Does induction require labeled training examples?

Its defining aim is to learn sense groupings without a fully prescribed labeled inventory. Particular methods may still use pretrained models or other resources.

Can it discover a sense missing from a dictionary?

It can reveal a recurring use in the data, but someone must assess whether that use warrants a separate sense description.

Must the number of senses be known in advance?

Not always. Some clustering methods require a chosen number of groups; others infer a structure. Either choice introduces assumptions to evaluate.

AI Glossary / Natural language and speech / Intermediate

Word-sense induction

Definition

Word-sense induction is the automatic discovery of different uses or senses of a word from the contexts in which it occurs. It groups usage patterns without requiring every occurrence to be assigned to a predefined dictionary sense during learning.

What it means

A word can have several meanings, and a specialist collection may contain uses absent from a general dictionary. Sense induction starts from examples and looks for recurring contextual differences.

This differs from word-sense disambiguation with a fixed inventory, where the task is to select among already specified senses. An induction system may discover clusters first and later assign new occurrences to them. The clusters do not necessarily come with readable definitions.

A discovered cluster is evidence of a pattern, not proof of a universally separate meaning. Topic, writing style or document source can create clusters that do not correspond to the sense distinctions a reader needs. The appropriate level of detail depends on the application and the data.

How it works

A workflow gathers occurrences of a target word and represents their contexts. Representations may use nearby words, grammatical relationships or contextual embeddings. A clustering method then groups similar occurrences.

Reviewers can inspect representative examples and identify whether groups have coherent meanings. A later model may map new occurrences to the induced groups. SemEval-2010's word-sense-induction task explicitly separated learning from unseen-instance tagging and evaluation.

Evaluation should consider both mixed clusters and excessive splitting. Putting all occurrences into one group loses distinctions; putting every occurrence into its own group avoids mixing but gives little reusable organization. Comparing clusters with labeled examples can help, but the comparison must account for arbitrary cluster identifiers and possible differences in sense granularity.

Three practical examples

A terminology team explores uses of “model” in documents about product designs, business forecasts and machine-learning systems.
A search researcher groups contexts of an ambiguous term to offer meaning-based result exploration.
A lexicographer examines whether a specialist corpus provides evidence for a previously unrecorded use of a familiar word.

Think of it this way

Think of sorting unlabeled photographs into groups based on what appears similar, then examining the groups to decide what their shared theme is. The labels emerge after the grouping rather than being supplied on every photograph beforehand.

The analogy is limited because word meaning depends on linguistic context and can have fuzzy boundaries. A visually tidy grouping, like a mathematically tidy cluster, is not automatically the most useful interpretation.

Worked example

Inspecting induced groups

An invented collection contains six occurrences of “port”. Three concern ships and cargo; three concern network connections. Suppose an induction system produces cluster A containing all three maritime occurrences plus one network occurrence, and cluster B containing the other two network occurrences.

A reviewer sees that A mixes two intended senses, while the network sense is split across A and B. The clusters' names A and B are arbitrary; renaming them would not change the result.

Consider pairwise grouping as a teaching check. The two gold groups of three contain 3 + 3 = 6 same-sense pairs. The predicted groups of four and two contain 6 + 1 = 7 pairs. Of those predicted pairs, four are genuinely same-sense: three maritime pairs in A and one network pair in B.

Pair precision is 4/7, about 57.1%, and pair recall is 4/6, about 66.7%. This example makes both mixing and splitting visible. It is one illustrative evaluation, not a complete judgment of linguistic usefulness.

Illustrative case study

A fictional engineering publisher studies the term “load” across its manuals. An initial clustering divides documents mainly by product family, even though several families use “load” in the same electrical sense.

The team examines sample contexts and changes its representation to focus more closely on the target word's local relationships. It then checks whether clusters distinguish relevant uses, such as electrical demand and placing material into a device, without simply reproducing document categories.

Editors give provisional descriptions to useful groups and retain ambiguous examples for review. They do not publish every cluster as a new dictionary sense. This case illustrates validation of an exploratory tool, not a measured deployment result.

Common misconceptions and mistakes

Equating every cluster with a distinct, established dictionary meaning.
Confusing induction with choosing from a fixed sense inventory.
Assuming cluster numbers have an inherent order or meaning.
Evaluating only on the same occurrences used to construct the clusters.
Ignoring rare senses because the largest groups look coherent.
Mistaking topic or source differences for differences in the target word's meaning.

Business use

Sense induction can help organizations explore terminology and discover how words are used in their own documents. It is particularly useful when a fixed vocabulary resource does not reflect local practice.

Choose a corpus that represents the intended domain and preserve enough context for review. Remove accidental duplication that would make repeated wording dominate the result. Test whether induced groups help a concrete task, such as routing searches or preparing a glossary.

If clusters will become published definitions, require linguistic and domain review. Check examples near boundaries, not only the clearest examples in each group. Keep the distinction between exploratory groups and approved terminology visible. A model can suggest distinctions, but the decision that a distinction is useful for readers remains an editorial judgment.

Check your understanding

A system groups every occurrence of a word by the website it came from. Is that sufficient evidence that it discovered word senses?

Answer: No. It may have learned source or topic differences. Representative contexts must show that the grouping reflects relevant differences in the word's use.

Related terms

Word sense disambiguation
Contextual embeddings
Clustering
lexical semantics
contextual embedding
sense inventory
polysemy
cluster evaluation
corpus analysis

Learn next

Read word-sense disambiguation and cluster evaluation. Collect a small set of contexts for one ambiguous word, group them without dictionary labels, and compare the result with a human interpretation. Explain any topic-based grouping or uncertain boundary.

Word sense disambiguation

Sources

Manandhar, Klapaftis, Dligach and Pradhan, SemEval-2010 Task 14: Word Sense Induction and Disambiguation: https://aclanthology.org/S10-1011.pdf. This primary task report describes learning senses from text and evaluating induced groups on unseen instances. The port-pair calculation and publisher scenario are original illustrations.

Previous: Word sense disambiguation Next: Word shape