Our approach

How does ModelFront catch bad translations?

The idea behind risk prediction is as old as machine translation itself: automatically catch bad translations. But for half a century it hardly existed outside of the lab.

To be useful for most real-world scenarios, a risk prediction system needs to be easy to customize, easy to evaluate, easy to integrate and easy to update.

How does ModelFront catch bad translations?

In research terms, we've built a system that generates "massively multilingual blackbox deep learning models for quality estimation, quality evaluation and filtering".

Massively multingual
One model for many language pairs
Blackbox
No reference translation required
Deep learning models
Based on global and client data, not hand-coded rules
Quality estimation
Line-level precision for production use cases
Evaluation and filtering
Also useful for offline use cases

Our core technology is built upon years of continued open scientific contributions by notable researchers in machine translation in industry and academia, as well as recent great advances in deep learning models, data and infrastructure.

We've productionized that risk prediction technology to make it accessible and useful to more players.


The right analysis

We love understanding what causes bad translations at the system and process level, as well as the challenges fundamentally inherent to natural human language.

The causes of translations errors are constantly evolving along with the systems, processes, languages and use cases, and we track them as they evolve.

The right input and output

Unlike BLEU, you can use ModelFront for new content or languages with no human reference translations, or on the human reference translations themselves.

We give you risk predictions that are engine-independent and customizable, and smart about valid differences in translations, like synonyms.

The right data and technology

We use large-scale monolingual open text in hundreds of languages, internally curated parallel datasets and hand-labelled datasets, and we continually invest in growing and updating those.

We actively develop technology focused on covering the known issues of the major types of translations outputs, like machine translation, crawled and aligned corpora, back-translation and human translation.


Since ModelFront was founded and the field of translation risk prediction has evolved, many of the early innovations at ModelFront like massively multilingual models have gained adoption across the top research organizations.

The technology and use cases in our space will continue to evolve quickly, and our approach will too.