ModelFront makes instant line-level risk predictions on machine translations. It tells you if a translation is good enough, or bad.
This is not a test.
Isto não é uma prova.
1% risk is very low - it means the translation is probably good.
This is not a test.
Isto é uma prova.
99% risk is very high - it means the translation is bad.
ModelFront is ideal for high-scale streams of text where the goal is to machine translate as much as possible, and skip any human review for most rows. Read more about hybrid translation.
It works with any machine translation engine, including custom engines, as well as human translations.
ModelFront is not an app or platform. We just provide you an API, which you're free to use as you wish. It's easy to integrate and highly scalable. We also provide a console where you can upload large files for evaluation and manage your API access and account.
Our Standard pricing plan is very simple: $200 for 1M source-side characters. We train and deploy custom models at no extra charge. We regularly refresh models and continue to support you as your content, translations and goals evolve.
The main production and offline use cases for enterprises and language service providers are:
Evaluation - evaluate and compare engines quickly and easily, whether on large corpora or specific projects
Hybrid translation - safely auto-approve raw machine translation, and only send the risky translations for post-editing
Validation - evaluate and compare engines quickly and easily, whether on large corpora or specific projects
Filtering - filter parallel data, like translation memories, for training better machine translation
ModelFront is built to work with many different types of text, whether professional, user-generated or company-internal content.
ModelFront can catch many types of critical and stylistic errors:
Named entities - literal translation of named entities like person names, product and brand names, business names, geographic names
Key information - failure to translate numbers, prices, codes, dates, emails and URLs
Negation - dropped or incorrect negation
Structure - structural issues like word order
Garbage - encoding issues and inserted words or characters
Idioms - literal translation of idioms
Grammar - grammatical errors like morphology and agreement
Casing - proper upper and lowercase
Punctuation - punctuation and formatting
Formality - tu/vous, tu/Lei, du/Sie and so on
A large online marketplace has 10M new items per year - 50M total. Each item has about 100 lines of about 10 words (50 characters), and should be translated into 10 languages.
The cost of human translation for those 1B lines per year would be about $1B. So the client uses machine translate, which is instant 1000x cheaper - only $1M - but 10% of the lines have a critical error.
When ModelFront catches those critical errors with 90% recall and 90% precision, and risky translations are hidden or left in English until a human translator can review them:
90% are never seen by a human.
90% of errors are caught. That's 99% correct total.