Unbabel started with a simple surf trip — just founders, great waves, and a vision for a world without language barriers. Back in 2013,Vasco Pedro, João Graça, Sofia Pessanha, Bruno Silva, and Hugo Silva were frustrated that technology had made this huge promise to solve machine translation (MT), but was still very far away from realizing that goal. The question they wanted to answer was “What would the world look like if there were no language barriers?”. It was during a surf trip that the would-be partners started to articulate the answer to this question and founded Unbabel with the vision of removing barriers to customer experience in any language by introducing a revolution in the way technology and humans collaborate in translation.
Today Unbabel has more than 350 employees in offices and hubs spread around the world (~150 in Lisbon, and ~200 in locations such as but also San Francisco, Pittsburgh, Timisoara, Edinburgh, Cebu, London, Berlin, Madrid, and New York) and it is building a language operations platform to enable teams across a business to interact with customers in any language. Powered by AI and refined by a global community of more than 100.000 human translators, Unbabel combines the speed and scale of machine translation with the authenticity that can come only from a native speaker to remove language barriers for some of the biggest and most well-known customer and B2B brands in the world. With Unbabel, you don’t have to speak the same language to be on the same page.
Despite the progress in the adequacy and fluency of MT systems, critical translation errors are still too frequent, including deviations in meaning through toxic or offensive content, hallucinations, mistranslations of entities with health, safety, or financial implications, and many others. These errors occur more often when the source sentence is out of domain or contains typos, abbreviations, or capitalized text, all of which are common with user-generated content. Observing these behaviors, Unbabel realized it would be a good idea to prevent the systems from generating such errors using our award-winning machine translation quality evaluation and estimation technology. That's how the QUARTZ: Quality-Aware Machine Translation project originated. It aims to enable machine translation systems to be more responsible while generating translations, and, in particular, to be more robust to critical errors. With ELISE funding, Unbabel plans to develop approaches to machine translation that are more robust to critical errors, evaluate these approaches with our community of translators on real-world data coming from different types of content and enable the usage of such approaches in production environments that can serve a large number of translation requests per day.
Maria Ana Henriques, Unbabel’s R&D Project Manager explained: “With QUARTZ, the translations generated by machine translation systems will contain less or no toxic text, critical errors, and spurious segments, allowing translators to focus on other types of errors while post-editing or evaluating automatically translated output. This leads to cheaper and better-quality translation processes. From a communication perspective, having MT output with fewer critical errors has a positive social impact as it enables easier communication between parties that are not fluent in each other's languages. This leads to better understanding and satisfaction, and less frustration.”
Henriques added: “QUARTZ will unlock opportunities in sectors where translation quality is key, such as life sciences, travel, fintech, and retail. With QUARTZ, we hope to make machine translation perform a few steps towards responsible machine translation systems. Such systems are more controllable and produce fewer or no critical errors.”
As an outcome of this project, the team has just presented one paper describing an approach originated within the QUARTZ project at the 2022 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). Furthermore, another paper describing the project has been published at the 23rd Annual Conference of the European Association for Machine Translation (EAMT).
Want to learn more and stay in touch?
Visit Unbabel’s website and follow them on Twitter and LinkedIn!
More about the ELISE 1st Open call and awarded companies is available here.
