AI Glossary / Generative AI and language / Intermediate

Machine translation

Definition

Machine translation is the automatic conversion of text or speech content from one human language into another by a computer system. Its goal is to preserve the source meaning in a suitable target-language expression. Translation is more than replacing individual words: grammar, context, terminology, tone, and ambiguity can all affect what a faithful result should say.

What it means

The original language is the source language, and the language being produced is the target language. A good translation preserves relevant meaning while sounding appropriate to its intended readers. These goals can conflict when a literal wording is unnatural or an idiom has no direct counterpart. Machine translation includes rule-based, statistical, and neural approaches. Large language models can also translate, but machine translation is not limited to conversational AI. Quality depends on the language pair, direction, subject matter, available context, and the kind of text being translated.

How it works

In a neural translation system, text is typically split into tokens and converted into numerical representations. An encoder-decoder model represents the source and generates target tokens while attending to relevant source information. Attention helps the decoder use different parts of the input as it writes different parts of the translation. A decoder-only language model can instead generate a translation from a prompt containing the source and instructions. Training may use aligned examples that express corresponding content in two languages, together with other language data. At deployment, decoding selects a target sequence from the model's predictions. That sequence can be fluent even when it mistranslates a number, omits a condition, or resolves an ambiguity incorrectly.

Three practical examples

• A retailer translates product descriptions for a new market, using an approved terminology list and checking measurements and compatibility statements.
• A support team translates incoming messages to identify the customer's issue, then has important outgoing responses reviewed for meaning and tone.
• A researcher translates foreign-language abstracts for preliminary discovery, consulting the original or a qualified translator before relying on a disputed detail. The required level of review depends on what the translation will be used to decide.

Think of it this way

Think of moving the contents of a room into a differently shaped room. The arrangement may need to change, but the important contents should remain. Translation similarly changes word order and expression while preserving meaning. The analogy is limited because meaning is not a set of physical objects: tone, implication, and cultural references may need interpretation, and the source itself may be ambiguous. A polished target text can therefore conceal a mistaken interpretation.

Worked example

Consider the Spanish sentence “El pedido no llegará antes del viernes.” A suitable English translation is “The order will not arrive before Friday.” Another acceptable wording is “The order won't arrive until Friday at the earliest.” Now compare “The order will arrive before Friday.” It is fluent English, but it loses the negation and reverses a central condition. A reviewer can check three elements: the subject is the order, the event is arrival, and Friday is the earliest possible day rather than a guaranteed delivery date. Even “The order will arrive on Friday” is too strong, because the original allows arrival later than Friday. This example shows why checking grammar or shared keywords is insufficient. A good translation preserves both the stated information and the limits of what the source claims.

Illustrative case study

Fictional case: Meridian Parts translates installation notes for a multilingual catalog. Reviewers discover that two different source terms are both rendered as the same target term, obscuring a distinction between a removable cover and a fixed housing. The team creates a terminology list with contextual examples and tests the translation system on representative instructions. It then reviews the revised output alongside the source, paying attention to negation, measurements, and component names. The case illustrates a workflow for detecting domain-specific mistakes; it does not claim that a glossary alone guarantees correct translation.

Common misconceptions and mistakes

One misconception is that a natural-sounding translation must be accurate. Another is that translating the result back into the source language proves fidelity. Back-translation can reveal issues, but both systems may make compatible mistakes. A single reference translation is also not the only valid wording. Automatic metrics based on reference overlap can help compare systems across test collections, but they do not replace meaning-based review of an individual consequential passage. Finally, translation and localization differ: localization may deliberately adapt formats, examples, or cultural presentation beyond transferring the original wording.

Business use

Before selecting a translation workflow, identify language directions, document types, terminology requirements, and the consequences of errors. Test actual content, including short fragments and longer passages where context matters. Separate quality checks for meaning, terminology, readability, and formatting. Measure post-editing effort if humans will revise the output, because an apparently fluent draft can take substantial time to verify. Preserve the source and version history so reviewers can compare changes. For repeat material, approved translations and terminology resources can improve consistency, but new contexts still require attention.

People also ask

Is machine translation the same as transcription?

No. Transcription represents speech in writing, usually in the same language; translation changes the language of the content.

Can it translate an idiom word for word?

It can produce a literal rendering, but that may fail to communicate the intended meaning.

Does one high benchmark score establish quality in every language?

No. Performance must be assessed for the relevant language pair, direction, domain, and use case.

Check your understanding

A source says a shipment may arrive after Monday. The translation says it will arrive on Monday. Is this merely a stylistic difference?

Answer: No. It changes both uncertainty and timing. A meaning-based assessment should flag the translation even if its grammar is correct.

Related terms

• Neural machine translation
• Parallel corpus
• Attention
• Localization
• Post-editing
• source language
• target language
• encoder-decoder model
• translation evaluation

Learn next

Compare two translations of a short paragraph. Mark names, numbers, negation, uncertainty, and technical terms before judging style. Then ask a competent bilingual reviewer whether each version preserves those elements and the overall intent. This makes translation quality more concrete than a general impression of fluency.

More in Generative AI and language

• LLM-as-a-judge - LLM-as-a-judge is an evaluation method in which a large language model assesses an answer, conversation, or other output using specified criteria.
• LLM evaluations (evals) - LLM evaluations, often shortened to evals, are structured tests used to measure how a large language model or a system built around it performs.
• LLM Temperature - LLM temperature is a generation setting that rescales a language model's next-token scores before sampling.
• Masked language model - A masked language model is a language model trained to predict selected hidden or corrupted tokens using the surrounding text.
• Masked language modeling - Masked language modeling is a self-supervised training task in which selected tokens in a text are hidden or altered and a model learns to predict their...
• Mixture of experts - A mixture of experts, or MoE, is a model design that combines multiple component networks called experts through a gating or routing mechanism.

Sources

Bahdanau, Cho, and Bengio, Neural Machine Translation by Jointly Learning to Align and Translate: https://arxiv.org/abs/1409.0473 Papineni and colleagues, Bleu: a Method for Automatic Evaluation of Machine Translation: https://aclanthology.org/P02-1040/ Vaswani and colleagues, Attention Is All You Need: https://arxiv.org/abs/1706.03762

Previous: Machine reading Next: Macroaveraging