Learn AI /Computer Vision and Language Technology /Machine translation
Machine translation
- Starter
- Computer Vision and Language Technology
- Communications teams
- Programme & data teams
Also called: automatic translation, MT.
Back to the milestone you came from
In five seconds
What it is
Machine translation uses computer systems to convert text from one language into another using learned patterns, statistical methods or hand-built rules.
In plain words
You give the system text in one language, and it gives back a version in another language. It can save time, but it can still miss meaning, so important translations should be checked by a person.
An everyday example
Imagine predictive text on a phone, but instead of completing a sentence in the same language, it builds a sentence in another language. It considers the source message and the words it has already produced, then chooses what is likely to come next. This often gives fluent wording, but fluent is not the same as faithful. An unusual expression, a specialised term or a small word such as “not” can shift the meaning, so an important message still needs someone who understands both languages to check it.
A human volunteer can ask questions, notice missing context and use real-world background knowledge in ways a machine system may not. Machine translation usually relies on statistical or learned patterns from training data and the immediate text context. Also, many systems process huge volumes very quickly, which a single volunteer cannot do. So the analogy helps with the idea of conversion between languages, but it can overstate human-style judgement.
Remember thisMachine translation predicts a plausible target-language version from learned or programmed patterns. It is fast, but consequential wording still requires comparison with the source and human review.
Understand it
Understand it
Why it matters
Machine translation can widen access across languages, reduce delays and help teams scan large amounts of multilingual material. For a non-specialist, the key question is not whether it sounds impressive, but whether it is reliable enough for the task. A rough translation may be good enough for triage, search or internal orientation. It may be unsafe for contracts, medical guidance, legal text, public warnings or politically sensitive wording. Knowing this helps you choose when to use it, when to require human review, and when not to use it at all.
How machine translation turns one language into another
After viewing this, a reader should be able to describe the main stages from source text to translated output and explain why review may still be needed.
01
1. Source input
A sentence or document starts in one language, such as French, Arabic or Nepali.
Draw a text card labelled 'Source text' with a short sample sentence and a language tag on the left.
02
2. Analyse context
The system breaks the text into parts and uses surrounding words to estimate meaning, grammar and likely equivalents.
Show a processing box with small tokens or word blocks entering it, plus icons for context and grammar.
03
3. Generate translation
The system produces a version in the target language, choosing wording that best fits the patterns it learned from many examples.
Draw a second processing box leading to a new text card labelled 'Target text' in another language.
04
4. Check before use
For important content, a person reviews the result for accuracy, tone, terminology and any safety or legal issues.
Add a final box with a human review icon, a checklist, and a note 'required for high-risk or public-facing use'.
A left-to-right flow diagram with four boxes and arrows. It starts with source language text, moves through analysis of words and context, then generation of target-language wording, and ends with human review for important uses. Read from left to right to see how input becomes output.
Hovering or tapping each stage reveals one common failure point, such as ambiguity, idioms, missing context or domain-specific terms.
Describe this without the diagram
The diagram is a four-step flow from left to right. Step 1 shows source text in one language. Step 2 shows the system analysing words and context rather than matching words one by one. Step 3 shows target text produced in another language. Step 4 shows human review, which is especially important for sensitive, legal, medical or public-facing content because fluent wording can still be wrong.
With reduced motion: A static four-box flow chart is shown instead of animation. Arrows connect source text, context analysis, translation output and human review, with short notes under each box about what can go wrong at that stage.
Machine translation takes language in one form and produces language in another language. The input might be typed text, a document, subtitles, or speech that has first been turned into text. The output is a translated version that aims to keep the meaning, tone and structure usable for the reader.
In simple terms, the system looks at the words and phrases, the order they appear in, and the wider context around them. Modern systems usually learn these patterns from very large collections of translated examples, often called a corpus. A corpus is a set of texts gathered for analysis or training.
The system does not translate by replacing each word one by one in a perfect way. Languages organise meaning differently. Grammar, idioms, politeness, legal terms and cultural references can all change the right answer. Because of this, the same source sentence can have more than one reasonable translation.
What comes out can be very helpful, but it still needs checking. A translation may sound fluent while changing a date, dropping a negation, or choosing the wrong sense of a word. That matters even more in health, law, public information and multilingual programme work, where a small wording error can change meaning or create risk.
How “Predictive text that changes languages” maps onto the real thing
| In the story | In the AI system |
|---|---|
| feedback cards in different languages | source text that needs translation |
| volunteer writes an English version | system produces output in the target language |
| choosing natural phrasing instead of word-for-word copying | translation based on meaning and context, not only dictionary matches |
| local expressions or unclear writing causing difficulty | idioms, ambiguity and poor input quality leading to errors |
A worked example
A desk officer at Headquarters is finalising talking points for front-office clearance before close of business and receives a short French update from an implementing partner.
Starting from: French source: “Le partenaire transmettra la version révisée vendredi, sous réserve de validation interne.”
- The system divides the source into words or subword units and identifies the planned action, the Friday deadline and the condition attached to it.
- Using patterns learned from matched French and English texts, a neural model predicts the English sequence one unit at a time.
- It produces: “The partner will send the revised version on Friday, subject to internal validation.”
- A bilingual reviewer checks the draft against the source and the office's usage. Here, “validation interne” means internal clearance, so the wording is revised before the talking points enter the file.
Result: The final text reads: “The partner will send the revised version on Friday, subject to internal clearance.” The machine produced a plausible draft, but the reviewer supplied the context-specific terminology.
In a United Nations workplace
A programme officer in a documentation and library service is screening multilingual applications against a rostered profile before the shortlist goes to the chief of section at close of business. An approved translation tool produces rough English versions for initial review.
Machine translation can help the programme officer understand applications submitted in different languages. Because the system generates likely target-language wording from learned patterns, specialist terms, qualifications and negation may be rendered incorrectly. The original text and bilingual review remain necessary for screening decisions.
Before you try thisApplications contain personal and employment information. They must not be uploaded to a public or otherwise unapproved translation service, because this may disclose personal data or place it outside authorised controls. Use only a tool approved for that material, minimise the data processed, and never reject or rank an applicant solely on a machine translation.
Where you may hear it
- During the first reading of multilingual draft text
- At a vendor demo for an ICT governance board
- While testing live captions before a side event
- During close-of-business document triage in a regional bureau
Avoid confusion
Avoid confusion
Often confused with
Text generation
A system produces new written words from a prompt, patterning its output on examples seen during training.
Low-resource language
This status applies to a language and task when useful data, software tools or evaluation material are limited for building and assessing technology.
Common misconceptions
Common misconception: It just replaces each word with the matching word in another language.
Correction: Useful systems do more than word swapping because meaning depends on grammar, word order and context. A direct dictionary replacement often breaks idioms, negation, gender, formality or sentence structure.
Common misconception: If the output sounds fluent, it is probably correct.
Correction: Fluent wording can still contain serious meaning errors. Systems may choose the wrong sense of a word, miss a domain-specific term, or quietly change who did what to whom. That is why important translations need review by someone qualified.
Common misconception: Machine translation and text generation are the same thing.
Correction: They overlap but are not identical. Machine translation is a task-focused system or capability for changing one language into another while preserving meaning. Text generation is a broader term for producing text, including summaries, answers or creative writing.
Quick check
When is machine translation most appropriate without relying on it as the final authority?
A rough draft for orientation or triage is a common low-risk use. The other situations involve high stakes, confidentiality, legal effect or public safety, where machine output alone is not enough and unapproved public tools may be inappropriate.
Go deeper
Go deeper
The simple version
You give the system text in one language, and it gives back a version in another language. It can save time, but it can still miss meaning, so important translations should be checked by a person.
The technical version
Machine translation maps a source-language sequence to a target-language sequence. Older approaches used hand-built rules or phrase statistics from bilingual text. Many current systems use neural models trained on parallel corpora, meaning matched translations of the same content in two or more languages. Autoregressive models often estimate which target word or subword unit is most likely given the source text and the target units already produced. In practice, quality depends on training data, domain match, language pair, terminology, input quality and available context. Speech translation adds extra stages, such as speech recognition before translation and sometimes speech synthesis after it. Evaluation can include automatic scores and human assessment, but neither guarantees fitness for a specific real-world use.
Reviewed external resources
EncyclopaediaMachine translation (opens in a new tab)Wikipedia
This link gives a broad overview of machine translation, its history and main approaches, and suits readers who want a general reference.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary gives short definitions of machine learning terms that may appear in explanations of translation systems, and suits readers who want quick clarification.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives practical machine learning background that can help readers understand how translation systems are trained, and suits beginners who want more technical context.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Machine translation (opens in a new tab)YouTube search
This is a YouTube search results page for video explanations of machine translation, useful for readers who prefer to compare several video styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Machine translation (opens in a new tab)Reddit search
This is a Reddit search results page for community discussions about machine translation, useful for readers who want to see common questions, experiences and debates from users and practitioners.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Machine translation predicts a plausible target-language version from learned or programmed patterns. It is fast, but consequential wording still requires comparison with the source and human review.