Must the substitute be one word?

Not necessarily. Some task definitions allow multiword alternatives, but the permitted replacement span and formatting should be specified.

Is lexical substitution the same as word-sense disambiguation?

No. They are related, but substitution outputs an alternative expression while disambiguation typically selects a sense from an inventory.

Can there be more than one correct substitute?

Yes. Several alternatives may preserve the intended meaning, sometimes with small differences that need contextual judgment.

AI Glossary / Natural language and speech / Intermediate

Lexical substitution

Definition

Lexical substitution is the task of finding a replacement word or expression for a target word in context while preserving the intended meaning. The replacement must fit the particular sentence, not merely resemble the target in a thesaurus. Several alternatives may be acceptable, and some contexts permit no straightforward replacement.

What it means

Consider how a word changes meaning across sentences. A substitute suitable for a physical object may be unsuitable when the same spelling describes an abstract idea. Context determines which alternatives preserve the message.
The task differs from assigning a dictionary sense label: its output is a usable substitute. It also differs from unrestricted paraphrasing, which may restructure the entire sentence.
Lexical substitution can support writing assistance or simplification, but a valid substitute is not automatically simpler, more formal, or more appropriate for a brand. Those are additional objectives that must be specified separately from meaning preservation.

How it works

Identify the target word and the context that constrains its interpretation. Determine whether it belongs to a multiword expression whose meaning cannot be handled by replacing one component independently.
Generate candidate replacements from lexical resources, corpus evidence, or a language model. Filter or rank candidates for the target's contextual meaning, grammatical role, and compatibility with surrounding words.
Adjust the form of a candidate where the task permits it, preserving tense, number, or other required features. Check the complete sentence after substitution.
Evaluate against human judgments or other suitable reference decisions. Because multiple substitutes can be reasonable, evaluation should account for alternatives rather than assuming that a single reference word exhausts all valid answers.

Three practical examples

A writing assistant suggests a context-appropriate alternative to a repeated word while keeping the statement's meaning.
A simplification experiment selects replacements that both preserve meaning and suit a defined reader vocabulary.
A semantic evaluation task asks annotators for possible substitutes, providing evidence about how they interpret a target word in its sentence.

Think of it this way

Think of replacing a component in an assembled object. The replacement must fit the surrounding parts and preserve the intended function, not simply look similar on a shelf.
The analogy has limits: language often permits several acceptable alternatives, and small differences in tone or emphasis can matter. There may be no perfectly interchangeable word for every context.

Worked example

Use this original toy sentence: “Mina gave a brief explanation.” The target is “brief,” and the candidates are “short,” “small,” and “concise.”
“Short” preserves the duration-or-length idea naturally in this sentence. “Small” is not a normal equivalent for the intended description of an explanation. “Concise” is plausible, but can suggest efficient expression without unnecessary words, adding a nuance beyond mere shortness.
Suppose a toy annotation exercise with five reviewers allows one preferred answer each: three select “short” and two select “concise.” Under a deliberately simple vote-share summary, their shares are 3/5 = 60% and 2/5 = 40%.
Those shares describe this invented panel's preferences. They are not universal correctness probabilities, and they are not the official SemEval scoring formula. A careful application can retain both candidates while recording the nuance that distinguishes them.

Illustrative case study

In a fictional project, Hazel Guides develops a rewriting assistant. Its first candidate list replaces a target with any dictionary neighbor, sometimes changing intensity or making a phrase unnatural.
The team introduces sentence-level review and separates two questions: does the replacement preserve meaning, and does it improve suitability for the intended readers? It allows “keep the original” when no candidate satisfies both.
Its test set includes words inside fixed expressions as well as ordinary uses. This invented workflow illustrates why generation of candidates and acceptance of candidates require different checks, without claiming measured gains.

Common misconceptions and mistakes

Do not assume that a dictionary synonym works in every context. Sense, grammar, and collocation constrain substitution.
Do not treat relatedness as interchangeability. A cause, effect, category, or opposite can be closely related to a word without replacing it appropriately.
Do not equate substitution with simplification. A meaning-preserving alternative may be less familiar to the reader.
Do not force replacement when the original is already the clearest choice.
Do not assume that disagreement among annotators always indicates an error. Different acceptable choices can reflect nuances and legitimate preferences, although clear meaning changes still need to be rejected.

Business use

For a business, substitution can support assisted editing, terminology adaptation, and controlled rewriting. Define whether the goal is clarity, consistency, reading accessibility, or stylistic variation before ranking acceptable replacements.
Evaluate changes in full context, including any facts, qualifiers, and commitments surrounding the target. An apparently harmless adjective or verb replacement can alter a promise or the strength of a claim.
Track rejection reasons such as wrong sense, changed intensity, grammatical mismatch, or inappropriate register. These distinctions reveal more about the tool's usefulness than simply counting how many alternatives it proposes.

Check your understanding

In the toy example, does the 60% vote share establish that “short” has a 60% probability of being correct in every sentence containing “brief”?

Answer: No. It records the choices of five invented reviewers for one sentence. Other contexts can require different substitutes, and the count is not a calibrated universal probability.

Related terms

Learn next

Choose one ambiguous word in two different sentences and propose three substitutes for each. Explain why an alternative that works in the first sentence fails in the second.
Then review whether accepted replacements change tone, intensity, or grammatical structure. Read the lexical substitution task paper to see why collecting multiple human alternatives provides a richer reference than selecting one dictionary label.

Sources

McCarthy and Navigli, SemEval-2007 Task 10: English Lexical Substitution Task, original task description: https://aclanthology.org/S07-1009.pdf Edmonds and Hirst, Near-Synonymy and Lexical Choice, original research on differences among related alternatives: https://aclanthology.org/J02-2001.pdf The sentence, candidate judgments, and vote-share exercise are original illustrations, not benchmark annotations or official scoring procedures. Independent review: original task setup, multiword issues and scoring consulted.

Previous: Lexical resource Next: Lexicalized PCFG