Ertas for Translation & Localization

The Challenge

Translation in specialized domains is fundamentally different from general-purpose translation. Medical documents require precise translation of drug names, anatomical terms, and dosage instructions where an error can endanger patients. Legal contracts must preserve the exact meaning of clauses across languages, including jurisdiction-specific legal concepts that have no direct equivalent. Technical documentation needs consistent translation of product terminology, API names, and UI labels that match the localized version of the software.

Generic translation models, even high-quality ones, consistently fail on these domain-specific requirements. They translate brand names when they should be left untouched, use colloquial medical terms instead of clinical ones, and produce legally ambiguous phrasings of contract clauses. The cost of post-editing generic translations in specialized domains often exceeds the cost of translating from scratch, because reviewers must check every term and phrasing against domain-specific glossaries and style guides.

The Solution

Ertas lets organizations fine-tune translation models on their own bilingual material: translation memories, previously approved translations and glossary-constrained examples. Training on correct translations from their own domain teaches the model the organization's preferred terminology, style conventions and handling of terms that stay untranslated. Ertas Studio accepts parallel text pairs as JSONL, or a dataset imported from Hugging Face, and runs the training on cloud GPUs.

The fine-tuned model keeps brand names intact, uses approved terminology consistently and follows the target market's conventions: formal or informal register, metric or imperial units, date formats and cultural references. Exported from Hub and run on the organization's own servers, it acts as the first-pass translation engine in a human-in-the-loop workflow, producing drafts that professional translators post-edit. Source documents translated after export stay on the organization's infrastructure. The translation memory used for training is uploaded to Ertas, so remove anything confidential before upload. As the glossary evolves and new products launch, the team retrains in Ertas Studio from its stored dataset with newly approved pairs added.

Key Features

Example Workflow

A global medical device manufacturer translates product manuals, regulatory submissions and training materials into 8 languages. Its current workflow uses a generic MT engine followed by heavy post-editing by specialized translators, at an average of 12 cents per word. The localization team exports 200,000 previously approved translation pairs from its translation memory system, removes patient data and unreleased product details, and uploads them to Ertas Studio. Linguists spot-check a sample of each language pair in Data Craft. The team fine-tunes language-pair-specific models that learn the company's product terminology, regulatory phrasing and style preferences, then compares each run against a held-out set of approved translations, scoring glossary adherence and post-editing effort. The chosen models are exported and served on internal servers connected to the CAT tool. Translators post-edit every draft, and the team tracks post-editing time and per-word cost against its old baseline before rolling out to all eight languages.