A working dictionary of the terms, frameworks, and vendor concepts for how AI translates, generates, evaluates, and governs multilingual content.
About this glossary
The language industry is inventing vocabulary faster than anyone can sensibly standardize it.
Marketing teams coin shiny terms. Researchers publish benchmarks. Practitioners test what actually survives contact with workflows, budgets, reviewers, and regulators.
Sometimes they all mean the same thing. Sometimes they don’t. That’s where the trouble starts.
This glossary is a working reference for the concepts now shaping multilingual AI: the practical terms, the research vocabulary, and the vendor frameworks worth understanding before they end up in an RFP, a roadmap, or (heaven help us!) a “transformational AI strategy” deck.
Table of contents
How to use it
Entries are grouped into eight categories reflecting where each term lives in practice. Each entry is a single paragraph combining a quotable definition and why it matters. Where a concept has multiple parents or spellings, the most-cited form is used as the headword.
Feedback and additions
This is still a working version. Terms are moving fast. Some will become standard. Some will quietly disappear. Some deserve to disappear faster.
Corrections, challenges, and nominations for the next version are welcome.
1. MT & LLM Foundations
Adaptive Machine Translation
An MT approach where the engine learns from a translator’s edits during the same session, rather than waiting for a separate retraining cycle. Popularized by Lilt and later adopted across enterprise TMS vendors, adaptive MT shortens the loop between human correction and machine output. Useful? Very. Magic? No. It depends on the quality of the edits, the domain, and whether the workflow actually captures learning instead of just talking about it.
Automatic Post-Editing (APE)
A dedicated model, often a purpose-tuned LLM, that improves raw MT output before a human sees it. APE can fix terminology, tone, gender, formatting, and other predictable issues. Its real value is economic: when paired with quality estimation, it helps decide which segments can move forward automatically and which still need human attention.
BYO-LLM / BYO-Engine
A vendor capability that lets customers connect their own MT engine, LLM, or model endpoint into a localization workflow. The platform wraps translation memory, glossaries, quality estimation, and CAT tooling around the customer’s chosen engine. This matters because enterprises increasingly want control over data, privacy, cost, and model choice. Vendor lock-in is expensive. So is pretending it isn’t.
CAMT / Context-Aware Machine Translation
Machine translation that uses surrounding sentences, document metadata, or dialogue history instead of translating each segment in isolation. Sentence-by-sentence MT is where many quality complaints are born: pronouns go vague, tone drifts, terminology wobbles. CAMT is now table stakes for serious LLM-based translation systems.
GEMBA (GPT Estimation Metric Based Assessment)
An LLM-based translation quality metric introduced by Kocmi and Federmann at EAMT 2023 that uses GPT models to score translations without requiring a human reference. GEMBA was the first demonstration that a general-purpose LLM could achieve accuracy on system-level translation ranking, previously the domain of dedicated metrics like COMET. It opened the door. What came through that door is GEMBA-MQM.
KGMT (Knowledge-Graph Mediated Translation)
An MT approach that uses a domain knowledge graph to constrain translation decisions around entities, relationships, and meanings. Apple ML’s XC-Translate benchmark is a related academic effort focused on cross-cultural entity translation. This is one answer to hallucinated entity translation. If “Prague” and “Praha” need to map to the same concept, a knowledge graph can make that explicit rather than hoping the model remembers.
L2T Prompting (Language-to-Thought)
A prompting technique where an LLM first reasons in an abstract “thought” representation before generating in the target language. It is designed to reduce language-dependent reasoning gaps, sometimes called the Translation Effect or Language Binding Effect. The practical point: LLMs do not always “think the same thing” across languages. Awkward, but important. This is still an emerging research framing, not yet standardized.
MAPS (Multi-Aspect Prompting and Selection)
A prompting framework that generates several candidate translations from different angles — literal, idiomatic, terminology-focused, and so on — then selects the strongest output. It turns a simple insight into a workflow: one prompt is often not enough.
MTPE (Machine Translation Post-Editing)
The workflow where a human linguist edits raw MT output to a defined quality level. ISO 18587:2017 formalized light and full post-editing. MTPE is still the unit many localization services are priced around, even as APE and quality estimation absorb more of the work. It is not dead. It is being repositioned.
Multi-Engine Routing
The practice of routing each translation request to the best engine or model for that content type, language pair, quality target, latency, and cost. The old question was “Which engine is best?” The better question is “Which engine is best for this job, right now?”
Quality Estimation (QE / MTQE)
A model that predicts translation quality without a reference translation. Long-standing research area; brought into mainstream localization by TAUS. Usually it produces a score that helps decide whether a segment is auto-approved, sent to APE, routed to a human, or rejected. QE is the gatekeeper of modern localization economics. Without it, “AI-first localization” is mostly vibes.
Red Teaming (AI)
Adversarial testing of AI systems to find safety failures, policy bypasses, and data leakage vectors before deployment. In regulated or multilingual contexts, red teaming should be language-conditional: a jailbreak that works in English may work differently in Arabic or Czech.
Token / Token Throughput
A token is the basic unit of LLM input and output, roughly a word fragment. Token throughput is how many tokens a system can process per second, determining how many users or segments a workflow can serve simultaneously. Enterprise localization buyers increasingly encounter per-token pricing; understanding tokens is how you model cost at scale.
Translation Memory (TM)
A database of previously approved source-target segment pairs. TM is old, useful, and very much not obsolete. In LLM workflows, it becomes part of the retrieval layer: a source of approved terminology, phrasing, and consistency that the model should respect rather than creatively ignore.
2. Evaluation & Quality
AutoLQA
Automated localization quality assurance using AI, often mapped to MQM categories, to score translations at scale and triage what needs human review. AutoLQA matters because full human LQA everywhere is expensive. The trick is knowing when automation is good enough — and when it is just fast nonsense with a dashboard.
BLEU / COMET / TER / chrF / METEOR
The main family of automatic MT metrics. BLEU and TER look at lexical overlap. chrF works at character level. METEOR adds synonymy. COMET is a learned neural metric trained on human judgments. BLEU is famous but limited. COMET is currently more useful in many serious MT evaluations. None of them tells the whole story.
Cutscore / Passing Threshold (PT)
The minimum quality score a translation must exceed to be considered acceptable. Defined in the TRANQUALITY TQE tutorial and formalized as part of the ASTM/multi-range quality measurement work. A score without a cutscore is just a number looking for a meeting. Cutscores turn evaluation into a decision: publish, review, route, or reject.
Eval-as-a-Service (EaaS)
Translation-quality evaluation offered as an on-demand service or API. As companies use more engines and LLMs, they increasingly want a neutral evaluator rather than trusting each vendor’s self-reported quality. Sensible. Vendors grading their own homework is not exactly a new risk.
GEMBA-MQM
A follow-on framework by Kocmi and Federmann, published at WMT 2023, that extends GEMBA from scoring into error-span detection. A hybrid approach that prompts an LLM to annotate MQM-style errors and then aggregates weighted errors into a score. It gives MQM-like granularity at a scale human programs cannot easily match. Useful for triage. Still needs calibration.
LLM-as-a-Judge
The pattern of using one LLM to evaluate another AI output for translation quality, factual accuracy, tone, or brand fit. It makes evaluation cheaper and more scalable. It also introduces bias, prompt sensitivity, and false confidence. In other words: powerful, but not a free lunch.
MQM (Multidimensional Quality Metrics)
An error-typology framework for scoring translation quality across accuracy, fluency, terminology, style, locale conventions, and other dimensions. MQM is the closest thing the industry has to a shared quality vocabulary. Many serious vendor SLAs and audit conversations start here.
QPS (Quality Performance Score)
Phrase’s proprietary translation quality score, used in its Language AI stack to help route segments for auto-approval or review. QPS is one example of a wider pattern: every major TMS now has a branded quality score. Do not assume these scores are comparable across vendors. They usually are not.
Stripped MQM
A simplified subset of MQM for high-volume, lower-risk scoring. Full MQM is expensive. Stripped MQM is what teams use when they need directional quality signals without turning every review into a PhD defense.
TransEvalnia
A recent LLM-based translation evaluation framework that reasons over MQM error types and produces structured, explainable assessments. The interesting part is not only the score, but the explanation. Regulated content needs auditability, not just a green checkmark.
TQI (Translation Quality Index)
Transifex’s proprietary quality score combining context, brand profile, and terminology signals. Like QPS and other vendor scores, TQI is useful inside its own platform but should not be treated as a universal metric.
TTE (Time-To-Edit)
The average time a human spends editing MT output to acceptable quality. It is refreshingly practical because it measures what actually costs money: human time. Procurement tends to understand that.
3. Governance, Risk & Trust
AI Governance / Risk-Based AI Governance
The policies and controls an enterprise uses to decide which AI translation workflows are acceptable for which content. Legal copy, medical content, support chat, and SEO pages should not all run through the same risk model. If they do, that is not governance. That is optimism.
AI Slop
High-volume, low-effort AI-generated content that degrades channel quality, such as marketing spam, synthetic filler, SEO junk. Increasingly a brand risk: audiences and AI systems both develop tolerance to it. In multilingual contexts, AI slop in non-English markets often goes undetected longer, which compounds the damage.
Do-Not-Claim List
A list of claims an AI system must never make on behalf of the brand. Especially useful for legal, regulated, medical, financial, or reputationally sensitive content. LLMs are probabilistic. A Do-Not-Claim list is deterministic. That is the point.
Guardrails
Rules, models, or filters that constrain AI output so it respects terminology, brand, safety, and locale policies. Guardrails do not make AI perfect. They make failure less chaotic, which is already a meaningful improvement.
ISO 5060 / ISO 42001
ISO 5060:2024 (Translation and interpreting — Evaluation of translation output — General guidance) covers translation services quality evaluation. ISO 42001:2023 covers AI management systems. Expect both to show up more often in procurement, audits, and enterprise AI governance. Boring? Maybe. Important? Definitely.
Localization Bias
The systematic gap where non-primary-language experiences perform worse than the primary-language experience on conversion, comprehension, satisfaction, safety, or other metrics. This reframes localization from “translation quality” to product equity. That is a much bigger conversation.
Locale Liability Ledger
A register of legal, regulatory, and reputational exposure created by AI-generated content in each locale. It gives legal, localization, and marketing teams a shared object to manage risk. Less hand-waving, more accountability.
Semantic Debt Ledger
A register of unresolved terminology conflicts, ontology drift, retired concepts, and inconsistent meanings. Semantic debt behaves like technical debt: invisible until something breaks downstream. Then suddenly everyone cares.
Tiered Workflows / Content Tiering
Mapping content types to explicit quality tiers with different levels of automation, review, SLA, and cost. This is how AI localization becomes operational. The question is not “Can AI translate?” The question is “For which tier, with which controls?”
4. Workflow & Orchestration
Agentic Localization
Localization workflows where AI agents plan, execute, and check tasks by invoking tools such as MT engines, glossaries, QE, and human review. This could reshape the localization supply chain. It could also create new operational messes. Both can be true.
Agent Ops
The discipline of managing AI agents in production: versioning prompts, monitoring behavior, controlling costs, and running post-mortems. Agents do not remove operations. They create a new kind.
AI Orchestration Layer
Middleware that routes translation and content-generation work across engines, LLMs, and human workflows based on content type, quality target, latency, and cost. Engine choice is commoditizing. Routing logic is where the interesting differentiation is moving.
Collapsing the Content Supply Chain
Reducing the steps between source content creation and multilingual publication. In the more radical version, teams generate in multiple languages in parallel rather than translating English after the fact. That changes what LSPs sell, what marketing owns, and what “source” even means.
Continuous Improvement Loop
A workflow where every human edit improves the engine, glossary, style guide, or prompt context. Without the loop, quality plateaus. With it, quality can compound. Assuming, of course, someone actually maintains the loop.
Distillation
Compressing a large model’s behavior into a smaller, cheaper model. Relevant when enterprise teams need to run localization or QE models on-premises or at lower cost without retraining from scratch.
Human-in-the-Loop / Human-at-the-Core
Workflow patterns that keep humans involved as reviewers, decision-makers, or owners of the quality bar. “Human-in-the-loop” can mean a person clicks approve at the end. “Human-at-the-core” implies more authority. The distinction matters.
Locale Operating Modes
A framework where each locale has an explicit mode: translation-first, generation-first, or hybrid. Not every market deserves the same workflow. Pretending otherwise is how teams overspend in one locale and under-serve another.
Model Context Protocol (MCP)
An open protocol introduced by Anthropic to standardize how LLMs receive tools, resources, and context from external systems. For localization, MCP could make TMSs, glossaries, QE systems, and agents easier to connect. “USB-C for AI agents” is the catchy version. The less catchy version is: integration economics may change.
Parallel Multilingual Generation
Generating content directly in multiple target languages from the same brief or knowledge base, instead of creating English first and translating later. It is powerful. It also forces the organization to maintain a real knowledge layer, because otherwise every language will drift in its own charming direction.
Shadow Localization
Localization happening outside approved workflows: engineers using ChatGPT for UI strings, marketers running one-off DeepL passes, regional teams improvising translations. Fast, cheap, and risky. Like Shadow IT, but multilingual.
5. Strategy, Ownership & Roles
Cultural Intelligence
The capability that helps content, and the AI systems producing it, behave with local cultural sense, not just grammatical accuracy. RWS uses the term for a layer of in-market experts, domain specialists, and multilingual data that keeps enterprise AI culturally fluent and context-aware. As a concept, cultural intelligence promises to push localization beyond language into local values, emotional nuance, and the messy human context AI does not reliably understand on its own.
Global-First Architecture
A product and content architecture that treats every locale as a first-class output from the start. It is the structural answer to Localization Bias. If global is bolted on at the end, the user can usually tell.
Human as a Feature (HaaF) / Translation as a Feature (TaaF)
Framings where human expertise or translation capability becomes a differentiated product feature, not a hidden production cost. This changes pricing, positioning, and go-to-market. It also forces vendors to explain what the human actually adds.
Language Intelligence
A category framing that positions language services as an intelligence layer built on knowledge, terminology, ontology, and quality data. It is broader than “translation services” and speaks to product, marketing, and data buyers. A useful rename, if the substance is there.
Market DNA
The cultural, regulatory, competitive, and channel-specific factors that define how content should behave in a market. LLM-generated content only works when conditioned on this reality. Otherwise it produces fluent content with no market fit. Lovely sentences. Wrong market.
Multilingual AI
The capability stack that uses AI to translate, generate, evaluate, route, and govern content across languages. It spans MT, LLMs, orchestration, evaluation, and knowledge infrastructure. Multilingual AI is replacing localization in many buyer conversations because it names the operating system, not just the deliverable.
Shippability
A framing that asks whether a translation is ready to ship, not whether it is abstractly “correct.” Shippability combines quality, brand, compliance, locale fit, and user risk into one decision. Product teams understand this language.
6. AEO, GEO & AI Search
AEO / GEO (Answer Engine Optimization / Generative Engine Optimization)
The discipline of making content discoverable, quotable, and trustworthy for AI answer engines such as ChatGPT, Perplexity, Google AI Overviews, and Copilot. Think SEO, but for systems that summarize and cite instead of sending people neatly to your page. Multilingual AEO adds another layer: how do you get cited accurately in German, Japanese, Portuguese, and not just English?
Belief Provenance
The traceable chain showing how an AI system came to hold or repeat a claim: source, retrieval, prompt, model, and output. “Where did that come from?” is becoming an audit question. Belief provenance is the answer.
Citation Supply Chain
The chain of sources, citations, re-citations, and reformulations that causes a claim to show up in an AI answer. In AEO, visibility depends not only on your content, but on where it sits in the wider citation network.
Claim Volatility Index
A measure of how quickly factual claims change and therefore how aggressively content must be maintained. Pricing, regulation, product features, and legal claims decay quickly. Evergreen educational content less so. The index makes that maintenance decision explicit.
Contested-Citation Ratio (CCR)
The share of citations around a claim that contradict or dispute it. A high CCR means AI systems may hedge, surface competing claims, or cite your competitor alongside you. Annoying? Yes. Useful signal? Also yes.
Reverse Content Supply Chain
Publishing content in the formats, languages, and places most likely to be retrieved and cited by AI systems. Traditional content marketing asks, “What does the reader want?” Reverse content supply chain asks, “What will the model retrieve before the reader ever sees anything?”
Structured Intent Payload
A structured representation of user intent — goal, constraints, audience, tone, locale — passed into generation or translation workflows. It replaces guesswork with declared intent. That matters because LLMs are very good at filling gaps, including the gaps you did not want filled.
Truth Source
The authoritative artifact that AI systems should use for factual claims in a domain: a glossary, knowledge graph, canonical page, product database, or legal source. Without a truth source, contradictions multiply. With one, at least the system has something to obey.
7. Data & Knowledge Layer
Context Engineering
The discipline of designing, retrieving, and shaping the context an LLM sees at inference time. Detailed by Anthropic in Effective Context Engineering for AI Agents. This includes retrieval, memory, tool schemas, style guides, glossaries, and structured payloads. For multilingual AI, context engineering is where quality is made. The model matters. The context often matters more.
Context Envelope / Intent Context / Metadata Envelope
A structured bundle of context attached to each translation or generation request: source metadata, locale, tone, brand voice, glossary, TM matches, content type, and constraints. Passing “just the text” to an LLM leaves too much to chance. The envelope reduces that chance.
Knowledge Layer
The organizational layer holding terminology, glossaries, ontologies, style guides, brand voice, market knowledge, and approved facts. Whoever owns the knowledge layer has unusual leverage over multilingual AI. It is not glamorous. It is the moat.
Locality Effect / Mother Tongue Effect
Findings from the MultiLoKo benchmark showing that LLMs can answer differently depending on language and local relevance. The practical implication is uncomfortable: a great English answer does not guarantee a great Vietnamese, Czech, or Arabic answer. Testing has to be language-conditional.
Multilingual Benchmarks (MultiLoKo, MMLU-ProX, XC-Translate)
Benchmarks such as MultiLoKo, MMLU-ProX, and XC-Translate that test multilingual and cross-cultural performance more seriously than English-centric evaluations. Any enterprise buying multilingual AI in 2026 should know these names, or at least know that English-only benchmarks are not enough.
Source Quality Improvement (SQI)
AI-driven pre-editing of source content before translation: fixing ambiguity, terminology, typos, structure, and clarity. SQI may be the highest-leverage intervention in the workflow. One bad source sentence becomes twenty bad target sentences. Fix it upstream.
Translation Effect / Language Binding Effect
The phenomenon where LLMs produce different answers depending on the input language, even when the meaning is supposedly equivalent. This is one of the more underappreciated risks in multilingual AI. Translation can be correct and the answer can still drift.
8. Emerging Vendor Frameworks
AIDAterm
A multi-agent, terminology-constrained MT framework using coordinated agents such as terminology verifier, translator, and post-editor. It is an early proof that agentic orchestration could improve glossary adherence. That matters because terminology is still where many LLM translation workflows go to die.
Alconost.MT/Evaluate
A free LLM-based translation evaluation tool from Alconost, available at alconost.mt, that scores translations without heavy setup. It is a clear example of Eval-as-a-Service moving toward commodity access. What used to look like consulting can now look like a web form.
Beehyve / Algebras Guardrails, Confidence Score, Autofixes
Vendor-branded bundles for agentic translation QA: guardrails, confidence scores, and automated fixes. The pattern is clear: vendors are not really selling “the LLM.” They are selling the layer that makes the LLM less reckless.
ContentQuo
A SaaS platform for running structured LQA programs and, increasingly, for developing and benchmarking multilingual AI. Its AI Studio module adds multilingual prompt engineering, context engineering, and AI benchmarking. The shift is worth noting: LQA is moving upstream, from checking finished work to shaping how AI systems produce multilingual content in the first place.
Custom.MT Console
A platform for managing, comparing, and routing multiple MT and LLM engines across localization workflows. Custom.MT lets teams connect stock engines alongside custom-trained models through a single interface, with BLEU, WER, and COMET scores surfaced by language pair and domain. In plain English: it makes BYO-LLM and multi-engine routing something teams can operate without turning every model choice into an engineering project.
Intelligent Score / Language Guard / SmartContext / Intelligent Post-Editing (XTM)
XTM’s Advanced AI Pack bundles scoring, guardrails, context, and automated post-editing into one workflow layer. Intelligent Score predicts segment quality. Language Guard catches sensitive or biased language. SmartContext feeds TM matches into the prompt. Intelligent Post-Editing applies controlled LLM fixes with an audit trail. The practical point: fewer manual routing decisions, more explicit quality gates. Assuming, of course, the thresholds are set by someone who knows what “good enough” actually means.
Language Weaver Pro / Language Weaver Edge
RWS’s Language Weaver stack is moving from classic MT into heavier enterprise AI: a large Cohere-built model in Pro, plus Edge for on-premise and hybrid deployments. The practical point is control. Security-sensitive buyers still want quality, QE, post-editing, and Trados integration without sending every workflow into someone else’s cloud. Less shiny demo, more model-layer ownership.
Lokalise AI Orchestration
Lokalise’s AI stack routes translations across multiple LLMs using Thompson Sampling, then grounds output in brand context through Custom AI profiles. The useful bit is the feedback loop: model choice is based on real acceptance data, not someone’s favorite engine. Lokalise also exposes this through a native MCP server, which brings TMS workflows closer to Claude, ChatGPT, Cursor, and similar tools.
ModelFront
An independent quality prediction and automatic post-editing platform that operates inside existing TMS setups. ModelFront positions itself as a neutral quality layer: it checks MT output, flags what falls below threshold, corrects what it can, and escalates what it cannot. The useful bit is the monitoring. If automation is going to replace review hours, someone still needs evidence that quality has not slipped.
MosAIQ LQA
Argos Multilingual’s AI language quality assessment framework. MosAIQ LQA uses a two-agent AI review process to check terminology, style, grammar, and compliance risk, then scores content against MQM and customizable quality frameworks before human reviewers validate the flagged issues. It sits in the AutoLQA and Eval-as-a-Service pattern: broader coverage, clearer quality signals, human validation where it matters, and reporting that different teams can actually read without needing a translation-quality decoder ring.
Phrase’s Language Intelligence Platform
Phrase’s visual workflow engine routes content across MT engines, LLMs, QPS, human review, and tools like Auto Adapt. The important bit is not the diagram. It is the operating logic: context in, model choice, quality check, adaptation, then the right handoff. Phrase also exposes Orchestrator through a native MCP server, which makes TMS workflows available to AI agents through a standard protocol. A clear reference implementation of the AI orchestration layer pattern.
TAUS EPIC
TAUS’s commercial API product combining Quality Estimation and Automated Post-Editing into a continuous quality control loop for MT output. TAUS EPIC builds on the QE work TAUS helped develop since 2004 and packages it as an integration layer: QE scores segments, APE corrects the ones that need it, and the workflow routes the rest. Not glamorous. Quite useful. This is the kind of plumbing that makes “AI localization” operational rather than aspirational.
Corrections, additions, and challenges are welcome.