knowledge representation
/ NOL-ij rep-ri-zen-TAY-shun /
Knowledge representation is the problem of writing down what a machine knows about the world in a form it can actually use to reason. A fact obvious to you — 'a bird usually flies, but a penguin doesn't; your sister's son is your nephew' — has to be encoded in some explicit, structured shape before a computer can draw conclusions from it. Knowledge representation is the study of which shapes work well: how to capture facts, categories, relationships, rules, and exceptions so a machine can connect them.
There are many forms. Logic uses statements like 'all humans are mortal' that can be combined by inference. Semantic networks and knowledge graphs lay out concepts as nodes joined by labeled links ('Paris —is-capital-of→ France'). Frames and ontologies organize things into categories with properties and defaults. Each choice is a trade-off between how expressive it is (can it say subtle things?), how easy it is to reason with efficiently, and how naturally it handles the awkward realities of knowledge — uncertainty, exceptions, and change.
It became famous as the heart of symbolic AI and expert systems: capture human expertise as represented knowledge, then reason over it. Its hard-won lesson is humbling. Representing everyday common sense explicitly — the vast web of unspoken facts a child knows — proved extraordinarily difficult and never fully succeeded. Modern neural systems sidestep hand-built representations by learning their own internal ones from data, which is flexible but opaque. The field hasn't vanished, though: knowledge graphs quietly power search engines and assistants, and combining learned and explicit knowledge is an active frontier.
A knowledge graph stores facts as linked triples: (Penguin —is-a→ Bird), (Bird —can→ Fly), (Penguin —cannot→ Fly). The exception in the third triple is exactly the kind of thing that makes representation hard: the machine must know to let the specific fact override the general rule, or it will happily conclude that penguins fly.
Facts as linked triples — and the exceptions that make representing common sense so hard.
The enduring lesson of knowledge representation is how staggeringly much 'obvious' common sense humans carry without ever stating it — and how hard that is to write down. Neural networks learn their own representations instead, gaining flexibility but losing the readability symbolic ones had.