Source-linked AI summary
"I don't believe in word senses"
Adam Kilgarriff
TL;DR
WSD commonly takes dictionary senses as its targets, while theoretical work questions whether word sense is a workable basic unit. The paper treats contextual corpus citations as basic units and derives senses by clustering them; it concludes that word sense itself is not sufficiently well-defined as a basic unit of meaning.
Problem
WSD targets dictionary-like senses, but theoretical accounts question whether word sense is sufficiently well-defined to serve as a basic unit.
Source: §2.2 · p. 5, §7 · p. 20Method
The paper treats occurrences of a word in context, operationalized as corpus citations, as basic units from which sufficiently distinct clusters can form senses.
Source: §7 · p. 20Results
The paper concludes that word sense was not sufficiently well-defined to be a workable basic unit of meaning.
Source: §7 · p. 20Takeaways & Limitations
In this account, corpus citations are basic units, while senses arise from distinct clusters and many uses may be aberrant.
Source: §7 · p. 20The proportion of anomalous uses remains an open question, and the paper expects it to vary substantially by corpus.
Source: §5.2 · p. 17
Abstract
from arXiv · showhide
Word sense disambiguation assumes word senses. Within the lexicography and linguistics literature, they are known to be very slippery entities. The paper looks at problems with existing accounts of `word sense' and describes the various kinds of ways in which a word's meaning can deviate from its core meaning. An analysis is presented in which word senses are abstractions from clusters of corpus citations, in accordance with current lexicographic practice. The corpus citations, not the word senses, are the basic objects in the ontology. The corpus citations will be clustered into senses according to the purposes of whoever or whatever does the clustering. In the absence of such purposes, word senses do not exist. Word sense disambiguation also needs a set of word senses to disambiguate between. In most recent work, the set has been taken from a general-purpose lexical resource, with the assumption that the lexical resource describes the word senses of English/French/..., between which NLP applications will need to disambiguate. The implication of the paper is, by contrast, that word senses exist only relative to a task.
1 Introduction
The introduction questions whether dictionary-style senses provide a principled basis for WSD and presents an alternative in which senses are clusters of corpus citations whose relevance depends on purpose.
1 Introduction
- WSD is generally framed as choosing among the senses listed in a dictionary or similar resource.The introduction compares this computational task with selecting an appropriate meaning when understanding an ambiguous word.Source: §1 · p. 1
- Philosophical accounts have not connected dictionary sense distinctions to naturally occurring language data, a gap this paper aims to address.The paper also notes that these accounts lack a taxonomy of the processes behind intuitions about multiple meanings.Source: §1 · p. 2
- The paper argues that Cottrell’s construal conflicts with lexicon theory and that attempts to ground the concept of word sense have been unsuccessful.It then examines lexicographers’ decisions and develops an alternative account based on clusters of citations exhibiting similar patterning and meaning.Source: §1 · p. 2
- Different corpora and purposes can yield different senses, so dictionary and thesaurus inventories need not suit NLP applications.The paper relates a word’s meaning potential to dictionary senses using corpus evidence.Source: §1 · p. 2
2 Thesis and antithesis: practical WSD and theoretical lexicology
Practical WSD needs a target inventory, often supplied by dictionaries or user-defined seed senses, while theoretical lexicology questions whether a stable set of word senses exists. The section sets these approaches against each other and asks what a word sense is.
2.1 Thesis
- WSD systems moved from limited, researcher-selected lexicons toward dictionaries as relatively principled sources of wide-coverage lexical information.Dictionary machine-readable versions enabled experimental WSD programs based on dictionary sense inventories.Source: §2.1 · p. 3
- Corpus-based approaches either cluster contextual vectors into senses or let users define senses by classifying seed collocates.Schütze and Pederson cluster context vectors; Clear and Yarowsky use seed collocates to establish the distinctions for disambiguation.Source: §2.1 · p. 3
2.2 Antithesis
- Cognitive linguistics treats metaphor and polysemy as central to how words develop additional meanings and uses.The section connects these processes to the structures underlying distinct word meanings.Source: §2.2 · p. 4
- Cruse (1995) distinguishes polysemy from polylexy and accounts for lexical relations, antagonism, facets, and semi-distinct readings.In this account, predictable non-base meanings need not be stored as separate lexical entries.Source: §2.2 · pp. 4–5
- Theoretical work on lexical ambiguity considers processes that generate additional meanings or uses and systematic relations among them.The section mentions generative lexicon processes, regular polysemy, and verb-class patterns of subcategorisation and alternation.Source: §2.2 · p. 5
- WSD treats ambiguity as a practical risk of choosing the wrong lexical entry, while theoretical accounts question whether a dictionary-like inventory of senses exists.The paper presents this contrast as an impasse requiring closer examination of what a word sense is.Source: §2.1 · p. 3, §2.2 · p. 5
3 What is a word sense?
The paper questions whether word senses can be identified by stable operational criteria: contexts may select or modulate meanings, while tests for ambiguity have significant limitations. Priming offers evidence that polysemous meanings are not functionally independent, but the evidence does not yield a simple, stable sense distinction.
3.1 Selection and modulation
- Word-sense identity needs operational criteria, but distinguishing a separate sense from contextual modulation is not straightforward.Cruse distinguishes contexts that select a distinct sense from those that modulate a word’s meaning.Source: §3 · p. 6, §3.1 · p. 6
3.2 Ambiguity tests
- Crossed-readings tests can probe ambiguity, but intermediate cases require crossed readings to be prototypical.The possibility of intermediate readings does not itself rule out ambiguity, though the added constraint is difficult to make precise.Source: §3.2 · p. 7
3.2.1 Criticisms of the tests
- Cross-reading tests fare tolerably well on consistency, but there is little evidence they resolve cases where intuitions are unclear.The paper also notes that test-sentence construction can affect outcomes, although the crossed-reading test is least affected by this problem.Source: §3.2.1 · p. 8
- Acceptability judgments cannot reliably isolate lexical ambiguity from interpretation, and informants’ judgments may be unstable or conflicting.The tests also require syntactically similar uses, limiting their application when a word’s uses differ syntactically.Source: §3.2.1 · pp. 8–9
3.3 Psycholinguistics and ‘semantic priming’
- In ambiguous-word priming, both bank readings are initially primed, but context leaves only the appropriate target primed after 50–200 ms.Williams’s results also suggest that polysemous meanings are not functionally independent and are unequal in mental-lexicon representation.Source: §3.3 · p. 10
- Williams found asymmetric priming: a central-reading context did not prime non-central targets, while a non-central-reading context primed central targets.The experiment used firm, with solid and strict as its readings.Source: §3.3 · p. 10
- Priming may help ground a distinction between ambiguity and generality, but further work is needed and it cannot practically determine sense counts for many words.The results are not sufficiently stable; Williams reports substantial item and subject variability.Source: §3.3 · p. 10
4 Lexicographers, dictionaries, and authority
Dictionary sense divisions reflect the purposes and constraints of lexicography, not necessarily discrete, non-overlapping meanings. The paper presents corpus-based clustering as the basis from which dictionary senses are abstracted, with the resulting distinctions depending on the clustering purpose.
4 Lexicographers, dictionaries, and authority
- Lexicographers’ lump-or-split decisions are subjective, and dictionary sense inventories are shaped by their target audience and editorial philosophy.The alternative decision may be equally valid, and dictionaries encode relations in the grey area between same and different senses.Source: §4 · p. 11
- Dictionaries are pressed to present clear, discrete meanings for settling disputes, but this format does not show that words actually have discrete senses.Lexicographers may be obliged to divide meanings without believing that the divisions reflect discrete, non-overlapping senses.Source: §4 · pp. 11–12
4.1 Lexicographical literature
- Ayto’s genus-and-differentiae approach leaves open questions, while Stock criticizes its circularity and limited applicability.Their discussion illustrates difficulties in giving a general account of how dictionary senses should be divided.Source: §4.1 · p. 12
- Corpus lexicography shifted the lexicographer’s primary evidence from introspection to word occurrences shown in context.Large computerized corpora made it possible to examine many ordinary and extraordinary instances together.Source: §4.1 · p. 13
4.2 A corpus-based model of word senses
- The corpus-based procedure clusters concordance lines by usage patterns, identifies what members share, and then expresses the conclusions as dictionary definitions.Clustering is data-driven; the dictionary’s senses are several steps removed from this basic process.Source: §4.2 · p. 13
- Under this account, usages belong to the same dictionary sense when the lexicographer would place them in the same cluster.The account also allows less distinct clusters to be described as vague or unspecified rather than forcing a clear ambiguity distinction.Source: §4.2 · pp. 13–14
5 Use, frequency, predictability, and the word sense
The section treats word senses as patterns abstracted from corpus use, with dictionary inclusion shaped by how frequent and predictable those patterns are. Its handbag examples show how varied non-standard uses draw on linguistic, collocational, and world knowledge.
5 Use, frequency, predictability, and the word sense
- Dictionary senses depend on whether a usage pattern is frequent enough and not sufficiently predictable from other uses; dictionaries apply different thresholds.A first-time reader’s ability to understand a use, including its connotations, is the paper’s measure of predictability.Source: §5 · p. 14
- Handbag uses draw on interacting knowledge sources, including metonymy, world knowledge, and lexical associations beyond the word’s denotation.For example, “handbags at ten paces” relies on the lexical association of handbags with weapons, not just their denotation.Source: §5 · p. 16
- “Dance round your handbag” licenses the music-genre use, but the collocation is more widely known than the genre label, which was a 1990–91 fashion.The collocation also appears in speech and in written music journalism; the genre use treats “handbag” as a music-genre noun.Source: §5 · pp. 15–17
5.1 Frequency
- Around 97% of BNC handbag citations use the base sense, helping explain why dictionaries omit the rare weapon and music-genre senses and the twice-attested collocation.The paper also cites limited space in single-volume dictionaries as a reason to omit partially compositional collocations.Source: §5.1 · p. 17
5.2 Analytic definitions and entailments
- A strictly denotational dictionary definition can omit handbag’s weapon association because the base and weapon-use denotations are the same.The paper notes that a more pragmatic lexicographer might still include the association, particularly when space permits.Source: §5.2 · p. 18
- Rare citations with unspecified “sanction” do not justify a dictionary sense for the abstract meaning “control”; some dictionaries instead use nested entries for related meanings.A “lowest common denominator” reading can cover uses that fall into distinct corpus clusters.Source: §5.2 · pp. 18–19
6 Implications for WSD
The paper argues that word senses are undefined without a rationale for distinguishing them, so no task-independent set is coherent. For NLP, senses should reflect the task and corpus, and unhandled uses can be added as lexical entries.
6 Implications for WSD
- Word senses require a rationale for clustering, so a task-independent set of senses for a language is not a coherent concept.For people, the fish-versus-weapon distinction for “pike” is worth making in almost any situation where they attend to a sentence containing the word.Source: §6 · p. 19
- The handbag data suggest that generative principles may add little coverage because non-standard uses often have word-specific histories and connections.One non-standard use may build on another, with connections specific to a word or lexical field.Source: §6 · pp. 19–20
- The relevant senses depend on the corpus: a general-English corpus may support a general-purpose dictionary, while music journalism makes the handbag music-genre sense prominent.The paper suggests a 1990s music-journalism dictionary would include that sense.Source: §6 · p. 20
- For near-term NLP coverage of a misinterpreted use, the paper recommends adding a lexical entry rather than computing a new meaning.Examples include adding verbal “handbag” as a synonym for “beat” when sufficiently frequent.Source: §6 · p. 20
7 Conclusion
The conclusion rejects word senses as basic units of meaning, treating them instead as abstractions over contextual corpus citations. For WSD, sense inventories are purpose-relative, so dictionary distinctions may not suit an NLP application.
7 Conclusion
- The concept of word sense is not sufficiently well-defined to serve as a basic unit of meaning.Source: §7 · p. 20
- Corpus occurrences in context, operationalized as citations, are the basic units; sufficiently large, distinct clusters of citations can form word senses.The paper argues that word senses belong in scientific vocabulary only as abstractions over clusters of usages, not as objects in the ontology.Source: §7 · p. 20
- Many uses do not fit the simplest clustering case: even handbag has aberrant citations, and word use can involve lexical creativity.Source: §7 · p. 20
- For WSD, senses are defined relative to interests, and a dictionary’s set may not match the set relevant to an NLP application.Source: §7 · p. 20