Machine translation & localisation

Every organisation that sells, supports or governs across language borders has to translate: contracts, manuals, websites, app UIs, product data, tickets and official letters. Translation agencies are slow and expensive, and free consumer tools leak confidential data. Current neural MT engines and LLMs produce near-professional drafts in seconds. Glossaries, translation memory and automatic quality estimation keep terminology consistent and show which segments still need a human. The open problem is no longer whether MT is good enough. It is deciding which content can go out unreviewed and which content still needs expert post-editing or a certified human translator.

Problem & value

• Today: SMBs either pay agencies per word with turnaround times of days, or staff paste text into free web translators. The second option is a GDPR and trade-secret problem. Enterprises run translation management systems (TMS) with agency workflows, but the volume of content (support, e-commerce, UI releases, internal communication) keeps growing faster than budgets.
• AI value: instant first drafts, consistent terminology across all EU languages, continuous localisation in software release pipelines, and multilingual customer service without native staff for every language. Linguists move from translating to post-editing and quality review.
• Quality level (sourced): in the WMT25 general MT task, professional annotators evaluated 60 systems on deliberately difficult, document-level test sets. Gemini 2.5 Pro was in the top cluster for 14 of 16 language pairs and was on par with or better than the human reference translation in 10 of them. The human references made the winning cluster in only 6 of 15 pairs [1]. The same study found that state-of-the-art systems still struggle with non-standard input, domain terminology and gender agreement, and that automatic metrics can rank systems very differently from humans [1].

Variants

• Gisting / internal communication: e-mails, chats, intranet pages; raw MT usually acceptable.
• Document translation with layout: contracts, offers, presentations and PDFs in which formatting is preserved (DeepL Document, Google Cloud Translation document API, Azure Document Translation).
• Website, shop & product-data localisation: CMS/PIM connectors, SEO keywords per market and locale-specific units and formats. Combine with LANG-001/LANG-005 for transcreation of marketing copy.
• Software & app localisation (developer variant): continuous localisation of string files in CI/CD via Lokalise, Crowdin, Phrase Strings or self-hosted Weblate, with placeholder and length checks.
• Regulated technical content: operating manuals and safety instructions in all EU languages for machinery (Machinery Regulation 2023/1230), and IFU/labelling for medical devices (MDR). MT plus mandatory expert review, fed from the CCMS (see LANG-022).
• Legal, financial and patent translation: post-editing by subject experts. Certified translations for courts and authorities (beglaubigte Übersetzungen) must be done by sworn human translators.
• Public sector: multilingual citizen information, translation of incoming applications and letters, and EU-funded eTranslation, which is free for eligible public administrations, SMEs, academia and NGOs in the EU [8].
• Image/scan translation: OCR + MT, or VLM-based translation of text in images (TranslateGemma accepts image input) [3][4].
• Private use: browser and phone translation. Firefox Translations runs Marian-based models locally in the browser [11]. A fully local setup is covered by PERS-005.
• Neighbours, not variants: speech-to-speech interpreting → SPEECH-003; subtitles → SPEECH-004; AI dubbing → SPEECH-007; Leichte Sprache (Easy Read) → LANG-010; sign language → LANG-030.

How it works

1. intake (file, CMS/PIM/repo connector, API call, chat message)
2. parsing & segmentation (Okapi/XLIFF filters, tags & placeholders protected; OCR/VLM for scans)
3. PII / secrecy check (route sensitive content to EU or on-prem engine; optional pseudonymisation DOC-018)
4. TM lookup (exact / fuzzy matches reused)
5. MT engine (NMT or LLM) with glossary, formality, style rules and document context
6. automatic QA (terminology, numbers, tags, length limits) + quality estimation (QE) per segment
7. routing: high QE
8. publish | medium
9. post-edit (human-in-the-loop) | low/critical
10. human translator
11. review in CAT tool / TMS, corrections update TM & termbase
12. delivery back to source system + quality reporting

• Engines: dedicated NMT (DeepL, Google, Microsoft, Amazon, Language Weaver, Systran) is fast and cheap per character. LLMs (Gemini, GPT, Claude, Mistral) are better at document context, tone, instructions ("keep the brand voice, use Sie-form") and rare language pairs, but are slower and need guards against omissions and additions. Many TMS platforms now route each job to the best engine for the language pair and domain.
• Quality estimation (reference-free metrics such as CometKiwi, xCOMET, MetricX-QE or LLM-as-a-judge with MQM/ESA error categories) decides which segments a human must see [5][3].
• Human-in-the-loop: post-editors (ISO 18587 workflow) for published and regulated content, terminology owners for the glossary, and final sign-off by subject experts for legal, medical and safety texts.

Data preparation

Typical input is spread across several places: Word and PowerPoint files on file shares, web content in the CMS, product data in the ERP or PIM, software strings in repositories, and old translations in the mailboxes of agencies and staff. The most valuable asset is often missing: a clean translation memory and a termbase.

Typical gaps are scanned PDFs without text, layout-heavy documents, several versions of the same file, inconsistent product names across languages, and no approved glossary. Access rights matter too, because contracts and HR texts must not be sent to a consumer tool.

Preparation steps: 1. Inventory what is translated per year, in which formats and languages (see PREP-001). 2. Run OCR on scanned originals (PREP-006, DOC-014). 3. Export old bilingual documents and agency deliveries into TMX with an alignment tool, and deduplicate. 4. Build a short glossary of 100 to 300 brand and product terms (PREP-015) and a style guide per locale. 5. Classify documents by confidentiality (PREP-024) so that sensitive ones are routed to an approved tool.

Effort: about 20 to 30 percent of the project, mostly the glossary and translation memory. A pilot works without them, but quality and consistency improve visibly once they exist.

Implementation paths

Which path for whom

• SMB: Path A, because mature EU-hosted MT products with glossary and document formats exist at SMB prices and translation is a standard process.
• Enterprise: Path B, because the value comes from wiring MT, translation memory and quality estimation into the CMS, PIM and TMS landscape with confidence-based routing.
• Developer: Path B, because MT APIs fit directly into i18n and CI pipelines for software strings.
• Private: Path A, because consumer and browser translators cover the need without any setup.
• Public: Path C, because sovereignty, procurement rules and confidential case documents favour open-weight MT on German or EU infrastructure.

Path A — Buy (off-the-shelf, recommended for most SMBs)

• MT products: DeepL (Germany, Cologne) Pro/API with glossaries, style rules, document translation, DeepL Voice and DeepL Agent (generally available since Nov 2025) [6][7][13]. Microsoft Translator and Google Translate inside Microsoft 365 and Workspace. EU options: Systran (France), Pangeanic (Spain) and the Commission's free eTranslation [8].
• Translation management / localisation platforms: Phrase (Phrase TMS, Strings, Language AI), memoQ (Hungary/EU), RWS Trados and Language Weaver, XTM, Smartling, Lokalise, Crowdin. Unbabel (Portugal) combines MT, QE and human post-editing as a managed service.
• Check: EU data processing and DPA; no training on customer data (DeepL Pro states texts are not stored or used for training without consent [6]); glossary support; file-format filters; SSO; connectors to CMS, PIM and Git.
• Choose A when: you need office and document translation, website/app localisation through standard connectors, or a combined MT + agency workflow.

Path B — Build on cloud APIs

• NMT APIs: DeepL API (text, document, glossary, formality), Azure AI Translator (EU data boundary, custom glossaries), Google Cloud Translation (Advanced, adaptive translation), Amazon Translate (Frankfurt region).
• LLM APIs for context-aware or style-constrained translation: Mistral (EU provider), Gemini on Vertex AI in EU regions, Azure OpenAI in EU data zones, Claude via EU-region cloud platforms. Pass glossary terms, the previous segments and the style guide in the prompt; use structured output to keep segment alignment.
• Orchestration: n8n or Make for e-mail and ticket translation; GitHub Actions with the Crowdin or Lokalise CLI for strings; Okapi Framework for file conversion to and from XLIFF. Add a QE step (COMET models or an LLM judge) to route segments.
• Choose B when: translation has to be embedded into your own product, ticket system or portal, or you want engine routing per language pair.

Path C — Self-host (open source)

• Open-weight translation models: TranslateGemma 4B/12B/27B (Gemma licence, 55 languages, image input, 2K-token context) [3][4]; EuroLLM-22B and EuroLLM-9B (EU consortium, all 24 official EU languages plus 11 more, fully open release) [9]; Tower+ (Unbabel, Qwen2.5-based, 2B/9B/72B, but CC-BY-NC, so non-commercial only) [10]; Tencent Hunyuan-MT (its WMT25 entry was the best constrained system in human evaluation, winning 11 language pairs) [1]. General multilingual LLMs such as Qwen3, Mistral Small or Gemma 3 also work well for major EU languages.
• Classic NMT on CPU: Marian NMT (MIT licence) and OPUS-MT models, CTranslate2 for fast inference, LibreTranslate/Argos Translate (AGPL-3.0) as a ready-made self-hosted API [12][14]. Meta NLLB-200 covers about 200 languages but is CC-BY-NC and labelled "research model, not for production" [15].
• Stack: vLLM or Ollama for LLM serving; Weblate (self-hosted localisation platform) or a TMS with a custom MT connector; COMET/CometKiwi for QE [5]; Qdrant/pgvector for retrieval of TM and glossary examples into the prompt.
• Hardware: HW1 (CPU server) is enough for Marian/Argos gisting. HW2 (one 24–32 GB GPU) runs TranslateGemma-12B or EuroLLM-9B for team-scale document volumes. HW3 is needed for 27B–72B models or high throughput.
• Choose C when: content falls under professional secrecy (§203 StGB: lawyers, doctors, tax advisers), classified or defence material, or works-council-sensitive HR texts.

Path D — Custom training

Worth it when there is a large, clean, domain-specific translation memory (typically hundreds of thousands of segment pairs or more) and recurring content: engineering, pharma, legal, patents, or a company's own product language. Options: DeepL glossaries and style rules (no training); custom-model features in Azure Custom Translator and Google AutoML Translation; or LoRA fine-tuning of an open model (EuroLLM, TranslateGemma, Qwen) on TM data plus post-edits. The TranslateGemma recipe of supervised fine-tuning followed by reinforcement learning with QE reward models shows the state-of-the-art approach [3]. Clean the TM first, because misaligned and outdated segments hurt more than they help.

State of the art (2026)

• LLMs lead overall quality: in WMT25, frontier LLMs (Gemini 2.5 Pro first) beat most dedicated systems in human evaluation, and the majority of participating teams built their systems on fine-tuned LLMs [1].
• Small specialised models close the gap: TranslateGemma-12B outperforms the larger Gemma 3 27B base model on WMT24++ (MetricX) [3]. Tencent's Hunyuan-MT and EuroLLM show that 7–22B models can reach competitive quality, which makes on-prem MT on a single GPU realistic [1][9].
• Document-level and multimodal translation: WMT moved to whole-document test sets with speech and screenshot sources [1]. Engines increasingly use the full document context instead of isolated sentences.
• Agentic localisation: TMS vendors and DeepL (DeepL Agent, generally available since Nov 2025) move from translating segments to running multi-step workflows: translate, check terminology, QA, file back [13].
• Evaluation: ESA/MQM human protocols and learned metrics (xCOMET, MetricX, CometKiwi) are standard. WMT25 shows that systems tuned against an automatic metric can top automatic rankings and still fall back clearly in human evaluation, so always validate with human review of your own content [1].
• Limits and failure modes: omissions and hallucinated additions (LLMs), wrong terminology, gender and formality errors, broken tags and placeholders, numbers and units, low-resource languages, ASR-noisy input, and over-fluent output that hides meaning errors. Raw MT must not be used for safety instructions, contracts or medical content without expert review.

Shared resources & combinations

• Shared building blocks: the translation engine (DeepL/Azure API or a self-hosted TranslateGemma/EuroLLM server) serves LANG-008 (multilingual summaries), LANG-010 (simplification pipelines), LANG-011 (proofreading after post-editing), CONV-001 and KNOW-002 (multilingual chat and RAG), SPEECH-003/SPEECH-004/SPEECH-007 (ASR → MT → TTS chains) and LANG-030.
• One GPU, many uses: a single HW2 GPU running vLLM with a multilingual LLM can handle translation, summarisation (LANG-008) and rewriting (LANG-011) for an SMB. ocr-docparse is shared with DOC-014 for scanned sources, and DOC-026 for accessible multilingual PDFs.
• Data assets: translation memory and termbase are reusable company assets. They feed brand voice (LANG-001/LANG-005), technical writing (LANG-022), creative versioning across markets (GEN-004) and embeddings for multilingual search.
• Privacy layer: DOC-018 anonymisation in front of external engines; guardrails for PII routing.
• Private: PERS-008 (foreign-language official letters), PERS-005 (local assistant).

Effort, cost & KPIs

• Path A: days. Per-user subscriptions or per-character API pricing; TMS licences per seat or word volume.
• Path B: 2–6 weeks for an integration with glossary, QE routing and connectors. Cost is driven by characters/tokens and by the human post-editing share.
• Path C: 1–2 months including model evaluation per language pair; one GPU server plus model updates and monitoring.
• Path D: 2–4 months, mostly TM cleaning and human evaluation.
• KPIs: post-editing effort (edit distance / time per segment), MQM error rate per 1,000 words, share of segments published without edits, turnaround time, cost per word versus the agency baseline, terminology adherence, number of languages served, and customer complaints about language quality.

People & change

Involve management first for budget and the rule which documents may be machine translated. The department lead (marketing, support, legal) owns the glossary and decides which content needs human post-editing. IT and the DPO check hosting, AVV and retention before the pilot; contracts and personal data need a clear rule. Legal defines which texts always get a human review.

Works council: depends. A plain translation tool without per-user evaluation rarely triggers §87(1) no. 6 BetrVG, but a cloud suite or CAT tool that logs user activity can, because the Federal Labour Court (BAG) accepts objective suitability for monitoring. Inform the council in good time under §90(1) no. 3 when work processes change; translators' jobs are the sensitive point. Public bodies involve the staff council (Personalrat).

AI literacy: support staff with a short training on limits (false friends, numbers, legal terms) and on what must never be pasted into free tools; Art. 4 AI Act asks employers to support AI literacy, see GUIDE-006. Communicate that translators move to post-editing and quality control, not out of work (GUIDE-008). Pick a pilot group of two or three language-savvy users, measure edit distance, and keep a human sign-off for contracts, legal and safety-relevant text. A company policy (GUIDE-007) keeps shadow tools out.

Risks, regulation & compliance

• EU AI Act: minimal risk. Translation is not an Annex III category and not an Art. 5 practice; the AI literacy duty (Art. 4) applies. Art. 50(2) marking duties for synthetic text apply from 2 August 2026 to providers of generative systems, but exclude systems that perform an assistive function for standard editing or do not substantially alter the input or its meaning [2]. Whether LLM translation falls under this exemption is not settled; GPAI-based providers should follow the Code of Practice. If the translation is part of a high-risk system (e.g. translating applicants' CVs for HR screening, or benefit decisions), that system's high-risk duties apply.
• GDPR: e-mails, HR files, medical and legal texts contain personal data. Use a DPA, EU processing and no-training clauses, and ban free consumer tools for business content. Use on-prem engines for special-category data (Art. 9) and for professional-secrecy holders (§203 StGB).
• Sector rules: Machinery Regulation 2023/1230 and IEC/IEEE 82079-1 (instructions in the official language(s) of the user's country). MDR Art. 10(11) and Annex I §23 (IFU and labelling in the required languages; the manufacturer remains responsible for accuracy). Pharma product information (EMA QRD templates). Court and authority documents need sworn translators. Quality standards: ISO 17100 (translation services) and ISO 18587 (MT post-editing).
• Typical risks & mitigations: meaning errors in safety-critical text → mandatory expert review; terminology drift → enforced glossaries and automated checks; prompt injection in LLM translation of untrusted text → strict system prompts and output checks; leaking secrets → engine routing by sensitivity class; misleading users → label machine-translated pages ("automatically translated") where the text has not been reviewed.

Getting started

1. Inventory translation content by type and risk: gisting, internal, published, and regulated or legal. Define which classes may go out as raw MT, which need post-editing and which need human translation.
2. Stop shadow use of free tools. Roll out an approved engine (DeepL Pro/API, Azure Translator in the EU, or eTranslation for eligible bodies) with an organisation-wide glossary.
3. Benchmark 2–3 engines (one NMT, one LLM, one open model) on 200–500 of your own segments per key language pair, using xCOMET/CometKiwi plus blind human review.
4. Connect the systems where text originates (CMS, PIM, Git, ticketing) through a TMS or n8n flow with TM reuse and QE-based routing.
5. Measure post-editing effort monthly and feed corrections back into TM, glossary and, at enough volume, a fine-tuned model.

Sources

1. Kocmi et al., "Findings of the WMT25 General Machine Translation Shared Task: Time to Stop Evaluating on Easy Test Sets", WMT 2025 — https://aclanthology.org/2025.wmt-1.22
2. EU AI Act, Article 50 (transparency obligations) — https://artificialintelligenceact.eu/article/50/
3. Finkelstein et al., "TranslateGemma Technical Report", arXiv:2601.09012 — https://arxiv.org/abs/2601.09012
4. Google, TranslateGemma 12B model card — https://huggingface.co/google/translategemma-12b-it
5. Unbabel COMET (COMET, xCOMET, CometKiwi metrics) — https://github.com/Unbabel/COMET
6. DeepL Pro data security — https://www.deepl.com/en/pro-data-security
7. DeepL API documentation — https://developers.deepl.com/docs/getting-started/intro
8. European Commission, AI translation and language tools (eTranslation) — https://translation.ec.europa.eu/tools-and-resources/ai-translation-and-language-tools_en
9. EuroLLM-22B Technical Report, arXiv:2602.05879 — https://arxiv.org/abs/2602.05879
10. Unbabel, Tower+ 72B model card — https://huggingface.co/Unbabel/Tower-Plus-72B
11. Mozilla Translations (engine behind Firefox Translations) — https://mozilla.github.io/translations/
12. Marian NMT — https://github.com/marian-nmt/marian
13. Tech.eu, "DeepL launches debut AI agent" (Sept 2025) — https://tech.eu/2025/09/03/deepl-launches-ai-agent-for-businesses/
14. LibreTranslate — https://github.com/LibreTranslate/LibreTranslate
15. Meta, NLLB-200 3.3B model card — https://huggingface.co/facebook/nllb-200-3.3B
16. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj