Semantic Network

M. Ross Quillian's 1968 dissertation, "Semantic Memory" (Carnegie Institute of Technology), proposed Teachable Language Comprehender (TLC) — a program that stored definitions as networks of nodes connected by typed pointers. The system handled questions by spreading activation from query nodes outward through the network until paths intersected. Quillian's framework directly inspired Marvin Minsky's frame theory (1974) and indirectly the entire knowledge-representation tradition in AI. Quillian's design also influenced cognitive psychology — Allan Collins (his colleague) carried the spreading-activation model into experimental work.

Allan Collins and Elizabeth Loftus's 1975 paper in Psychological Review extended Quillian's mechanism. When a concept is activated, activation spreads along network links, decaying with distance. Closer concepts are primed more strongly. In a lexical decision task, target words preceded by semantically related primes ("doctor" → "nurse") are recognized 50-100 milliseconds faster than unrelated controls. The network distance predicts the priming gradient. James McClelland and David Rumelhart's connectionist models (1981 onward) reformulated the same insights in distributed neural-network terms.

WordNet, started by George Miller at Princeton in 1985, organizes English vocabulary into synsets — groups of synonyms ({car, auto, automobile, motorcar}). Synsets are linked by IS-A relations forming hierarchies (car → motor_vehicle → vehicle → conveyance → instrumentality → artifact → entity). Other links: meronymy (HAS-PART: car HAS engine), antonymy (large/small), troponymy for verbs (manner-of: limp is a troponym of walk), and entailment (snore entails sleep). WordNet 3.1 (2011) contains 117,659 synsets and 206,941 word-sense pairs across noun, verb, adjective, and adverb networks. It is open-source and underpins thousands of NLP papers.

Word2vec (Mikolov et al., 2013) and GloVe (Pennington, Socher, Manning, 2014) replaced symbolic links with dense vectors learned from co-occurrence. Words become 100- to 300-dimensional vectors; geometric relationships approximate semantic ones — king − man + woman ≈ queen. Contextualized embeddings (ELMo 2018, BERT 2018, GPT 2018+) produce per-token vectors sensitive to context. Embeddings outperform WordNet on many tasks but lack explicit relations; one cannot ask "what is the IS-A parent of dog" of a vector. Hybrid systems (knowledge-augmented BERT, KnowBert 2019) combine WordNet structure with embedding flexibility.

Many. Meronymy (PART-OF: wheel partOf car). Substance-of (oak substanceOf furniture). Member-of (player memberOf team). Antonymy (hot/cold), with directional, complementary, and gradable subtypes. Synonymy (begin/start), recognizing context dependence. Troponymy (verb manner: amble is a manner of walking). Causation (kill → die). Entailment (divorce entails being-married). Frame-relatedness (FrameNet adds relations across frames). Knowledge graphs (Freebase, Wikidata) extend semantic networks with rich event and entity relations across millions of nodes.

David Meyer and Roger Schvaneveldt's 1971 lexical decision study showed faster responses to "butter" after "bread" than after "doctor." Subsequent work catalogued associative priming (table-chair), categorical priming (apple-pear), and antonym priming (hot-cold). Priming dissipates within 1-2 seconds and is sensitive to depth of processing. Neuroimaging (Federmeier and Kutas, 1999, N400 component) shows priming effects in EEG. The experimental literature largely supports a graded, relational structure of semantic memory, consistent with network models — though distributed embeddings predict the same effects.

WordNet is integrated into many systems for synonym expansion, hypernym lookup, and word sense disambiguation. The classic Lesk algorithm (1986) used WordNet glosses for disambiguation. Knowledge graphs (Google's Knowledge Graph, launched 2012; Wikidata, 2012; Freebase, 2007-2016) extend networks to entities and events at scale. Retrieval-augmented language models (RAG, 2020) and graph neural networks (GNNs) bring relational structure to deep learning. Semantic networks remain useful for explainability and structured reasoning even as embeddings dominate raw performance benchmarks.