How can we train natural language models to be accurate when human bilinguals can't even agree on what's wrong?

If you ask two people whether a translation is good or bad, in all likelihood you’ll get back two very different answers. How can we objectively measure translation quality when it seems so inherently subjective?

At Unbabel, we’re not just leveraging AI — we’re advancing the state of the art. Discover the world’s first native-quality translation pipeline.

Discover OpenKiwi: an open-source framework for Machine Translation that allows the MT community to implement our award-winning Quality Estimation systems.

There’s one reason most customer service managers won’t consider automating translation: quality. Discover how Unbabel is challenging that assumption.

If you believe the hype coming out of certain tech companies’ PR departments, you’ll hear all sorts about how recent advances in “neural” mean that machine translation is “nearly indistinguishable” from professional quality. But of course, you’ve actually used those services and know that’s absolute nonsense. Because machine translation engines are trained on simplistic and …

So said João Graça, Unbabel’s co-founder and CTO, to me the other day as we were having coffee. Truth is, machines don’t always get it right and we all know how embarrassing a bad translation can be. But how do we know when an automatic machine translation is bad? Is it possible to know where …

Translation is about finding the equivalent meaning to a sentence in a different language but more often than not you don’t have the exact equivalent words.

Understand with Unbabel is a series where we dive deep into the topics and challenges faced as we accelerate towards a world without language barriers.

Quality and speed might be perceived as contradictory choices, and scaling both is an immense technological challenge, but it’s key to our success.