Are all idioms multiword expressions?

Many familiar idioms are, but terminology and coverage vary. The broader MWE category also includes expressions with relatively transparent meanings.

Must the words be adjacent?

No. Some expressions allow intervening material, as in a separable phrasal verb.

Does tokenizing an expression as one token solve it?

No. That may help one representation, but it does not automatically resolve meaning, grammatical variation, or literal-versus-idiomatic ambiguity.

AI Glossary / Natural language and speech / Intermediate

Multiword expressions

Definition

Multiword expressions are combinations of words with conventional lexical, grammatical, or meaning properties that require treatment beyond an unrestricted combination of the individual words. They include many idioms, phrasal verbs, fixed expressions, and conventional collocations, with boundaries depending on the linguistic framework and task.

What it means

Some expressions mean something that is hard to predict from their parts. Others are fairly transparent but favor particular wording or grammatical patterns. The category is therefore broader than colorful idioms whose meanings cannot be guessed.

A system translating “call off the inspection” should recognize cancellation, rather than translate “call” and “off” independently. It must also recognize “call the inspection off,” where the relevant words are separated. Treating every multiword expression as a single uninterrupted string misses this flexibility.

Conversely, merely appearing together does not make a sequence a special expression. “The blue folder” can be an ordinary compositional phrase. Corpus frequency helps discover candidates, but frequent word sequences and conventional lexical expressions are not identical sets.

How it works

A language resource can record an expression's components, meaning, allowed inflections, and permitted variation. A detector may combine lexical entries with grammatical patterns or learn contextual representations from annotated examples.

For a fixed expression, exact matching may be a useful starting point. Variable expressions require more: intervening words, tense changes, or alternative word orders may preserve the same expression. Literal and idiomatic uses may share the same words, so context still matters.

Copestake and colleagues' research on reusable lexical resources emphasizes both variability and compositionality. A practical design should preserve the parts when their individual contributions matter, while also representing the conventional whole. Collapsing every expression into an opaque token can discard useful grammar.

Three practical examples

Support translation: interpret “the battery gave out” as a failure in the relevant context rather than a literal act of giving something away.
Search: connect “call off the visit” with “cancel the visit” when the document describes a cancellation.
Writing assistance: recognize a conventional combination such as “heavy rain” and assess suggested substitutions in context rather than assuming every near-synonym fits equally well.

Think of it this way

Think of a familiar tune made from several notes. Recognizing the pattern adds information beyond identifying each note, and some performances allow variation while preserving the tune. The analogy is limited because words also participate in grammar, and the same sequence may support a literal reading in another context.

Worked example

Compare three invented service messages: A, “Please call off the inspection”; B, “Please call the inspection off”; and C, “Please call the inspector from the office.” A and B contain the cancellation expression “call off”; C does not.

An exact adjacent-word search for “call off” finds A but misses B. In this three-message teaching set, there are two intended positive cases, so that search has recall 1/2 = 50%. Its single retrieved case is correct, giving precision 1/1 = 100%.

A pattern that accepts any “call” followed later by “off” could capture B, but would need careful tokenization and syntax to avoid unrelated occurrences or misleading fragments such as “office.” The example does not establish a production detector's accuracy. It shows why flexible recognition and contextual validation are separate from simply widening a string search.

Illustrative case study

Fictional case: Orchard Assist translates equipment-support conversations. Reviewers see that a customer saying a motor “gave out” receives instructions about locating an object that was supposedly handed over.

The team adds contextual examples of failure expressions and their literal alternatives. It tests both fixed and separated constructions rather than building a list of isolated word replacements. Reviewers check whether the resulting advice preserves the customer's meaning. This scenario illustrates error analysis; it makes no measured claim about translation gains.

Common misconceptions and mistakes

Not every multiword expression is completely fixed, contiguous, or semantically opaque. Nor does every frequent bigram belong in the category. Different annotation projects may include different classes, so labels must be documented before comparing systems.

Do not assume a dictionary entry settles the reading in every sentence. A literal use can resemble an idiomatic one. Also avoid replacing each component with a synonym independently: conventional combinations may resist substitutions even when the overall meaning seems transparent.

Business use

For a business glossary or translation workflow, collect expressions that occur in actual customer and internal language. Store examples, intended meanings, accepted variants, and contexts where the same words have a different reading.

Evaluate whether errors change actions, sentiment, or domain terminology. A cancellation expression misread as a phone call can send a workflow in the wrong direction. Keep original wording available for review. When comparing models, include both familiar expressions and less frequent variants so success does not depend only on memorizing an exact phrase.

Check your understanding

A detector recognizes “call off the visit” but not “call the visit off.” What limitation does this reveal?

Answer: It likely relies on contiguous wording or lacks the construction's permitted variation. The meaning can remain cancellation despite the changed position of the object.

Related terms

Learn next

select one expression and write a fixed form, a permitted variant, and a superficially similar literal sentence. Explain which properties remain constant. Then compare phrase matching with syntactic or contextual recognition, checking both missed expressions and false positives.

Sources

Sources: Copestake et al. (2002), Multiword expressions: linguistic precision and reusability, lexical representation, variability, and compositionality: https://aclanthology.org/L02-1145.pdf . The three-message retrieval calculation and Orchard Assist scenario are original. The source describes a specific research approach; this entry does not treat its historical implementation as a current universal standard.

Previous: Multiway data analysis Next: Music alignment