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.
Machine translation - simply hide content or translations when the translation is risky
Hybrid translation - send to human translators when the translation is risky
Evaluation - evaluate engines quickly and easily, whether on large eval sets or specific documents
Comparison - compare translations from multiple engines and pick the best translation, at the document or even line level
Engine development - find errors at scale when training and deploying your own machine translation system or customized system like AutoML
Creating datasets in many language for training for other downstream machine learning tasks
ModelFront is built to work with many different types of text, often user-generated or company-internal content:
Product titles and descriptions
Reviews and comments
News headlines and articles
Notifications and alerts
Emails, messages and chats
Legal contracts and patents
Scientific papers and reports
Search indices and queries
Keywords and tags
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
Soon there we'll return separate scores for each error type, so that clients can choose which are critical.
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.