Textbook of AI

A semantic network is a knowledge representation in which concepts are nodes and relations are labelled, directed edges between them. The classic example is the chain canary --is-a--> bird --is-a--> animal, with property edges such as bird --can--> fly and canary --colour--> yellow. Inheritance is handled by traversal: to answer "can a canary fly?" the system follows the is-a link to bird and finds the can-fly property. The framework was introduced by Ross Quillian in his 1968 paper Semantic Memory, a chapter in Marvin Minsky's edited volume Semantic Information Processing, as a computational model of human associative memory.

Spreading activation

Reasoning in a semantic network proceeds by spreading activation from a query node along edges to retrieve related concepts. Activation decays with graph distance (typically multiplicatively, $a_v \leftarrow a_v + \lambda \cdot a_u$ for each edge $u \to v$, with $\lambda < 1$). Cumulative activation at a target node measures the strength of association, and intersecting activation fronts from multiple query nodes find shared meaning. Quillian's original program could answer questions like "what do cry and comfort have in common?" by spreading activation from each word and looking for an intersection node, in this case, sad. The spreading-activation idea is a clear ancestor of message passing in graph neural networks and, more loosely, of attention mechanisms in modern Transformers.

Descendants and influence

Semantic networks fanned out into half a century of representation languages:

• Conceptual dependency theory (Schank, 1972), a small inventory of primitive acts (PTRANS, ATRANS, MTRANS, INGEST...) into which all sentences are decomposed.
• Frames (Minsky, 1975), bundles of slots-and-fillers that capture stereotyped situations (a restaurant frame has slots for waiter, menu, bill).
• KL-ONE (Brachman, 1979) and the description logics that descend from it, a formal core that became the basis of OWL and the Semantic Web.
• Conceptual graphs (Sowa, 1984), bipartite graphs of concept and relation nodes, with a fully worked-out logical semantics.
• WordNet (Miller, 1985–), a hand-built lexical semantic network of English with around 117,000 synsets connected by hypernym, hyponym, meronym and antonym links.
• Modern knowledge graphs, Google's Knowledge Graph (2012), Wikidata, YAGO, DBpedia, Freebase. Wikidata alone has over 100 million items and 1.4 billion statements.

In cognitive psychology, Collins and Loftus's spreading-activation model (1975) and the propositional networks of Anderson and Bower (1973, in their HAM and ACT theories of memory) are direct heirs.

Modern relevance

Knowledge graphs underpin search engines, virtual assistants and biomedical reasoning systems. Within deep learning, knowledge-graph embeddings (TransE, ComplEx, RotatE) map nodes and relations to vectors so that $\mathbf{h} + \mathbf{r} \approx \mathbf{t}$ for true triples $(h, r, t)$. Graph neural networks generalise spreading activation by learning the message and aggregation functions. And the rise of retrieval-augmented generation has revived interest in symbolic knowledge stores, semantic networks, in essence, as a complement to the parametric knowledge of large language models, particularly where verifiability and update are required.