Word Sense Disambiguation (WSD)
- Introduced by
- Navigli (survey, 2009)
- Year
- 2009
- Type
- NLP semantic-disambiguation task
What does it do?
Word sense disambiguation (WSD) is the natural-language-processing task of choosing the correct meaning of a polysemous word from its context.
Surveyed by Navigli (2009), it resolves which sense of a many-meaning word applies in a given sentence, improving the quality of information retrieval, machine translation, and question answering.
The intuition behind the method
Many everyday words carry more than one meaning — "bank" can be a riverside or a financial institution, and only the surrounding words tell us which. A reader resolves this automatically, but a machine sees the same string of characters either way. WSD makes that choice explicit: it looks at the words around an ambiguous term, compares them against the dictionary definitions of each candidate sense, and selects the sense whose meaning best fits the context.
History
Word sense disambiguation is a long-standing problem in computational linguistics. Navigli (2009) consolidated the field in a comprehensive survey, while Mihalcea (2006) set out the knowledge-based methods that draw on lexical resources such as WordNet rather than labelled training data.
When to use it
WSD applies when you have text containing words with multiple meanings and a lexical knowledge resource — WordNet or a target-language ontology — that lists their senses, together with a defined context window. It suits explanatory and classification work on text data, and a modest corpus (around 20 units) is enough to begin. Without an available sense inventory for the language, knowledge-based WSD cannot run.
Limitations and considerations
- Requires an available WordNet or target-language ontology; languages without a sense inventory cannot be disambiguated this way.
- Accuracy depends on how well the defined context window captures the disambiguating evidence.
- Fine-grained sense distinctions in a resource can be hard to separate even with rich context.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Resolves polysemy so downstream tasks read the intended meaning rather than a surface string.
- Improves the quality of information retrieval, machine translation, and question answering.
- Knowledge-based approaches work directly from a lexical resource without needing large labelled training sets.
Applications
- Improving information-retrieval relevance by matching the intended sense of a query term.
- Raising machine-translation quality by selecting the correct sense before translating.
- Supporting question-answering systems that must interpret ambiguous terms correctly.
- Semantic analysis of corpora across the humanities, social sciences, and other text-heavy fields.
How it works
How it works
The text is scanned for polysemous words — terms that have more than one possible meaning in the knowledge resource. Each such word becomes a target for disambiguation, since unambiguous words need no further analysis.
A context window is defined around each target word — the neighbouring words and phrases that signal which meaning is intended. The size and shape of this window is a defined parameter of the method, because the surrounding evidence is what drives the decision.
Candidate senses for the target word are retrieved from a lexical knowledge resource such as WordNet or a target-language ontology. Each candidate carries a gloss or definition that describes that particular meaning.
The sense whose definition best overlaps with the context is chosen. Knowledge-based approaches such as the Lesk algorithm compare the words in the context window against each sense's gloss and select the best match, assigning a single sense label to the word in context.
Common pitfalls
- Applying a resource built for one language to text in another — the sense inventory must match the corpus language.
- Using a context window that is too narrow to carry the disambiguating signal, or so wide it pulls in irrelevant words.
- Assuming every word is ambiguous; restrict disambiguation to genuinely polysemous targets to avoid wasted effort and noise.
Frequently asked questions
What knowledge resource does WSD need?
Knowledge-based WSD relies on a lexical resource that lists the senses of words — WordNet for English, or an equivalent ontology for the target language. The resource supplies the candidate senses and their glosses that the method compares against the context.
How does the context window matter?
The context window is the set of neighbouring words used to decide which sense applies. It is a defined parameter: too narrow and it misses the disambiguating signal, too wide and it adds irrelevant words. Choosing an appropriate window is central to the method.
What is the Lesk algorithm?
Lesk is a classic knowledge-based approach that compares the words in the context window against the dictionary gloss of each candidate sense and picks the sense with the greatest overlap. It works directly from the lexical resource without labelled training data.
Does WSD work for languages other than English?
Yes, provided a sense inventory or ontology exists for that language. Applying an English resource such as WordNet to text in another language will not produce correct senses; the resource must match the corpus language.
Related methods
Sources
Navigli, R. (2009). Word Sense Disambiguation: A Survey. ACM Computing Surveys (CSUR), 41(2), Article 10, 1-69.
Open DOI referenceCopied.Mihalcea, R. (2006). Knowledge-Based Methods for Word Sense Disambiguation. Computational Linguistics.
Open publication referenceCopied.