Localization & language-technology terminology

Rendering meaning from a source language into a target language. A translation brief still needs audience, purpose, subject matter, format, and quality expectations; a word-for-word substitution is not the professional definition.

Adapting a product, service, or content for a specific locale. It can include translation, formats, layout, imagery, legal requirements, product behavior, testing, and release work.

Designing and engineering a product so it can support different languages, scripts, regions, and cultural conventions without one-off rewrites. It happens before and alongside localization.

The broader business and operating system for taking a product across markets. Organizations use the word differently, but it commonly spans internationalization, localization, market strategy, and global operations.

A defined combination of language and regional or cultural conventions used by a product. A locale is not always the same thing as a country, and a language can have several market-specific locales.

The language of the material being translated or adapted. The source may itself be a translation, so source provenance and authority still matter.

The language into which material is translated. Professional briefs normally specify the market or variety as well, such as French for Canada rather than only French.

A source-to-target combination, such as English to Amharic. Direction matters because comprehension, writing ability, specialization, and market knowledge are not automatically symmetrical.

Recreating the intended effect of high-visibility content for a target market rather than following source syntax closely. It usually requires a creative brief, options, rationale, and market review.

Changing language or content to fit a new audience, channel, constraint, or culture. The term is broad; a scope should say whether it includes rewriting, cultural change, timing, legal review, or only language changes.

Localizing marketing, interface, or editorial copy while preserving voice, intent, and channel constraints. It sits between straightforward translation and fully open-ended transcreation.

Reviewing and adapting references, symbols, narratives, visuals, interaction patterns, or assumptions that may not work in the target culture. It is not a license to stereotype a market.

Translating target content back into the source language, usually by an independent linguist, to expose meaning shifts. It is a diagnostic tool, not proof that the target text is natural or effective.

Resolving differences between translations, reviews, back translations, or evaluators into one controlled decision. The record should capture the final choice and its rationale.

Representing text from one writing system in another, such as converting a name from Ethiopic or Arabic script into Latin characters. It does not translate the meaning.

Converting speech or other audible content into written text. A specification should define verbatim level, speaker labels, timestamps, non-speech sounds, language variants, and treatment of unclear audio.

A language professional who produces target-language meaning from source-language material. Specialization, writing ability, research, tools, and target-market knowledge all affect fit.

A practitioner who adapts content or products for a locale and may handle files, context, product constraints, terminology, and QA in addition to translation.

A second qualified person who evaluates or edits work against defined requirements. Reviewer independence, authority, and whether changes are mandatory must be explicit.

A person who checks near-final content for errors such as spelling, punctuation, formatting, or omissions. In some markets the title is used more broadly, so the scope must define the level of intervention.

A target-market writer who recreates persuasive or creative content from a brief and source intent. The work often includes multiple options and commentary rather than one literal target.

A specialist who prepares, transforms, validates, and reintegrates localizable content and language assets. The role often covers file formats, automation, builds, fonts, scripts, and technical QA.

The person coordinating scope, schedules, assets, languages, suppliers, reviews, risks, and delivery. Project management is a delivery function, not merely forwarding files.

A person with domain expertise who resolves technical or professional questions. An SME may know the subject without being qualified to judge target-language writing, so responsibilities should be separated.

A database of source and target segments from previous work. It suggests reusable matches but does not determine whether an old translation is correct for the new context.

A unit presented for translation, often a sentence, heading, or software string. Segmentation is a tool decision; meaning may cross segment boundaries.

A translation-memory match whose stored source is identical under the tool's matching rules. Context, metadata, product changes, and previous quality can still make review necessary.

A partial translation-memory match expressed as a percentage or band. Percentages are tool-specific similarity signals, not universal measures of editing effort or quality.

A structured terminology database containing approved concepts, terms, languages, status, definitions, context, and sometimes forbidden forms. It is more controlled than a simple word list.

A list of important words or expressions with preferred translations or explanations. Glossaries vary from simple spreadsheets to governed multilingual resources.

Rules for voice, grammar, punctuation, formatting, inclusivity, product naming, and locale-specific decisions. It should include examples and an owner, not only abstract principles.

A collection of language data used for research, search, training, or evaluation. Its value depends on provenance, representativeness, annotation, licensing, and fitness for the task.

Software that helps people translate using bilingual editors, translation memory, terminology, QA checks, and file filters. CAT does not mean the translation is automatically machine-generated.

A platform for managing localization projects, workflows, users, content, language assets, integrations, and reporting. Products differ widely, so TMS is a category rather than one standard feature set.

Software-generated translation. The label covers different architectures and providers; output quality varies by language, domain, prompt or settings, context, and input quality.

Machine translation based on neural-network models. It improved fluency over many earlier systems but can still omit, invent, mistranslate, or flatten terminology and style.

Translation generated or revised by a large language model, often with instructions or context. Fluency can be high while faithfulness, terminology, consistency, privacy, and reproducibility remain open risks.

An integration that moves content or metadata between systems, such as a repository, CMS, design tool, and TMS. A connector's behavior around deletion, branching, status, and retries must be tested.

A defined interface through which software requests data or actions from another system. API access does not by itself provide a safe workflow, permissions model, or idempotent behavior.

A protocol for exposing tools, resources, and prompts to compatible AI applications. MCP can connect an agent to localization operations; it does not guarantee translation quality or authorize uncontrolled write access.

A working file that keeps source and target segments plus tags, status, and metadata together. It is useful for review and exchange but is not always the final deliverable.

An OASIS XML standard for exchanging localizable content and related metadata between tools. Version and implementation support matter; a file labeled XLIFF may use custom extensions or lose data in round trips.

An XML exchange format commonly used to move translation-memory content between systems. Export a sample and test import because metadata and segmentation do not always transfer identically.

An XML standard family for exchanging structured terminology. A spreadsheet glossary is not automatically equivalent to a governed TBX termbase.

A text-based gettext catalog containing source messages, translations, comments, and status. It is common in open-source and software localization workflows.

A collection of locale-specific messages or assets loaded by an application. The exact format and fallback behavior depend on the framework or platform.

A stable identifier used by software to look up a localized message. Keys should not encode fragile English wording or be silently reused for a different meaning.

A runtime value inserted into a message, such as a name, count, date, or link. Translators need its type, examples, constraints, and safe placement rules.

Moving user-facing text out of source code into managed message resources. It is necessary but not sufficient for internationalization.

Building a sentence from separately translated fragments. This often breaks grammar and word order; use complete structured messages with explicit variables and selection logic instead.

The rule used when a requested locale or message is unavailable. Fallback must be deterministic and observable so users do not receive silent language mixtures or unsafe defaults.

Transforming source strings into artificial variants that expose expansion, hard-coded text, encoding, directionality, or layout problems before real translation.

The change in length or geometry that occurs after translation. Design for variable content rather than relying on one universal expansion percentage.

A writing direction used by scripts such as Arabic and Hebrew. Supporting RTL includes layout, navigation, icons, text direction, mixed content, and input—not simply mirroring every screen.

Text containing both left-to-right and right-to-left runs. Correct display depends on the Unicode Bidirectional Algorithm plus appropriate markup and isolation.

A writing system whose display may require shaping, reordering, combining marks, or context-sensitive behavior. The term describes technical processing, not the difficulty or value of a language.

The standard that assigns characters code points and defines properties and algorithms used to process text. Using UTF-8 does not automatically solve fonts, normalization, segmentation, input, bidi, or locale behavior.

A numeric value in the Unicode codespace. One visible user-perceived character may contain several code points.

An approximation of a user-perceived character made from one or more code points. Cursor movement, deletion, and length limits should not split clusters.

A visual shape used to render one or more characters. Unicode encodes characters, while fonts and shaping systems select and position glyphs.

Converting canonically equivalent Unicode sequences into a consistent form. Apply it deliberately at defined boundaries; indiscriminate normalization can damage identifiers or source evidence.

Finding boundaries between grapheme clusters, words, sentences, or lines. Rules vary by purpose and language; splitting on spaces is not a universal word algorithm.

Locale-sensitive comparison, sorting, and grouping of text. Binary or code-point order rarely matches user expectations for an alphabetical list.

The Unicode Common Locale Data Repository: structured locale data and specifications for formats, names, plural rules, collation, units, and other conventions used by modern software.

A controlled point after which source changes are restricted for a localization cycle. It reduces churn but should not prevent urgent corrections; late changes need an explicit delta path.

The transfer of source assets, context, instructions, references, access, deadlines, and acceptance criteria into production. A file alone is not a complete handoff.

A staged workflow in which translation, bilingual editing, and final proofreading are separate activities. Labels and independence vary, so the scope should define what each pass checks.

An operating model that moves content changes through localization alongside frequent product releases. It requires source discipline, automation, state, review rules, QA, and exception handling.

Maintaining parallel content or code versions, such as release and development branches. Localization systems need rules for change propagation, conflicts, backports, and obsolete strings.

A documented adjustment to agreed scope, price, schedule, or assumptions. It protects both buyer and supplier from silently absorbing material changes.

A defined decision point with required evidence and an accountable owner. Passing a gate means agreed criteria were met; it does not mean the product is free of all defects.

Reviewing localized content in the product, design, document, or media where users encounter it. It reveals meaning, layout, interaction, and visual issues that a bilingual grid can hide.

Process-focused work designed to give confidence that requirements can be met. In localization, the label is often used loosely; distinguish preventive QA from checking a finished output.

Monitoring or checking work and process performance against requirements. QC catches problems during or after production; it does not replace a quality system.

Evaluation of localized language against defined dimensions, instructions, and severity rules. It may be performed in a bilingual tool, the product, or both.

Testing whether the localized product works: navigation, input, formats, links, variables, layout, fonts, directionality, and locale behavior, not only wording.

A framework for defining and categorizing translation-quality issues. Teams should tailor the taxonomy, severities, sampling, and thresholds to their content and decisions.

A controlled set of issue categories such as accuracy, terminology, style, locale convention, or formatting. Categories should help action and analysis rather than reward over-tagging.

The impact level assigned to an issue, typically based on meaning, usability, risk, reach, and recoverability. Severity is not the same as how much a reviewer dislikes wording.

The observable conditions a deliverable must meet to be accepted. They should name scope, tolerances, evidence, approver, and what happens when criteria are not met.

Human work on machine-translated output to meet a defined target quality. The post-editor must be able to retranslate or reject output when correction is not efficient or safe.

A limited correction pass intended for a specified use, often focused on intelligibility and serious errors. It is not a universal quality level and is inappropriate when polished or high-risk text is required.

Editing machine output toward requirements comparable to a professionally translated deliverable. Actual effort depends on source, language, system, domain, context, and error density.

A person assigned a real decision, correction, or approval role in an automated workflow. The phrase is meaningless unless authority, evidence, competence, time, and escalation are defined.

Output that introduces unsupported content or claims. In translation it can appear as invented details, omitted constraints, or fluent text that is not faithful to the source.

Providing a model with authoritative context or requiring claims to be tied to controlled sources. Grounding can reduce some errors but does not prove a translation is correct.

A pattern in which a system retrieves documents or data and supplies them to a generative model. Multilingual retrieval quality, language mismatch, permissions, and citation behavior all need evaluation.

A controlled set of inputs and expected evidence used to assess a system. It should represent languages, tasks, risk, edge cases, and real deployment conditions.

Adding labels or structured information to text, audio, images, or video. Annotation quality depends on the schema, instructions, tools, annotator fit, calibration, and adjudication.

The operational rulebook for label definitions, boundaries, examples, edge cases, escalation, and version history. It is part of the dataset, not administrative overhead.

A structured model of concepts and relationships used to organize or label information. Annotation projects often use lighter taxonomies; the correct complexity depends on the task.

A pilot round in which annotators and reviewers apply the rules, compare disagreements, and refine the guidelines before scaling production.

A measure of consistency among annotators under a defined method. High agreement can coexist with a bad schema, and low agreement may reveal ambiguity rather than poor workers.

A controlled decision process for resolving annotation or evaluation disagreements. Adjudicated outcomes and reasons should feed guideline updates.

A curated set of examples with trusted labels or judgments used for calibration, quality monitoring, or evaluation. Gold labels still need provenance, review, and maintenance.

A record of where data came from, how it was collected, transformed, labeled, licensed, and reviewed. Provenance supports legal, ethical, quality, and reproducibility decisions.

Technology that converts speech into text. Performance varies with language, accent, noise, channel, overlap, names, and domain.

Technology that generates spoken audio from text. Evaluation should cover intelligibility, naturalness, pronunciation, prosody, voice identity, and product context.

A common ASR metric based on substitutions, deletions, and insertions relative to a reference transcript. Tokenization and language structure affect interpretation, so WER is not directly comparable everywhere.

An ASR error metric calculated at character level. It can be useful when word segmentation is unstable, but it still does not describe semantic or product impact.

Determining who spoke when in an audio stream. It is distinct from identifying a person's real-world identity.

Assigning transcript content to speaker labels or identities defined by the project. Overlap and uncertain speakers need explicit conventions.

An average subjective rating used for perceived speech quality or naturalness under a defined test. Scale, question, listening setup, language, and rater population determine what the score means.

A resource connecting words or names to pronunciations, phonemes, or variants. It is especially useful for names, brands, abbreviations, and domain terms.

Timed text that generally conveys dialogue for viewers who can hear the audio but may not understand its language. Platform and market conventions vary.

Synchronized text representing speech and meaningful non-speech audio for people who cannot rely on the soundtrack. Captions commonly include speaker identification and sound cues.

Subtitle files that include dialogue plus accessibility information such as speaker labels, sound effects, and music cues. The exact term and convention vary by market.

Captions delivered as a selectable track that viewers can turn on or off. Player and format support are part of accessibility quality.

Captions permanently visible in the video image. They cannot be disabled and require visual-safe placement during production.

Replacing original dialogue with recorded target-language performance synchronized to the program. It includes adaptation, casting, direction, recording, editing, mixing, and QC.

Target-language speech recorded over or alongside original audio, often with less lip synchronization than dubbing. Narration, corporate voice-over, and unscripted voice-over use different conventions.

Narration of important visual information for people who cannot fully access the picture. It is written and mixed around existing dialogue and sound, and may require extended-description techniques.

Setting subtitle or caption in-times, out-times, and divisions against the media. Good spotting considers speech, shot changes, reading time, rhythm, and platform rules.

One timed subtitle or caption event containing text and timing metadata. A cue can contain one or more lines.

A time reference attached to media frames or samples. Frame rate, start time, drop-frame convention, and program version must match throughout the workflow.

The rate at which timed text asks a viewer to read, usually expressed in characters or words per second. It is one constraint among meaning, timing, line breaks, audience, and language.

A reading-speed measure dividing displayed characters by cue duration. Counting rules and acceptable ranges differ by language, audience, and delivery specification.

A line-length limit used in timed text. It affects readability and layout but should not force unnatural segmentation or meaning loss.

A cut or transition between shots. Timed text is often placed or adjusted around cuts to reduce visual distraction, according to the delivery specification.

A subtitle intended to convey essential on-screen text or foreign dialogue even when the main subtitle track is off. Selection and delivery rules must be defined by the platform.

A company that supplies and manages one or more language services. The label does not reveal actual specialization, ownership, worker model, or quality system.

A supplier focused on one language or locale, often with deep local operations. It may provide several services within that market.

A supplier coordinating work across multiple languages, often through internal teams and subcontractors. Buyers should understand which work is performed where.

The lowest charge for accepting and administering a project, regardless of a small word or minute count. It covers setup, communication, file handling, invoicing, and opportunity cost.

A rate applied to the count of source words. Counting rules, repetitions, tags, minimums, review layers, and non-word tasks must still be specified.

A rate applied to the runtime of delivered media. It does not describe whether transcription, spotting, translation, SDH, recuts, or QC are included.

A visible price or scope adjustment for compressed turnaround or out-of-hours capacity. A rush fee does not make an impossible schedule safe.

An approved third-party expense billed separately, such as studio, voice talent, secure transfer, travel, or platform charges. The estimate should state markup and approval rules.