Semantic networks are our first knowledge representation. We use them to represent problems and reason over them—especially for visual analogy tasks like Raven's Matrices.
Encode knowledge & problem → Use it to find the answer
How we encode what we know so an agent can use it.
Semantic nets work with Generate and Test, Means-Ends, Problem Reduction.
Core use: A is to B as C is to ? — complete the pattern.
Every semantic network has three parts:
Nodes — objects or concepts (e.g. shapes, entities). The vocabulary.
Directed links — how nodes connect. Relationships between objects.
Labels — what each link means (e.g. inside, above). Enables inference.
For a visual analogy: nodes = objects (circle, square, triangle); links = spatial relations (inside, above); labels describe transformations between frames (unchanged, expanded, deleted).
What makes a representation useful?
A is to B, as C is to ? Build semantic networks for A→B and C→D. Pick the D that best matches the A→B transformation.
Unchanged — same object, same relation. Expanded — object grows or multiplies. Deleted — object removed. Use consistently across A→B and C→D.
Classic state-space problem (also: cannibals & missionaries). Semantic nets model states and moves.
The network makes illegal and unproductive moves obvious. From initial state, ~5 moves are possible; only ~2 are both legal and productive. The representation exposes this without search.
When multiple candidates could fit, assign weights to transformations to prefer more specific or better matches.
Sum weights across transformations. Best D = highest total.
Semantic networks mirror theories of human memory—especially spreading activation.
Activating one node spreads to connected nodes. Explains why "John wanted to get rich. He got a gun." leads you to infer robbery—activation from rich and gun merges on paths through the net.