(4), 889–911.
  • Kitcher, P. (1981). Explanatory unification. Philosophy of Science, 48(4), 507–531.
  • Novak, J. D., & Gowin, D. B. (1984). Learning How to Learn. Cambridge University Press, Cambridge.
  • Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press, Cambridge.
  • Schaffer, J. (2016). Grounding in the image of causation. Philosophical Studies, 173(1), 49–100.
  • Sosa, E. (1980). The raft and the pyramid: Coherence versus foundations in the theory of knowledge. Midwest Studies in Philosophy, 5, 3–26.
  • Strevens, M. (2008). Depth: An Account of Scientific Explanation. Harvard University Press, Cambridge, MA.
  • Thagard, P. (1989). Explanatory coherence. Behavioral and Brain Sciences, 12(3), 435–502.
  • Woodward, J. (2003). Making Things Happen: A Theory of Causal Explanation. Oxford University Press, New York.