Machine Translation: How It Works, Types and Research Use

Machine translation (MT) is software that uses artificial intelligence to translate text from one language to another without a human translating it. This guide is for researchers who need to read, screen or prepare material in another language: it explains how MT works, when its output is good enough, and how to run a real translation request against a documented API.

Checked against vendor and professional-association documentation on 2026-11-07 (sources listed at the end).

What is meant by machine translation?

Machine translation is a natural language processing task: mapping text in one language to text in another. Amazon Web Services describes it as using AI to automatically translate text between languages with no human involved, in two steps: work out the meaning of the source text, then express that meaning in the target language. It is distinct from human translation, in which a translator produces the target text; the site’s dictionary entry on AI translation and the entry on translation (work) cover how translated output is treated in scholarly records.

The main types of machine translation

Type	How it works	Strengths and limits reported by the source
Rule-based (RBMT)	Dictionaries plus linguistic rules; IBM describes direct, transfer and interlingua levels	Predictable, but struggles with unknown words and needs manual dictionary updates (AWS); needs large dictionaries (IBM)
Statistical (SMT)	Learns from existing human translations; a translation model proposes phrases with probabilities and a target-language model checks fluency (IBM)	Needs millions of words per language pair (AWS); out-of-vocabulary words are a problem (IBM)
Neural (NMT)	Encodes the whole source sentence, then decodes it into the target language, often with transformer networks (IBM)	Described by AWS as the current standard; pronoun ambiguity remains a challenge (IBM)
Hybrid	Combines rule-based and statistical approaches (AWS)	Used to combine the strengths of both

Is ChatGPT a machine translation?

ChatGPT is a general-purpose large language model, not a dedicated translation system, but a language model can be asked to translate and then performs the same task. IBM notes that prompting large language models is being explored for MT, and Google’s Cloud Translation documentation offers a translation-specific LLM model alongside its default neural model. So “machine translation” names the task, and a chat-style LLM is one tool that can do it. Whichever tool you use, the cautions in the next section apply equally.

Can you give me an example of machine translation?

AWS gives the example of a multinational online store translating customer product reviews so that readers in every market can read them. The research equivalent is translating abstracts or full texts of non-English papers so that a reviewer can decide whether a study is relevant before commissioning a full human translation. The API walkthrough below shows a concrete request and response.

What machine translation is used for in research

• Screening and gisting. The American Translators Association (ATA) says generic MT can be an excellent tool for gaining an approximate idea of a text. That fits first-pass triage of foreign-language literature.
• Drafts for human post-editing. AWS lists machine output as a starting point for human post-editing. The ATA states that for customised, field-trained MT, post-editing by a professional translator is still required unless errors and non-idiomatic style can be tolerated.
• Speech. IBM describes speech-to-speech translation as speech recognition followed by neural translation.

When not to rely on it

The ATA advises against relying on MT where “the law, a life, or your bottom line is involved”, says MT does not understand meaning or context, and says creative texts should be translated by a human. For participant-facing materials such as consent forms, see the site’s guide to informed consent in research and the entry on translated informed consent materials; do not treat raw machine output as a finished translated document.

The ATA also notes that free online services may store and analyze the content you submit, so check a service’s terms before pasting in confidential or unpublished material.

IBM lists why translation is hard for software: ambiguity (polysemy, idioms, unclear pronouns), differing word order between languages, literal versus sense-based rendering, and poetry.

What is the best machine translation?

None of the sources read for this guide rank services, and no ranking is claimed here. The ATA’s position is that customised MT trained on a specific field is more reliable than generic MT. A practical consequence: test candidate tools on a sample of your own text, in your own language pair and subject area, and have a fluent reader compare the results.

Install or access (official source only)

The walkthrough below uses Google’s Cloud Translation documentation page “Translate text” (read 2026-11-07). Setup (a Google Cloud project, enabling the Cloud Translation API and credentials) is covered in Google’s own setup pages; use those and not a third-party mirror.

Option	Authentication shown in the docs	Notes from the docs
Basic (v2)	API key passed as ?key=	Default model is Google NMT; does not support custom models; omit source to auto-detect the source language
Advanced (v3)	Bearer token from gcloud auth print-access-token plus an x-goog-user-project header	Supports custom models; response includes a model field when a custom or LLM model is used

Free alternatives

Open-source tools can be run on your own machine, which also addresses the ATA caution about free online services storing submitted text. Both entries below are taken from each project’s official README (read 2026-11-07); quality for your language pair and subject area is not assessed here, so test on your own sample.

Tool	What the project says it is	Licence
Argos Translate	Open-source offline translation library; uses OpenNMT; usable as a Python library, command-line or GUI application	MIT or CC0 (dual licensed)
LibreTranslate	Free and open source machine translation API, entirely self-hosted and offline capable; engine powered by Argos Translate	GNU AGPL v3

Step by step: translate a sentence with the Basic API

1. Create or choose a Google Cloud project, enable the Cloud Translation API and create an API key, following Google’s setup documentation.
2. Send a POST request to https://translation.googleapis.com/language/translate/v2?key=YOUR_API_KEY with the header Content-Type: application/json; charset=utf-8.
3. Put the text in q (an array of strings) and the target language code in target. Optionally set source; leave it out to have the API detect the language.
4. To use the translation LLM instead of the default neural model, add model with the value projects/PROJECT_ID/locations/REGION_NAME/models/general/translation-llm. Leave model out for the default NMT model.

curl -X POST 
  -H "Content-Type: application/json; charset=utf-8" 
  -d '{
    "q": ["The old lighthouse stood on the edge of the cliff."],
    "target": "es"
  }' 
  "https://translation.googleapis.com/language/translate/v2?key=YOUR_API_KEY"

Output and how to read it

The Basic API wraps the result in data: {"data": {"translations": [{"translatedText": "..."}]}}. When you omit source, the response can include a detectedSourceLanguage field, which is worth checking against what you expected. The Advanced API returns translations with one translatedText per input, in input order.

The text is only a machine rendering. Before using it in a protocol, data extraction form or publication, have a competent human reader review it, as the ATA guidance above recommends.

Troubleshooting

• HTML tags in your input. With the format set to HTML, tags are kept and only the text between them is translated; the documentation says the API does not translate HTML tags. XML and other markup are not supported and the result is undefined. Use plain text (text/plain in v3, format set to text in v2) for ordinary prose.
• Wrong or unexpected source language. If you left source out, check detectedSourceLanguage; set source explicitly when the text mixes languages.
• Custom model not accepted. The Basic API does not support custom models; use the Advanced API.
• Authentication errors. Basic requests in the example use an API key; Advanced requests use a bearer token from gcloud auth print-access-token and the project header. Do not mix the two styles in one request.

Frequently asked questions

What is meant by machine translation?

Software translating text between languages with no human translator involved, in two steps: understand the source meaning, then express it in the target language (AWS).

Is ChatGPT a machine translation?

It is a general-purpose language model that can perform the machine-translation task. Dedicated systems such as neural MT services are built specifically for it.

What is the best machine translation?

No source read here names one. Field-trained MT is more reliable than generic MT (ATA); test on your own material.

Can machine translation replace a professional translator?

The ATA says post-editing by a professional translator is required unless errors and non-idiomatic style can be tolerated, and that creative texts should be translated by a human.

Sources

• Amazon Web Services, “What is Machine Translation?” aws.amazon.com/what-is/machine-translation (read 2026-11-07)
• IBM, “What is machine translation?” ibm.com/think/topics/machine-translation (read 2026-11-07)
• American Translators Association, “Machine Translation” atanet.org/resources/machine-translation (read 2026-11-07)
• Google Cloud, “Translate text (Basic)” docs.cloud.google.com/translate/docs/basic/translating-text (read 2026-11-07)
• Argos Translate README github.com/argosopentech/argos-translate (read 2026-11-07)
• LibreTranslate README github.com/LibreTranslate/LibreTranslate (read 2026-11-07)