Machine Translation — AI Glossary

Reference details

Related terms

• BLEU
• Multilingual Model
• Text Classification

How it works

1. Dedicated translation models are trained on aligned sentence pairs, and the training objective is the translation itself rather than general next-token prediction.
2. Source and target are segmented with a shared subword vocabulary so that related forms across languages can share pieces, which is what makes multilingual models practical at all.
3. Segmentation is a design decision with consequences: translating sentence by sentence loses the context needed to resolve pronouns, formality and terminology consistency, while translating whole documents costs more and risks drift.
4. General-purpose language models translate without being trained specifically for it and win where context decides — pronoun gender, register, domain vocabulary, terminology held consistent across a document.
5. Automatic evaluation has moved from surface-overlap metrics to learned metrics that correlate better with human judgement, because overlap scores penalise correct translations that use different words.
6. The reliability question is segment-level rather than corpus-level: a good average over a document still leaves individual sentences wrong, and the errors concentrate in named entities, numbers and negation.

Example: Why a 98% per-segment rate still means most documents contain an error

Take a document of 50 segments and suppose each is translated acceptably with probability 0.98 — a high rate for any system. The probability that the whole document is clean is 0.98^50 = 0.364, so 1 - 0.364 = 63.6% of documents contain at least one unacceptable segment. At 100 segments it is 1 - 0.98^100 = 86.7%, and at 200 segments 1 - 0.98^200 = 98.2% — effectively every long document. Raising the per-segment rate to 99.5% gives 1 - 0.995^50 = 22.2% of 50-segment documents with an error, which is better and still means roughly one document in five needs a human to find the problem. This is the arithmetic behind why professional workflows are post-editing rather than translation: the question is never whether the output is good on average but how expensive it is to locate the few bad segments, and that cost scales with document length exactly as the probability above does. It also explains why the useful product feature is confidence per segment rather than a better average — flagging the suspect 2% turns a full review into a spot check.

Commonly confused with

Translation quality and fluency are different properties and automatic metrics have historically rewarded the wrong one. A fluent output reads naturally in the target language; an accurate one preserves the source's meaning, including the parts that are awkward to express. Modern systems are fluent enough that errors no longer look like errors — a mistranslated clause arrives in confident, idiomatic prose, which is harder to catch than the broken syntax older systems produced. So fluency has become a reason for more careful review rather than less, and the errors that survive are concentrated exactly where a fluent reader would not look: numbers, names and negation.

When to use it

Reach for it when:

• High-volume content where post-editing is the workflow and the system's job is to reduce human effort rather than replace it.
• Gisting, where the reader needs to know roughly what a document says and approximate is adequate.
• Context-dependent translation — literary, marketing, domain-specific — where a general model's handling of register and terminology is worth its cost.
• With per-segment confidence, which turns a full review into a targeted one.

Reach for something else when:

• Legal, medical or safety-critical text without professional review, where a fluent mistranslation is more dangerous than an obvious one.
• Low-resource language pairs without evaluating that specific pair, since aggregate multilingual scores hide very wide variation.
• Judged by corpus-level averages when segment-level correctness is what matters, which the compounding above makes clear.
• Content where numbers, names or negation carry the meaning, which is precisely where surviving errors concentrate.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Machine Translation — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path