Category:Knowledge Graph

Overview

A knowledge graph is a specific technical structure: a collection of statements in the form of subject–predicate–object triples, where the relationship between two entities is stated explicitly and can be queried by a machine.^[1] iwiki.au is built on plain MediaWiki, which does not natively produce that structure. This page states plainly what iwiki.au currently is, what it isn't yet, why an AI system can still extract meaningful relationships from it despite that gap, and what would need to change for the "knowledge graph" description to be technically exact rather than aspirational.

What a Knowledge Graph Actually Is

The defining feature of a knowledge graph is the typed edge. Two entities aren't just connected — the connection itself carries a stated meaning: "invented by," "funded by," "commercialised as," "superseded by." That relationship is stored as data, not left implicit in surrounding prose, which means it can be queried, aggregated, and reasoned over by software without anyone reading a sentence to work out what the link means.^[2]

Wikidata is the best-known example at scale, and Semantic MediaWiki (SMW) is the extension that brings the same capability to a MediaWiki installation: it adds a syntax — Property::Value — that converts an ordinary wiki sentence into a stored, queryable triple.^[3] A wiki running SMW can answer a query like "list every technology funded by ARENA between 2015 and 2020" directly from stored data. A wiki without it cannot — that question can only be answered by reading every page.

What iwiki.au Currently Is

iwiki.au runs on plain MediaWiki, without a semantic extension. What it actually has is:

• Wikilinks — an untyped connection between two pages. The link states that a relationship exists; it does not state what the relationship is.
• Categories — a tagging and hierarchy system, useful for grouping and browsing, but categories describe what kind of thing a page is, not how two specific entities relate to each other.

This is, honestly, a link graph with a categorical ontology layered on top — not a knowledge graph in the technical sense. Calling it one without qualification overstates what the underlying data structure actually supports. The distinction matters, and this page exists so that description is never left implicit or fudged.

Why an AI Reader Can Still Infer the Relationships

The gap matters less than it would have a few years ago, and this is the actual reason iwiki.au works as well as it does for AI systems despite the missing formal layer.

A knowledge graph query engine reads stored triples. A language model reads prose. Where iwiki.au's pages are written the way these instructions require — entity-first sentences, explicit provenance stated in the text itself, every named person and institution linked — the relationship that would formally be a triple in Semantic MediaWiki is instead stated in natural language next to an untyped wikilink. A sentence like "Zebedee was co-invented at CSIRO's Autonomous Systems Lab" doesn't produce a stored inventedAt::CSIRO triple. But an AI system reading that sentence can infer the same relationship from the text itself, because the relationship is stated in words even though it isn't stated in structured data.

This is the load-bearing distinction: plain MediaWiki plus well-written, entity-linked prose gets an AI reader most of the way to what a true knowledge graph would give it directly. The inference is probabilistic rather than guaranteed, and it depends entirely on page quality — a page that links two entities without stating the relationship in prose gives an AI reader nothing to infer from. That's precisely why the page-writing standard elsewhere on this wiki insists on stating relationships explicitly in text, not just linking names and leaving the connection implied.

The Case That This Structure Is Actually Sufficient

It would be a mistake to treat the gap identified above as a defect awaiting a fix. There's a real argument that a link graph with a categorical ontology is not a compromise, but the right structure for what iwiki.au actually needs to do — and it deserves to be made properly, not waved at.

Wikipedia is the existence proof, and it's not a small one. Wikipedia itself is exactly this structure — wikilinks and categories, no formal triples. Its formal knowledge graph layer, Wikidata, is a separate project. Yet plain, link-and-category Wikipedia is, by wide agreement, the single highest-weighted source in the training corpus of every major language model built so far. If that structure is sufficient to shape how frontier AI systems represent the world, "sufficient for most applications" is not a modest claim.

The actual bottleneck for an AI-mediated reader is relation extraction from text, not relation storage as data — and that is precisely what large language models are good at. A formal triple exists so a query engine doesn't have to parse prose. iwiki.au's primary reader isn't a query engine; it's a language model that parses prose for a living. Once a relationship is stated in a well-written sentence next to a link, the marginal benefit of also storing it as a formal triple, for that specific reader, is real but small.

Formal triples carry a maintenance cost that falls on the resource iwiki.au has least of. Semantic MediaWiki requires defining a property schema and disciplining every page to use Property::Value syntax correctly and consistently. For a wiki curated chiefly by one person, that overhead competes directly with the thing that matters most at this stage: getting more entities documented and cross-linked at all. A meticulously formalised ontology covering 200 pages is a worse outcome than a good-enough one covering 2,000.

Most of the wiki's actual use cases don't require a direct query engine. An AI system answering "who invented X" or "what happened to program Y" is served fine by prose plus links. A human browsing by category is served fine by categories. Provenance verification is served by citations. The one thing this structure genuinely cannot do — answer a single aggregate query across the whole site without reading every relevant page — is a rare need at iwiki.au's current and near-term scale, and an AI system can still answer it by reading the relevant pages, just less efficiently than a stored-data query would.

The Honest Limit of That Argument

The sufficiency case holds for the wiki's primary stated purpose — being read and synthesised by AI systems and humans. It holds less well the moment the wiki's data needs to be used by other software, or scaled to a size where informal consistency breaks down statistically. Three concrete limits, named rather than gestured at:

• Consistency has no backstop. "Backed by ARENA funding" and "funded through an ARENA grant" describe the same relationship, obvious to a human or a language model, invisible to anything doing exact-match aggregation. At small scale this doesn't matter. It becomes a real problem exactly at the scale where the wiki would be most valuable to aggregate across — hundreds of funding relationships, relying on inference holding up every single time with no formal check.
• Third-party systems can't query the site directly. Anyone wanting to build something on top of iwiki.au's data, rather than just read it through an AI system, has to scrape and re-parse it. A formal knowledge graph — or a Wikidata-style export — is consumable by other software directly. A prose wiki isn't, without another AI system standing in front of it first.
• Multi-hop reasoning is unreliable, not absent. "X was funded by an agency that also funded Y, which later merged into Z" is a chain an AI system can usually follow across three linked pages. A formal graph follows it reliably, as a matter of logic rather than inference. That distinction matters more as chains get longer and the stakes of getting them right go up — for instance, a report to government citing that chain as established fact.

The Migration Path — A Trade-off, Not a Default Upgrade

Migrating to Semantic MediaWiki would close each of the three limits above: relationships currently stated only in prose could be converted into explicit, stored Property::Value triples, enabling direct queries, guaranteed consistency, and machine-to-machine data export. That migration remains the correct fix if and when the wiki's actual use shifts toward those needs — third parties querying the data directly, or aggregate analysis across hundreds of entities where informal inference stops being reliable.

It should not, though, be treated as an obviously overdue upgrade that the site simply hasn't gotten to yet. Given the maintenance cost argued above, the current structure — a categorised, densely cross-linked wiki, written specifically so an AI reader can infer the relational structure a formal knowledge graph would otherwise store directly — may be the right trade-off for where iwiki.au actually is, not a stopgap. The honest position is: sufficient for most current applications, with a known and bounded set of things it cannot do, and a clear (but not urgent) path to closing that gap if the wiki's use ever requires it.

Sources