WordNet: An On-line Lexical Database

An on-line lexical reference system whose design is inspired by psycholinguistic theories of human lexical memory: English nouns, verbs, adjectives and adverbs are organized into synonym sets (synsets), each representing one lexicalized concept, linked by semantic and lexical relations.

Miller and Johnson-Laird’s name for psycholinguistic research on the lexicon. Standard dictionaries like the OED focus on historical (diachronic) evidence and neglect the synchronic organization of lexical knowledge; WordNet instantiates psycholinguistic hypotheses about it, and exposes them to the full common vocabulary.

Putting alphabetical dictionaries on line uses computers merely to look things up faster. WordNet organizes the lexicon by meaning rather than form, with the bookkeeping (like two-way look-ups in a thesaurus) done by the computer.

Nouns, verbs, adjectives and adverbs. Function words are assumed to be stored separately as part of the syntactic component of language, supported by observations of aphasic speech.

Word association studies (Fillenbaum and Jones 1965): probes tended to elicit responses of the same category (nouns most strongly, then adjectives, then verbs), which can’t be association by contiguity since adjacent words in text rarely share a category.

single organizing principle for all categories would misrepresent the psychological complexity of lexical knowledge.

A table with word forms as columns and word meanings as rows; an entry means the form can express that meaning (in context). Two entries in one column: the form is polysemous. Two entries in one row: the forms are synonyms.

	F1 board	F2 plank	F3 committee
M1 piece of lumber	✓	✓
M2 group of people	✓		✓

Both arise from the many-to-many mapping between forms and meanings: a listener recognizing a form must cope with its polysemy; a speaker expressing a meaning must choose between synonyms.

A constructive theory represents enough to construct the concept from scratch (standard dictionary definitions often fail this). A differential theory needs only symbols that distinguish meanings. WordNet is differential: a synset like {board, plank} or {board, committee} identifies a concept for someone who already knows it.

A short definition used when no suitable synonym exists, e.g. {board, (a person’s meals, provided regularly for money)}. Like a synset it only needs to let an English speaker tell this sense from others, not teach the concept.

Relative to context: two expressions are synonymous in a context C if substituting one for the other there doesn’t change truth value (Leibniz’s context-free version makes true synonyms almost nonexistent). Plank for board rarely changes truth in carpentry contexts. In practice it is graded semantic similarity.

If synonyms must be interchangeable, words of different syntactic categories can’t be synonyms (can’t share a synset), consistent with evidence that nouns, verbs and modifiers are organized independently in memory.

{rise, ascend} and {fall, descend} are conceptual opposites, but the antonym pairs are rise/fall and ascend/descend; people hesitate over rise/descend or ascend/fall. Antonymy also isn’t simply “not-x”: not rich doesn’t imply poor.

A semantic relation between meanings: {x} is a hyponym of {y} if speakers accept “An x is a (kind of) y” (maple → tree → plant). It is transitive and asymmetric, and with usually one superordinate it forms a hierarchy (an inheritance system); it is the central organizing principle for nouns.

The part–whole (HAS-A) relation between meanings: {x} is a meronym of {y} if speakers accept “A y has an x (as a part)” or “An x is a part of y”. It is asymmetric and (with qualifications) transitive; a meronym can have many holonyms.

A user clicking on trees shouldn’t be told it isn’t in the database, so the interface strips inflections (plural, verb forms) and looks up the base form; derivational morphology was left for the future.

A hierarchy in which each word inherits the distinguishing features of all its superordinates: if Rex is a collie, Rex inherits the properties of collie, which inherits those of dog, canine, and so on. WordNet nouns are organized this way. E.g. canary → finch → passerine → bird → vertebrate → animal.

Attributes (modification, e.g. a canary is small, yellow), parts (meronymy, e.g. has a beak, wings) and functions (predication, e.g. sings). Only meronymy was implemented in the noun files at the time.

The topmost nodes of the noun hierarchies: instead of one hierarchy under an empty “entity”, nouns are partitioned into 25 semantic fields (act, animal, artifact, food, person, plant, …), each heading a separate hierarchy, which also helps lexicographers work in parallel.

Seldom more than about a dozen levels; the deepest (technical) chains run about eleven levels, artifacts six or seven, persons three or four.

A level somewhere in the middle where most distinguishing features are attached: people can list many features for basic-level categories, few for terms above it, and only a few extra ones below it.

Component parts (branch/tree), members (tree/forest) and substances (what something is made of), out of six types distinguished by Winston et al. The component relation is by far the most frequent.

Parts must distinguish an object from others: atoms are part of every concrete object, so as meronyms they distinguish nothing. Some parts, like point, are shared by many different things (arrow, knife, needle, pencil…).

Descriptive adjectives (big, heavy, interesting), which ascribe a value of an attribute to a noun (heavy: WEIGHT(package) = heavy), and relational adjectives (presidential, nuclear), which relate to a noun and are cross-referenced to the noun files.

Antonymy: most attributes are bipolar, and antonymous adjectives express opposing values of an attribute (heavy/light for WEIGHT). WordNet also links descriptive adjectives to the noun lexicalizing their attribute.

Direct antonyms are lexically paired opposites (wet/dry); indirect antonyms are conceptual opposites reached through similarity (moist → wet → dry). Adjectives form bipolar clusters: two focal antonyms, each surrounded by similar adjectives (satellites).

Direct antonymy is a lexical association between particular words, typically frequent (mostly Anglo-Saxon) adjectives learned together; less common near-synonyms (ponderous, weightless) get their opposition indirectly through the focal pair.

With a special cluster headed by angry/not angry, since such attributes have degrees (from no anger to fury) but no lexicalized opposite pole.

Verbs are more polysemous and their meanings more flexible: they shift with the nouns they combine with (have a Mercedes vs have a headache), and the most frequent verbs (have, be, run, make, take) are the most polysemous.

Into 15 files, mostly semantic domains (verbs of bodily care and functions, change, cognition, communication, competition, consumption, contact, creation, emotion, motion, perception, possession, social interaction, weather), plus one file of stative verbs.

The verb counterpart of hyponymy (Fellbaum and Miller 1990): “To V1 is to V2 in some particular manner”, e.g. limp/walk, lisp/talk, mumble/talk. “Manner” is loose: speed (stroll), force (chop, slam, tap), medium (fax, phone), intention (confess, preach).

It needs nouns (“to amble is a kind of to walk” is infelicitous), and even with gerunds people reject “Ambling is walking” without qualification; the difference between verbs lies in manner elaborations, hence troponymy.

Every troponym entails its superordinate (limping entails walking), and the two activities are temporally co-extensive: one walks every instant one limps. Pairs related by entailment with proper temporal inclusion (snore/sleep, buy/pay) are not troponyms.

The entailed activity always precedes the entailing one: hit and miss both entail aim, succeed and fail entail try, win and lose entail play. There’s no temporal inclusion, so no part–whole statement is possible.

It links lexicalized causative–resultative pairs (give/have, show/see, teach/learn, fell/fall, raise/rise); the subject of the resultative is the object of the causative. Periphrastic causatives (cause to die) aren’t treated as synonyms of lexical ones (kill), since they express more indirect causation.

Flatly, like adjectives: mostly by synonymy and opposition (live/die, wake/sleep, include/exclude), with almost no superordinates or troponyms.

Lexicographers write source files, one per part of speech and semantic field, listing synsets with relational pointers and glosses; the Grinder program checks their syntax, resolves the pointers and compiles them into the database used by the access tools.