Before machine learning dominated AI, researchers faced a direct question: how do you store what a machine "knows" in a form it can reason over? Knowledge representation (KR) is the subfield devoted to that question — designing data structures and formalisms that capture facts, relationships, categories, and rules in a way a computer can query, update, and draw inferences from.

The goal wasn't just storage. A good knowledge representation should support inference: if the system knows "a canary is a bird" and "birds can fly," it should be able to derive "a canary can fly" without that fact being explicitly stored. This inferential capability is what separates knowledge representation from a plain database.

Semantic networks: knowledge as a graph

A semantic network represents knowledge as a graph of nodes (concepts, objects, or categories) connected by labeled edges (relationships). A typical semantic network might have a node for "Canary" connected by an "is-a" edge to "Bird," which connects via another "is-a" edge to "Animal," alongside "has-property" edges like "Bird — can — Fly" and "Canary — has-color — Yellow."

The power of this structure is inheritance: properties attached to a general category automatically apply to everything below it in the hierarchy, unless explicitly overridden. "Canary can fly" doesn't need to be stored directly — it's inferred by traversing from Canary up to Bird and reading off Bird's properties. This is essentially object-oriented inheritance, developed independently in AI research before it became a mainstream programming paradigm.

text

Note

Penguins can't fly despite being birds, and real-world knowledge is full of such exceptions. Naive semantic networks need explicit override mechanisms (non-monotonic reasoning) to handle this — the system must be able to retract or suppress an inherited default when a more specific fact contradicts it. Getting this right without breaking other inferences turned out to be a genuinely hard logical problem, not a minor implementation detail.

Frames: bundling knowledge into structured objects

Marvin Minsky proposed frames in 1974 as a richer alternative to bare semantic networks. A frame is a structured record representing a stereotyped situation or object — like a "Restaurant Visit" frame — with named slots for its expected components (customer, waiter, order, bill, tip) and default values that fill in when specifics aren't known.

Frames also support procedural attachment: a slot can carry not just a static value but a small procedure that runs when the slot is accessed, read, or changed — for example, automatically computing a "total" slot whenever the "items" slot is updated. This made frames closer to object-oriented classes with methods than to a plain graph, and frame systems directly influenced the design of object-oriented programming languages.

Semantic networks vs. frames vs. modern knowledge graphs

Property	Semantic networks	Frames	Modern knowledge graphs
Structure	Graph of nodes and labeled edges	Structured records with slots and defaults	Large-scale graph, triples (subject-predicate-object)
Inheritance	Via is-a links, informal	Via frame hierarchies, with overrides	Via explicit ontology/schema classes
Scale	Hand-built, small (hundreds of nodes)	Hand-built, small to medium	Billions of entities (Google, Wikidata)
Population method	Manual encoding by knowledge engineers	Manual encoding by knowledge engineers	Automated extraction from text, web data, user edits
Use today	Pedagogical, legacy expert systems	Influenced OOP design, legacy systems	Search engines, recommendation, LLM grounding (RAG)

Why hand-built knowledge representation hit a wall

Semantic networks and frames worked well in narrow, carefully curated domains, but they had a scaling problem that mirrors the one expert systems faced: every fact and relationship had to be entered by a human knowledge engineer. Common sense knowledge — the kind a five-year-old has effortlessly, like knowing water is wet or that you can't be in two places at once — turned out to require an impractically large number of hand-coded facts and exceptions to capture reliably, a challenge the CYC project spent decades attacking directly.

This manual bottleneck is precisely what machine learning sidestepped: instead of hand-encoding facts, learn statistical patterns directly from large corpora of text or data.

Note

Modern knowledge graphs — Google's Knowledge Graph, Wikidata, enterprise knowledge graphs used in search and recommendation — are direct descendants of semantic networks, using the same node-and-edge structure (often formalized as subject-predicate-object triples, like "Paris — capital-of — France"). The difference is scale and population method: rather than hand-building hundreds of nodes, these systems extract millions of entities and relationships automatically from structured data and web text, then use the graph for fast relationship queries that would be slow or impossible for a pure neural model to answer reliably.

The modern synthesis: knowledge graphs plus neural models

Today's systems often combine both worlds rather than choosing one. A retrieval-augmented generation (RAG) pipeline might use a knowledge graph to retrieve precise, verifiable facts and relationships, then hand them to a language model to synthesize into fluent prose — getting the graph's precision and auditability together with the language model's fluency and generalization. This hybrid pattern — symbolic structure for facts, neural models for language and pattern recognition — is a recurring theme across modern AI, not a one-off trick.

What's next

Semantic networks were one half of classical symbolic AI; the other half was using represented knowledge to actually reason and make decisions. The next module, Expert Systems, covers how rule-based inference engines used exactly this kind of structured knowledge to replicate human specialist reasoning — and why that approach eventually hit its own wall.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →