Must a stored sentence exactly match the input?

No. Adaptation from similar examples is central to the approach, although exact matches can also be useful.

Can examples be shorter than a sentence?

Yes. Phrase-level and other paired units can support translation.

Does EBMT avoid grammar completely?

Not necessarily. Matching and adaptation can use grammatical or semantic information, and example-based components can be combined with rule-based methods.

AI Glossary / Natural language and speech / Intermediate

Example-based machine translation

Definition

Example-based machine translation, or EBMT, translates new input by retrieving relevant bilingual examples and adapting their target-language counterparts. The examples may be phrases, sentences, or larger paired texts. Its central idea is translation by analogy: reuse suitable previously translated material while accounting for the differences in the new input.

What it means

A bilingual example contains an expression and a corresponding translation. When a new expression resembles a stored one, that pair can provide evidence about wording, structure, and contextual choices.
The system must do more than notice shared words. It needs to identify which parts correspond and which differences require changes in the target language. Similar-looking source sentences can require different translations when their senses or grammatical relationships differ.
EBMT overlaps with other data-driven translation ideas, but not every system trained on examples is normally called example-based machine translation. The term emphasizes the operational use of retrieved examples and adaptation, rather than merely the existence of training data.

How it works

Build or obtain a collection of aligned source-target examples. Record enough information to identify corresponding parts and distinguish relevant contexts.
For a new input, retrieve examples using lexical, structural, semantic, or combined matching criteria. Rank the candidates and identify the useful correspondences.
Adapt a retrieved translation or combine compatible translated fragments to fit the new source. The required changes may involve word choice, agreement, morphology, or order rather than simple copying.
Check the resulting output and define behavior when no sufficiently useful example exists. Systems can combine example-based components with other methods; the exact retrieval and adaptation mechanisms vary across implementations.

Three practical examples

A research translator reuses a previously translated phrase pattern while adapting a place name and associated wording.
A controlled-domain system retrieves examples containing a specialized expression whose translation depends on context.
A hybrid translation pipeline uses bilingual examples for selected constructions and another method for the remaining material.

Think of it this way

Think of a translator consulting a set of worked examples and carefully adapting the closest relevant case. The analogy is bounded because a computer’s matching criterion may favor superficial resemblance, and a stored example does not explain every grammatical or semantic adjustment needed for a new sentence.

Worked example

Suppose a miniature English-to-French example store contains “The red car is here” paired with “La voiture rouge est ici,” and “The red bicycle is here” paired with “Le vélo rouge est ici.”
For the new input “The blue car is here,” a suitable adaptation can reuse the first sentence’s structure and produce “La voiture bleue est ici.” The adjective must agree with the feminine noun voiture; replacing rouge with an uninflected lookup form bleu would miss that requirement.
For “The blue bicycle is here,” the corresponding form is “Le vélo bleu est ici.” These simple sentences illustrate why the relevant example and target-language adjustments matter together.
The examples are constructed for teaching, not outputs from a named engine. They also do not define a universal algorithm: a real system must determine correspondences and agreement through its own representation and adaptation procedure, and more complex input can require changes beyond one adjective.

Illustrative case study

In a fictional project, Cedar Labels maintains a reviewed bilingual collection of short product descriptions. The team experiments with an example-based component because many new descriptions resemble earlier ones but differ in a few attributes.
Reviewers inspect both the retrieved source example and the adapted output. They find that some apparently small substitutions change target-language agreement, so the evaluation includes those cases rather than testing only exact matches.
The imagined project shows how reusable examples can provide evidence while still requiring adaptation. It does not claim a measured accuracy gain or that repeating sentence patterns eliminates translation ambiguity.

Common misconceptions and mistakes

EBMT is not just a dictionary that replaces each source word independently.
It is not automatically identical to a translation-memory tool that presents matches for a human to edit, although the ideas can overlap in a workflow.
A neural system trained on parallel examples is not necessarily EBMT in the traditional sense.
The nearest example under a chosen distance can still express the wrong sense or relation.
Adding more examples is not guaranteed to improve every result if the examples are inconsistent, misaligned, irrelevant, or poorly translated.

Business use

The approach is useful to understand when evaluating systems that reuse bilingual material, especially in repetitive or specialized content. Its practical value depends on example coverage, retrieval quality, and reliable adaptation.
Keep provenance and review status for the examples. Evaluate inputs that are similar but not identical to stored material, including changes affecting agreement, negation, quantities, or word sense. Separate retrieval failures from adaptation failures so the remedy targets the right component. Establish a fallback or review route for poorly supported inputs rather than assuming a nearest match is always acceptable.

Check your understanding

An English sentence differs from a stored example by one noun. Is replacing only that noun in the target translation always sufficient?

Answer: No. The change can affect agreement, articles, word order, or the intended sense. The adaptation must be checked against the target language and context.

Related terms

Parallel corpus
Alignment
Translation memory
Hybrid machine translation
Analogy-based reasoning
Machine-translation evaluation

Learn next

Start with a few reviewed bilingual pairs and distinguish exact reuse from adaptation. Change one feature at a time and identify which target-language elements must change. Then study retrieval similarity and alignment before comparing traditional EBMT with translation memory and other data-driven methods.

Sources

Sumita and Iida, Experiments and Prospects of Example-Based Machine Translation, original research describing retrieval and adaptation: https://aclanthology.org/P91-1024.pdf Nagao, A Framework of a Mechanical Translation between Japanese and English by Analogy Principle, historical original-paper record: https://aclanthology.org/www.mt-archive.info/70/Nagao-1984.pdf Independent source review: Sumita–Iida’s retrieval/adaptation definition and Nagao’s analogy-based fragment translation passage were read. French agreement was cross-checked against OQLF’s noun-group guidance: https://www.oqlf.gouv.qc.ca/francisation/ordres_prof/capsules-grammaticales/accord-groupe-nominal.aspx

Previous: Exact test Next: Exclusion Bias