Semantic Networks in AI: Explanation and Worked Example

A semantic network is a knowledge representation formalism where entities and concepts are represented as nodes and their relationships are represented as edges in a graph. This representation is widely discussed in AI knowledge representation literature because it supports modeling what is related to what, and it can support inference by traversing the network and (in some models) propagating relevance via activation.

At a high level, semantic networks let you express facts like “a sparrow is a bird” and connect these to broader categories and attributes such as “birds can fly,” using relationships such as is-a (class membership / hierarchy), has-part (part-of), and related-to (general association). A crucial idea is that inferential links can be derived by moving through the graph (e.g., from sparrow → bird → “can fly”), often with assumptions about inheritance and default reasoning depending on the specific model.

Below is a conceptual schematic of how facts are encoded:

To connect this to classic AI terminology: some early “semantic memory” proposals used network structures with nodes and directed links, aiming to model meaning in terms of how concepts are related, and later work introduced ideas like spreading activation where activating one concept can increase the likelihood of retrieving related concepts.

Limitation note: Semantic-network inference quality depends on how you model relationships (direction, type, and constraints) and how you handle defaults vs exceptions (e.g., “birds can fly” vs “penguins can’t fly”).

Core components of a semantic network

A semantic network typically contains:

• Concept node: the meaning-bearing units.
• Relationship edge: e.g., is-a, part-of, can-do.
• Hierarchy: enables inheritance-style reasoning.
• Activation / relevance: often modeled as spreading activation.

Even if your course focuses on “representation” rather than “neural-style activation,” the graph view is fundamental: semantics come from structure (connectivity) and labels on edges.

Worked example: “Birds that fly” (including an exception)

1) Knowledge you want to represent

Suppose you want to represent:

1. Sparrow is a Bird
2. Penguin is a Bird
3. Birds can Fly
4. Penguins cannot Fly

This is a classic example because it demonstrates both:

• inheritance from a superclass (Bird → Fly), and
• overriding by a subclass (Penguin → not Fly).

2) Semantic network diagram

3) Query: “Can Sparrow fly?”

Reasoning path:

• Start at Sparrow
• Follow is-a to Bird
• Follow can to Fly
• Conclude: Sparrow can fly (given no exception edge for Sparrow)

4) Query: “Can Penguin fly?”

Reasoning path:

• Start at Penguin
• Follow is-a to Bird → suggests “can fly”
• Then note the exception edge Penguin → cannot → Fly
• Conclude: Penguin cannot fly

Pro Tip: In semantic networks, exceptions are typically represented by explicit edges at the more specific node (e.g., Penguin) so they can override inherited defaults from higher-level nodes (e.g., Bird).