NLLB Translation Glossary: CAT Tools, TM, MT, Post-Editing
Last updated: March 2026
Translation Glossary: CAT Tools, TM, MT, Post-Editing
The translation industry has its own vocabulary, and it can be intimidating. CAT, TM, TMS, MT, MTPE, NMT, BLEU, COMET, MQM, TBX, XLIFF — the acronyms alone can fill a page. Whether you are a business buyer commissioning translation services, a developer integrating translation into your product, or a new translator entering the profession, this glossary covers the essential terms you need to know.
Terms are organized by category and cross-referenced with our in-depth guides for further reading.
Core Concepts
Machine Translation (MT)
Automatic translation of text by software, without human involvement. Modern MT uses neural networks trained on billions of sentence pairs. The main approaches in 2026 are Neural Machine Translation (NMT) engines like DeepL and Google Translate, and Large Language Models (LLMs) like ChatGPT and Claude.
MT quality varies dramatically by language pair and content type. For a comparison of leading MT systems, see Google Translate vs DeepL vs ChatGPT.
Computer-Assisted Translation (CAT) Tools
Software that helps human translators work faster and more consistently. A CAT tool is not a machine translator — it is a workbench for human translators that provides features like translation memory, terminology management, quality assurance checks, and integrated MT suggestions.
How it works: The translator imports a source document. The CAT tool segments the text (usually into sentences). For each segment, the tool displays translation memory matches, glossary entries, and optionally an MT suggestion. The translator reviews, edits, or writes the translation for each segment. The final output is an assembled, formatted document.
Leading CAT tools (2026): memoQ, Trados Studio, Smartcat, Phrase, Wordfast, MateCat (open source).
Translation Memory (TM)
A database of previously translated segments (usually sentences) stored as source-target pairs. When a translator encounters a segment similar or identical to one already in the TM, the tool suggests the stored translation.
Match types:
- 100% match (exact match): The segment is identical to one in the TM. The stored translation can often be reused directly.
- 101% match (context match / ICE match): The segment and its surrounding context are identical. Even more reliable than a 100% match.
- Fuzzy match (70-99%): The segment is similar but not identical. The translator uses the stored translation as a starting point and edits the differences.
- No match (< 70%): The segment is too different from anything in the TM. The translator translates from scratch.
TM reduces costs by 20-60% over time as the database grows. For a deeper comparison, see Translation Memory vs AI.
Terminology Management / Glossary / Termbase
A controlled database of approved translations for specific terms. Unlike TM (which stores full sentences), a glossary stores individual terms or short phrases with their approved translations, definitions, usage notes, and sometimes visual references.
Why it matters: If your product is called a “dashboard” in English, you need the same term used consistently in every French translation — not “tableau de bord” in one document and “panneau de contrôle” in another. Glossaries enforce this consistency.
Standard format: TBX (TermBase eXchange) is the ISO standard for exchanging terminology data between tools.
Translation Workflow Terms
Post-Editing (MTPE / PE)
The process of a human translator reviewing and correcting machine translation output. MTPE (Machine Translation Post-Editing) is the industry-standard hybrid workflow in 2026.
Light post-editing (LPE): Fix only critical errors — factual mistakes, obvious grammar problems, nonsensical passages. The result is understandable but may not be perfectly polished. Used for internal content and high-volume low-stakes material.
Full post-editing (FPE): Edit to publication quality. The result should be indistinguishable from human translation. Used for customer-facing content.
Post-editing costs 30-60% less than translating from scratch, because the MT output provides a usable starting point. See our Human vs AI Translation guide for when MTPE is appropriate.
Translation Management System (TMS)
A platform that manages the entire translation workflow — not just the linguistic work, but project management, file handling, translator assignment, review routing, quality tracking, and delivery. A TMS integrates with CAT tools, MT engines, and content management systems.
Leading TMS platforms (2026): Phrase (Memsource), Smartcat, Crowdin, Lokalise, Transifex, XTM.
For platform comparisons, see Best Localization Platforms.
Localization (L10n)
Adapting content for a specific cultural and market context, going beyond translation to include design, formatting, cultural references, and business logic adaptation. The “10” in L10n represents the 10 letters between L and n in “localization.” See our complete Localization vs Translation Business Guide.
Internationalization (i18n)
Engineering a product to support localization — separating text from code, supporting Unicode, designing flexible layouts, enabling RTL rendering. The “18” represents the 18 letters between i and n. Internationalization happens once; localization happens per market.
Transcreation
Creative rewriting of content for a target market, going beyond translation or localization. The source content serves as a brief, and the target content is created fresh to achieve the same emotional and commercial impact. Used primarily for marketing campaigns, slogans, and advertising.
Quality and Evaluation Terms
BLEU Score (Bilingual Evaluation Understudy)
The most widely used automatic metric for measuring MT quality. BLEU compares machine-generated translations to human reference translations by counting matching n-grams (word sequences). Scores range from 0 to 100 (higher is better). For a full explanation, see Machine Translation Quality: BLEU Scores Explained and try our BLEU Score Calculator.
COMET (Crosslingual Optimized Metric for Evaluation of Translation)
A neural MT quality metric developed by Unbabel that uses language models to evaluate translations. COMET considers semantic meaning (not just word overlap) and correlates better with human judgment than BLEU. See Translation Quality Metrics.
xCOMET
An extension of COMET that provides both sentence-level quality scores and error span detection — identifying which specific parts of a translation contain minor, major, or critical errors.
MQM (Multidimensional Quality Metrics)
A framework for human evaluation of translation quality. Trained evaluators classify errors by type (accuracy, fluency, terminology, style) and severity (critical, major, minor). MQM is the gold standard for quality assessment but requires trained evaluators and is expensive to implement at scale.
Quality Estimation (QE)
Predicting translation quality without a human reference translation. QE models evaluate the source and candidate translation directly, making them useful for production environments where reference translations are not available.
Technical Terms
Neural Machine Translation (NMT)
The dominant MT architecture since approximately 2016. NMT uses deep neural networks (typically transformer architectures) trained on parallel corpora to translate text. All major commercial MT engines (Google Translate, DeepL, Microsoft Translator) use NMT. For a deeper understanding, see How AI Translation Works.
Large Language Model (LLM)
General-purpose AI models (GPT-4, Claude, Gemini, Llama) that can perform translation among many other tasks. Unlike dedicated NMT engines, LLMs are not specifically designed for translation but often produce competitive or superior results, especially for context-heavy and creative content.
XLIFF (XML Localization Interchange File Format)
The OASIS standard format for exchanging localization data between tools. XLIFF files contain source and target text segments with metadata. Most CAT tools and TMS platforms support XLIFF import and export.
TMX (Translation Memory eXchange)
The standard format for exchanging translation memory data between tools. A TMX file contains paired source-target segments that can be imported into any CAT tool.
Parallel Corpus
A collection of texts in two (or more) languages that are translations of each other, aligned at the sentence level. Parallel corpora are the training data for MT systems. The quality and size of available parallel corpora directly affects MT quality for a given language pair.
Tokenization
The process of breaking text into units (tokens) for processing by an MT system or LLM. Tokenization strategies differ between models and affect both cost (charged per token) and quality (especially for morphologically rich languages).
Fuzzy Match
A TM match that is similar but not identical to the current source segment, typically scored as a percentage (70-99%). Fuzzy matches save translator time by providing a close starting point that needs only partial editing.
Industry and Business Terms
LSP (Language Service Provider)
A company that provides translation, localization, interpretation, and related language services. LSPs range from small specialized agencies to large multinational companies. They typically manage translator networks, use TMS platforms, and handle project management.
SLV (Source Language Volume)
The word count of the source text — the standard basis for pricing in the translation industry.
TEP (Translation, Editing, Proofreading)
The standard three-step quality process: (1) a translator produces the first translation, (2) an editor reviews for accuracy and fluency, (3) a proofreader checks for final errors. TEP is the minimum quality standard for professional translation.
DTP (Desktop Publishing)
The process of recreating document layouts in the target language. Necessary for brochures, manuals, presentations, and any content with complex formatting. DTP is particularly important when target text is significantly longer or shorter than the source (German text is roughly 30% longer than English).
CAL (Computer-Assisted Localization)
A broader term encompassing CAT tools, MT integration, QA automation, and localization engineering. Less commonly used than CAT but technically more comprehensive.
FAQ
What is the difference between a CAT tool and machine translation? A CAT tool is a workbench for human translators that provides TM, glossary, and QA features. Machine translation generates translations automatically without a human. Modern workflows often integrate MT into CAT tools — the MT provides a suggestion, and the human translator edits it.
Do I need a TMS if I already have a CAT tool? A CAT tool handles the linguistic work; a TMS handles the project workflow. Small teams (1-3 translators) can manage with a CAT tool alone. Organizations with multiple languages, many translators, and continuous content streams need a TMS for project management, routing, and quality tracking.
What is the difference between TM and MT? TM (Translation Memory) stores and retrieves previous human translations. MT (Machine Translation) generates new translations using AI. They complement each other: TM ensures consistency with past work, while MT handles new content that has no TM match.
Is BLEU or COMET a better quality metric? COMET correlates better with human quality judgments. BLEU is simpler, faster, and more universally understood. Use both: BLEU for backward compatibility and quick checks, COMET for more reliable quality measurement.
What does a CAT tool cost? Prices range from free (MateCat, OmegaT) to $50-$300/month for professional tools (memoQ, Trados). Cloud-based platforms like Smartcat and Phrase offer tiered pricing starting from free tiers for individuals up to enterprise plans at $1,000+/month.
Sources: