Can a language model perform WSD?

Yes, language models can support contextual sense selection, but the task still needs clear labels, evaluation and uncertainty handling.

What happens when the correct sense is missing?

A fixed-choice model may choose an unsuitable label. The workflow should provide a way to flag missing coverage or abstain rather than concealing the mismatch.

AI Glossary / Natural language and speech / Intermediate

Word sense disambiguation

Definition

Definition: Word sense disambiguation, or WSD, is the task of deciding which meaning of a word is intended in a particular context. In its conventional computational form, it selects a label from a specified inventory of possible senses.

What it means

What it means: A message saying “The seal is damaged” could concern a closure, an official mark or an animal, depending on the surrounding discussion. Identifying the word seal does not settle the meaning. WSD connects that occurrence to the interpretation supported by its context.
The output depends on the inventory. A broad animal-versus-object distinction and a detailed dictionary classification are different tasks, even when they examine the same sentence. WSD also differs from discovering new senses: choosing among supplied labels is not the same as constructing those labels from examples. A system may need both capabilities, but they require different evaluation.

How it works

How it works: The system receives a target word and its context, retrieves candidate senses and compares the evidence for each candidate. Methods can use overlaps with dictionary definitions, annotated training examples, semantic relationships or contextual representations produced by a language model.
For example, a model may encode the target occurrence and compare it with representations built from labeled examples of each sense. The closest candidate becomes a prediction, subject to the application's uncertainty policy. Common-sense frequencies can provide a baseline, but always selecting the most frequent meaning ignores the actual sentence. Proper evaluation uses held-out examples and the same inventory version as the system, while separately examining rare meanings and domain changes.

Three practical examples

Three practical examples:
1. A maintenance search engine interprets seal in a discussion of a leaking pump as a component-related meaning. It should still check context rather than treating every maintenance document as unambiguous.
2. A translation workflow distinguishes an animal-related use of bat from a sporting implement. The relevant meaning helps select an appropriate expression in the target language.
3. A research archive separates plant as an organism from plant as an industrial facility. Ambiguous headings are reviewed with their abstracts instead of being classified from the isolated keyword.

Think of it this way

Think of it: WSD resembles choosing the appropriate definition while reading a dictionary beside a passage. The passage supplies evidence; the dictionary supplies available choices. The analogy is limited because an automated model can rely on misleading statistical cues, and neither the inventory nor the passage necessarily supports a unique answer.

Worked example

Worked example: Suppose a small test set contains ten occurrences of plant, with agreed labels O for organism and F for industrial facility. There are six O examples and four F examples. Always predicting O yields six correct answers out of ten, or 60% accuracy.
A contextual system correctly labels five of the six O examples and three of the four F examples. Its accuracy is 8/10, or 80%. Because it misses one example from each class, the O recall is 5/6, approximately 83.3%, and the F recall is 3/4, or 75%.
These invented figures show why comparing against a majority-sense baseline matters. They do not demonstrate general performance: ten examples are too few to represent all uses, and different inventories can change the task. If a new deployment contains mostly facility-related reports, the earlier 60% baseline does not transfer automatically. Keep the test distribution and meaning labels visible when interpreting the result.

Illustrative case study

Illustrative case study: Fictional Harbor Manuals uses WSD to organize technical documents. A prototype performs well on long passages but struggles with headings such as “seal inspection.” Reviewers find that the heading alone often lacks the clues present in the body text.
The team includes surrounding paragraphs, establishes a review option for unresolved cases and reports results separately for headings and full passages. It also checks whether the sense labels improve users' actual search results. This fictional case illustrates evaluation design and context selection. It does not claim that adding a WSD component necessarily improves every search engine or that confidence scores establish correctness.

Common misconceptions and mistakes

Common misconceptions and mistakes: WSD is not just part-of-speech tagging; multiple senses can share the same grammatical category. It is not identical to entity linking, which connects mentions to particular knowledge-base entries or entities. A common sense is not always the correct sense. A model can also select an inventory label confidently when the appropriate domain meaning is absent. Finally, treating every occurrence within a document as having the same meaning is a useful possible heuristic in some settings, not a universal linguistic rule.

Business use

Business use: Apply WSD where a specific meaning distinction changes a useful decision, such as retrieval, translation or routing. Document the inventory, context window and treatment of missing or ambiguous senses. Compare against a simple frequency baseline and evaluate uncommon meanings separately. Measure the downstream result as well as label accuracy: an accurate distinction may still be irrelevant to a search task, while a rare error could send a request to the wrong service. Preserve text evidence for editorial review.

Check your understanding

Check your understanding: A classifier gets 90% accuracy by always selecting the dominant sense in a test set that contains that sense 90% of the time. Has it demonstrated contextual disambiguation?

Answer: No. It has matched the majority baseline without showing that it uses context. Examine minority-sense performance and context-dependent examples.

Related terms

Word sense
Contextual embeddings
Entity linking
Word-sense induction
sense inventory
polysemy
contextual embedding
most frequent sense baseline
WordNet

Learn next

build a small labeled set for one ambiguous word, reserving some examples for evaluation. Compare a majority-label rule with a context-based decision. Inspect disagreements, ambiguous cases and missing meanings before interpreting any single accuracy figure as evidence of reliable understanding.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Appendix I, task definitions, inventories, baselines and contextual methods: https://web.stanford.edu/~jurafsky/slp3/I.pdf . Princeton WordNet, official description of sense-specific lexical representations: https://wordnet.princeton.edu/ . The plant calculations and Harbor Manuals case are original teaching examples, not published benchmark results.

Previous: Word sense Next: Word-sense induction