4.12 Semantic networks, extended semantic networks and frames

What this is and why it exists

A semantic network is concepts joined by labelled links. A frame is the same knowledge turned inside out — grouped by object into named slots with defaults. Both get their compactness from inheritance, and both get their characteristic bug from it too, because exceptions break inheritance and the world is full of exceptions. These notations are the direct ancestors of today's knowledge graphs and of every ontology you will meet, so the strengths and the failures both carry forward.

The vocabulary

• Node — a concept or an individual in the network.
• Link — a labelled relationship between two nodes.
• is-a link — a link asserting membership or subtype, along which properties are inherited.
• Property link — a link attaching an attribute value to a node.
• Inheritance — properties stated high in a hierarchy applying to everything below.
• Exception — a case contradicting an inherited property.
• Frame — all knowledge about one concept, grouped into named slots.
• Slot and default — a named attribute, and the value assumed when none is given.

The mental model

A semantic network is a graph of meaning. Nodes are concepts or individuals; links are labelled relationships between them. Two link types do most of the work. Property links attach attributes: a bird node has a link labelled "covering" to a feathers node. The is-a link asserts that one thing is a kind of another, or an instance of it, and it is the link along which properties travel downward.

That single mechanism is where the compactness comes from. State once that birds have feathers, that they lay eggs, that they have two legs, and then every kind of bird and every individual bird has all of it without repetition. A hierarchy of a hundred concepts can carry what would take thousands of separate statements, and asking whether a particular sparrow has feathers is a short walk up the links rather than a lookup that must have been prepared.

Building one from a paragraph of English is the exercise, because of what it reveals. Take an ordinary paragraph and draw it. Two things become apparent almost immediately. First, most of the content is relationships rather than words — the nouns are nodes and the interesting information is in the verbs and prepositions connecting them, which is a genuinely surprising ratio the first time you count it. Second, a great deal is left implicit, and you cannot draw it without deciding what the sentence assumed. "The manager approved the request on Tuesday" needs nodes for a manager, a request and a day, links for the approval and its time, and a decision about whether approval is a relationship between two things or an event with participants. The English did not say, and the network must.

That difficulty is the honest lesson: converting language into a representation forces choices the language avoided, and different people produce different networks from the same paragraph. Making those choices consistently across a large body of text is what an ontology is, and it is why ontologies are expensive.

Inheritance is the strength and the failure, and the failure has a name. Properties passed down a hierarchy save enormous repetition and break the moment a subordinate contradicts an inherited property. Birds fly; penguins are birds; penguins do not fly. The network says they do, confidently and compactly.

The obvious patch is to let a more specific statement override a more general one — attach "cannot fly" directly to the penguin node and let the nearer value win. That works for the simple case and creates two harder ones. Multiple inheritance: if a concept inherits from two parents with conflicting values, which wins? Nothing in the structure decides it, and it must be settled by an ordering rule that is itself a design choice. And cancellation without limit: if any property can be overridden anywhere below, then nothing inherited is reliable, and a system cannot conclude that something flies from its being a bird — which was the entire point of the hierarchy. Where you draw the line between an exception and a wrong hierarchy is a modelling judgement, and the practical rule is that when exceptions become common the hierarchy was wrong: split flying birds from flightless ones and state the property where it is actually true.

Ordinary networks cannot express quantification or negation, which is the other structural limit. Drawing a link says something is so. There is no way in the basic notation to say "no student passed", or "every employee has exactly one manager", or "some machine in the room is faulty but not which one". These are not exotic statements — they are ordinary business rules — and the notation cannot hold them at all, which is a representational adequacy failure in the previous topic's terms.

Extended semantic networks add machinery for exactly this: partitioning the network into spaces so that a statement can be about a whole space rather than about individual nodes, letting a link be quantified or negated as a unit. The mechanism amounts to importing the expressiveness of logic into the graph notation, which is why the extension is fair to describe as a graph-shaped logic. The cost is the one the representation topic predicted: as expressiveness rises, inference gets harder and the readability that made the plain network attractive is largely spent.

Frames are the same knowledge grouped the other way. Instead of a graph seen link by link, a frame collects everything known about one concept into named slots with values, and a slot may carry a default used when nothing more specific is known. A bird frame has slots for covering, locomotion and number of legs, with defaults of feathers, flying and two. A penguin frame is-a bird and overrides the locomotion slot. This is very close to how an object is described in programming, and the resemblance is not accidental — this line of work fed directly into object-oriented design.

The differences from a network are practical rather than fundamental: the same knowledge is expressible in both, and the grouping changes what is convenient. Frames make "tell me everything about this concept" immediate and make "find everything related to that concept" a scan. Networks are the reverse. Frames also make defaults explicit and first-class, which is useful, and they add procedural attachment — a slot whose value is computed by a procedure when asked for, or that triggers one when set — which quietly brings in the procedural knowledge from the previous topic.

The comparison on the four adequacy properties, which is the standard examination question.

Read the table as a trade rather than a ranking. The plain network wins on readability and acquisition, which is why knowledge graphs — its modern descendant — remain popular for exactly that reason. The extension buys expressiveness at the cost of both efficiency and acquisition. Frames win on organisation and on the practical matter of getting knowledge in, because the mental model is familiar. No column dominates, which is the point the representation topic made in general and this table makes concrete.

What you should now be able to explain or do

Draw a semantic network with concept nodes, relationship links and property links. Build one from a paragraph of English and say what the exercise reveals about relationships and implicit content. Use inheritance through is-a links and explain the compactness it buys. State the exception problem, the two harder cases the obvious patch creates, and the practical rule for when the hierarchy is wrong. Say what plain networks cannot express and what the extension adds, with its cost. Represent the same domain with frames, slots and defaults, and say what procedural attachment brings. Compare all three on the four adequacy properties.

Check yourself