ModelFront works for many types of content. We built it to support customization from day one.
We're currently serving accurate risk predictions for everything from human translations to high-scale datasets of machine translated user-generated text.
It's not a great fit for very one-off, artistan content like marketing slogans, which require careful handcrafted translation or transcreation of every line.
The API returns a risk prediction from 0.0 to 1.0 (0% to 100%) for each line - an original segment and its translation.
See our API docs
To get started, all you need to send is a few hundred lines of sample data in the source language, and tell us your target languages.
Yes! In fact, we ask that clients send at least a bit of sample data - a few hundred rows of source-language text - to guarantee them accuracy and satisfaction. The more data we have, the better!
Like with machine translation, more is better. 100K segments is a good minimum. It can be from a mix of language pairs. If you're really lacking data we can get some more for you.
We work with clients to define evaluation sets. We work to achieve the target accuracy, and then notify you to try it with another set that we and our models have never seen.
The two key numbers here are: recall - how many of the errors it catches - and precision - how many of the errors it caught are actually errors.
For more information, read about our metrics!
A risk prediction score is just the risk of a critical error.
For example, su in su carro can be translated from Spanish to English as her, his, its, their or your. Without more context, we can't say more about the quality, but we can say that it's risky.
Traditional quality estimation or post-editing effort estimation correlates mostly with the edit-distance to a reference translation.
That's not always useful, because certain types of errors are much more painful than stylistic variations - especially mistranslated brand, product, person or place names, mistranslated prices, sizes and codes, changes in negation and dropped, inserted or offensive words.
We can shape risk models to your exact needs - whatever bad means for your application or users.
Translation memory matching requires a translation memory with significant coverage over the new content. Perfect translation memory matches may still be high risk.
Risk prediction works even independent of any translation memory or reference translations.
Yes. This is a very interesting topic with no simple answers. We've found that the best way to control for these factors is to work with clients on their custom definition of quality for each type of text.
Like most binary classification models, our models implicitly make the cutoff at 0.5 (50%).
We recommend using different threshold for different types of text, based on the value to the user. For example, some products generate much more revenue than others, so the risk threshold should be lower. Zooming in, the title of a product is more important than the description, so the risk threshold should be even lower.
Our API returns the raw risk score precisely so that your application can use it in the best way for you and your users.
Yes. They are not required, but custom models will give you an accuracy boost.
Yes, it can. The models are not specific to any engine. If you have an internal engine or a custom flavor of an engine like AutoML, then for best accuracy we recommend a custom model trained on its output.
Yes, it can. Clients routinely find errors in human translation with ModelFront. That said, human translations have a very different signature and error typology than machine translations. For example, humans make typos, confuse false friends or simply do not know some very obscure words. Literary human translation or transcreation often involves significant restructuring.
No. ModelFront currently returns segment-level scores, which can also be aggregated into paragraph-, document- and corpus-level scores.
Not yet. It is on our roadmap. This will let clients set specific risk thresholds for specific error types, and understand their translation engine's overall error typology on a given content type.
We do offer strict and tolerant models. Tolerant models ignore certain types of cosmetic errors. For example, for user-generated content it can make sense to allow inconsistent casing, punctuation, locale formatting or orthographical variants.
Our technology can handle any of the top 100 languages. Our generic base model currently supports only those for which we have significant high-quality data.
Our technology can handle any pairing of the more than 100 languages we support.
However, it's important to know that top machine translation systems, like Google Translate, Microsoft Translator and DeepL, translate via English. This means that the quality is much lower for most pairs.
Read more about direct translation.
Our current technology can handle markup languages like XML and HTML. It does not check the correctness of the tags, only of the text.
To ensure users' privacy and comply with the EU General Data Protection Regulation, we automatically delete request data.
We also recommend that clients pseudonymize sensitive data before sending it for training and do the same before each prediction.
Just as fast as translation APIs.
The API supports batch requests. We can also support very large offline jobs, and datacenter co-location.
Yes. It's a JSON REST API, so it's easy to call from any production-strength programming language.
See our API docs