Word sense disambiguation (WSD) is the task of determining which meaning of a polysemous word is intended in a given context, by assigning each occurrence a sense from a predefined inventory such as a dictionary or lexical database. In Speech & language AI, it is one of the oldest problems of natural language understanding, because words such as “bank”, “bass” or “interest” carry several unrelated or loosely related meanings that only the surrounding text can separate.
WSD rarely appears as a standalone product. It works as a hidden capability inside larger systems: machine translation, search, information extraction and question answering all fail in predictable ways when they choose the wrong sense. Modern language models resolve much ambiguity implicitly, yet explicit WSD remains relevant wherever an auditable, inventory-based sense label is required.
How word sense disambiguation selects a meaning
A WSD system combines three ingredients: a target word, the context in which it occurs, and a sense inventory that lists the candidate meanings. The system scores each candidate sense against the context and returns the best match, or a probability distribution over senses when downstream components can use uncertainty.
Context evidence comes in several forms. Local collocations (the words immediately next to the target), topical cues from the wider paragraph, syntactic relations such as the verb that governs a noun, and selectional preferences (for example, “drink” prefers liquid objects) all point toward a sense. The classic heuristic of one sense per discourse observes that a word repeated within one document usually keeps the same meaning, while one sense per collocation observes that a fixed neighboring word strongly predicts the sense.
The following table shows how different evidence types resolve the noun “bass” in practice.
| Evidence type | Example context | Selected sense |
|---|---|---|
| Local collocation | “bass guitar” | Low-pitched instrument |
| Topical cue | Paragraph about rivers and bait | Freshwater fish |
| Syntactic relation | “caught a bass” | Fish (object of catch) |
| Selectional preference | “the bass was too loud” | Low-frequency sound |
Sense inventories and granularity
The choice of sense inventory determines what counts as a correct answer. The most widely used inventory is WordNet, which groups synonyms into synsets and distinguishes many senses per word; the noun “line” alone has more than two dozen. Such fine-grained inventories make the task harder, because human annotators themselves often disagree on closely related senses.
Inter-annotator agreement on fine-grained senses typically falls between about 70 and 80 percent, which sets a practical ceiling for automatic systems. Coarser inventories, which merge related senses into broader groups, raise agreement above 90 percent and are often more useful for applications such as translation, where only distinctions that change the output matter.
The table below compares common inventory styles and their typical uses.
| Inventory style | Granularity | Typical use |
|---|---|---|
| WordNet synsets | Very fine | Research benchmarks, lexical studies |
| Coarse sense clusters | Medium | Translation, search, reliable evaluation |
| Multilingual networks | Fine, cross-lingual | Multilingual WSD and entity linking |
| Domain glossaries | Narrow, task specific | Clinical, legal or technical text |
Approaches from knowledge bases to neural models
WSD methods fall into four families, which differ in the data they require and in how they generalize to rare senses. Knowledge-based methods such as the Lesk algorithm compare the context with dictionary glosses or traverse a semantic graph; they need no labeled data but reach modest accuracy.
Supervised methods train classifiers on sense-annotated corpora such as SemCor. Before deep learning, a separate classifier per word was common; current systems use contextual embeddings from transformer encoders and compare them with sense embeddings or gloss encodings. Gloss-aware neural models read the definition text itself, which lets them handle senses that are rare or absent in training data. Large language models can perform WSD through prompting, either by choosing among listed definitions or by paraphrasing the intended meaning.
| Approach | Data required | Strength | Weakness |
|---|---|---|---|
| Knowledge-based | Dictionary or semantic graph | No annotation needed | Lower accuracy |
| Supervised feature-based | Sense-annotated corpus | Strong on frequent words | Poor on unseen words |
| Neural gloss-aware | Annotated corpus plus glosses | Handles rare senses better | Depends on gloss quality |
| Prompted language model | Instructions and sense list | Flexible, few examples | Cost, less predictable labels |
On standard English all-words benchmarks, a simple most frequent sense baseline reaches about 65 percent F1, and modern neural systems report results around 80 percent F1 or slightly higher, close to the agreement ceiling of human annotators. Results drop considerably for low-resource languages and specialized domains.
Relation to neighboring language tasks
WSD is closely related to several other tasks in the field, but each resolves a different kind of ambiguity. Entity linking maps mentions of named entities to entries in a knowledge base, so it disambiguates names such as “Paris” rather than common words. Part-of-speech tagging separates grammatical categories (“book” as noun or verb), which is often a prerequisite step, because senses are usually defined per part of speech.
- Named entity recognition finds and types names but does not choose a specific referent.
- Coreference resolution links expressions that refer to the same entity, not to a sense.
- Lemmatization reduces inflected forms to a dictionary form, which WSD needs before lookup.
- Semantic textual similarity compares whole sentences and depends indirectly on correct senses.
Limitations and common mistakes
The most serious limitation of WSD is that sense boundaries are not natural facts; they are lexicographic decisions. A system can be penalized for an answer that a human reader would accept, which makes benchmark scores hard to translate into application value.
| Symptom | Likely cause | Remedy |
|---|---|---|
| Always predicts the dominant sense | Skewed training distribution | Gloss-aware models, rebalanced data |
| Fails in medical or legal text | Domain shift in senses | Domain glossary and adaptation |
| Scores well, helps little downstream | Inventory too fine-grained | Coarser sense clusters |
| Inconsistent labels in one document | No discourse-level constraint | One sense per discourse heuristic |
A frequent misconception is that large language models have solved WSD. They handle common ambiguity well, but they still struggle with rare senses, metaphorical uses and domain-specific meanings, and they do not output a stable inventory label unless they are constrained to a fixed list.
Stripped to the essentials
Many words have more than one meaning. A “bank” can be the side of a river or a place that holds money. People rarely notice this, because the rest of the sentence tells them which meaning fits.
Word sense disambiguation is teaching a computer to make that same choice. The software looks at the nearby words, compares them with the possible meanings listed in a dictionary, and picks the one that fits best. When it chooses well, translations, search results and automatic summaries make sense; when it chooses badly, they can become confusing or wrong.
Where it is applied
A pharmaceutical company runs a system that scans thousands of clinical reports each week to find adverse drug reactions. The word “cold” causes trouble: it may describe a viral infection, a feeling of low temperature, or the phrase “cold extremities”, which can signal a circulation problem. Early versions of the pipeline counted every mention as an infection and flooded reviewers with false alarms.
The team added a domain-adapted WSD component that uses a medical glossary as its sense inventory and a transformer encoder fine-tuned on a few thousand annotated sentences. It labels each occurrence with a sense and a confidence score, and low-confidence cases go to a human reviewer. After deployment, the share of irrelevant alerts fell substantially, and reviewers spent their time on genuine safety signals instead of filtering out ordinary uses of the word.