Translation Industry Glossary

Agentic translation is when translating a document is handled not by one model call but by several specialized agents, each responsible for one part of the task — tone, terminology, cultural adaptation, review.

AI orchestration is a control layer that decides which model or service handles a given request, what context to add, and which checks the result passes through, instead of sending everything through the same model on the same path.

BLEU is the oldest widely used machine translation metric: it scores how much a translation overlaps with a reference translation, word by word and phrase by phrase.

Software that helps translators work more efficiently by providing translation memory, terminology management, and quality assurance features.

COMET is a neural network trained on human ratings that scores translation quality more accurately than classic metrics like BLEU, including reference-free scoring.

Data localization is a statutory requirement that certain categories of data be stored, or first collected, within a country's borders, and that transfers out of it follow a defined procedure.

Data residency is the physical or geographic location where data is stored at rest, stated as a fact about infrastructure rather than about who may access the data.

Data sovereignty is the principle that data is subject to the laws of the jurisdiction with authority over it, which is not always the jurisdiction where it is stored.

A translation memory match where the source text is similar but not identical to a stored segment, typically measured as a percentage of similarity.

GEMBA is a way to score translation quality by asking a large language model directly, instead of relying on a metric purpose-trained for the task.

ISO 11669 is a standard listing the translation project parameters that need to be explicitly agreed before work starts: audience, purpose, terminology requirements, and delivery format.

ISO 18587 is a standard setting requirements for machine translation post-editing: what counts as full versus light post-editing, and what competencies a post-editor needs.

ISO 5060:2024 is an international standard describing the translation quality evaluation process — pre-evaluation, evaluation, and post-evaluation — with an error typology harmonized with MQM.

LLM-as-a-judge is an evaluation method in which a large language model scores another model's output against a written rubric, instead of comparing it to a reference answer.

LQA is the stage where a linguist reviews a finished translation against the source and against agreed criteria for style, terminology, and local conventions.

A company that offers professional translation, localization, interpretation, and related linguistic services to clients across industries.

MQM is an industry-standard framework for evaluating translation quality that categorizes errors by type (accuracy, fluency, terminology, style) and severity (critical, major, minor).

MQM Core is an updated version of the translation error classification that merges MQM and DQF into a single taxonomy aligned with the ISO 5060 standard.

The process of human translators reviewing and correcting machine translation output to achieve publication-quality translations.

MTQE scores machine translation quality with a model that looks only at the source-target pair and needs no pre-made reference translation to compare against.

NMT is machine translation built on neural networks that learn to translate whole sentences at once, instead of assembling a translation from individual phrases the way older statistical systems did.

QE (Quality Estimation) is the umbrella term for automated translation quality scoring by a model, covering both reference-based and reference-free approaches.

RAG is when a model searches an external database for relevant snippets before responding and adds them to its context, instead of relying only on what it memorized during training.

A discrete unit of text that is translated as a single entity, typically a sentence or paragraph, forming the basic building block of translation memory.

The source language is the original language of content being translated, while the target language is the language into which content is translated.

The systematic process of identifying, storing, and managing specialized terms and their translations to ensure consistency across all content.

An XML-based standard format for exchanging translation memory data between different CAT tools and translation management systems.

The systematic process of evaluating translation quality using standardized metrics, frameworks, and error categorization to ensure translations meet specified quality requirements.

A database that stores previously translated segments for reuse, ensuring consistency and reducing translation costs through leverage of existing work.

An XML-based format standardized by OASIS for exchanging localizable data and its translations between tools in the localization workflow.