Lexical semantics is the study of word meaning and the systematic relationships between word meanings. It addresses questions such as: What does it mean to know the meaning of a word? How are word meanings related to each other through synonymy, antonymy, hyponymy, and meronymy? How do words acquire multiple senses (polysemy), and how do speakers select the appropriate sense in context? In computational linguistics, lexical semantics informs the design of lexical resources, word sense disambiguation systems, and the lexical entries that drive compositional semantic analysis.
Semantic Relations
Antonymy: hot ↔ cold (opposite meaning)
Hyponymy: dog ⊂ animal (IS-A relation)
Meronymy: wheel ⊏ car (PART-OF relation)
Polysemy: bank → {financial institution, river bank}
Homonymy: bat → {animal, sports equipment}
Lexical relations structure the mental lexicon and are formalized in computational resources like WordNet. Hyponymy (IS-A) organizes words into taxonomic hierarchies: "dog" is a hyponym of "animal." Meronymy (PART-OF) captures partonomic structure: "wheel" is a meronym of "car." Synonymy groups words with similar meanings into synsets (synonym sets). These relations support inferences: knowing that "dogs are animals" licenses the inference from "every dog barked" to "some animal barked."
Polysemy and Sense Enumeration
Most content words are polysemous, having multiple related senses. The word "run" has dozens of senses in English dictionaries, from physical movement to operating a machine to a score in cricket. The sense enumeration approach, used in WordNet and traditional dictionaries, lists each discrete sense. However, the boundaries between senses are often fuzzy, and the number of senses depends on the granularity of the inventory. Alternative approaches include prototype theory, which represents word meaning as a central exemplar with family resemblances, and Pustejovsky's Generative Lexicon, which derives senses compositionally from a structured lexical representation.
Distributional semantics offers an empirical complement to classical lexical semantics. Word vectors capture graded similarity (not just discrete relations), and methods like clustering can automatically discover sense inventories from usage patterns. Multi-prototype models (Reisinger and Mooney, 2010) learn multiple vectors per word, one per sense. Contextualized embeddings from BERT can be clustered to induce word senses that correlate with human judgments. These data-driven approaches are beginning to bridge the gap between handcrafted lexical resources and corpus-based models.
Computational Applications
Lexical semantic knowledge is essential for numerous NLP tasks. Word sense disambiguation (WSD) systems use lexical resources and context to determine which sense of an ambiguous word is intended. Semantic role labeling identifies the thematic roles (agent, patient, instrument) that words play in a sentence, drawing on the argument structure specified in lexical entries. Textual entailment systems leverage hyponymy and synonymy relations to determine whether one sentence logically follows from another.
The interaction between lexical and compositional semantics remains a central concern. Pustejovsky's Generative Lexicon proposes that word meanings include rich type structures (qualia roles) that support compositional operations like type coercion and co-composition. FrameNet represents word meanings in terms of semantic frames — schematic representations of situations — that specify the participants, props, and relations associated with each word. These resources provide the fine-grained lexical knowledge needed for deep semantic analysis.