How we made Clarify
Clarify, the on-demand translation expert that helps you communicate exactly what you mean, isn't just a groundbreaking feature for DeepL. It represents a new approach to using AI. It’s a fundamentally interactive experience that delivers more human-like results than AI working alone.
Building a genuinely interactive AI experience — one that can understand when and how to engage with a user — represented a new kind of challenge that required many different perspectives to solve. We asked three people at the heart of the project to tell the story of how we built Clarify. Please meet Product Designer Freddie Sukprasong, Research Scientist Danielle Saunders and Engineer Simon Lenz!
Tackling the unspoken issue in machine translation
Simon Lenz, Front-end Engineer: "Work on the Clarify project started around a year ago, but the idea of solving ambiguities and assumptions in translation is much older. It’s something we had looked at prototypes for in the past. After all, we knew from the way that people edited our translations that a need existed. With the emergence of DeepL’s own LLMs, it was suddenly possible to solve this issue – and that’s what we set out to do.”
Danielle Saunders, Research Scientist: "Ambiguity is a known problem for machine translation, but it’s the aspect of machine translation that nobody traditionally talks about. People have side-stepped the issue because they haven’t had a way of knowing what the user wants. If you think about what Clarify does to solve this, there’s a lot of subtly different things that need to happen. It had to identify ambiguities, ask the right questions about those ambiguities, and know how to solve them.”
Focusing on the right questions
Freddie Sukprasong, Product Designer: "Prior to introducing Clarify, one of the points of feedback we tended to get was around gender questions. In English, a doctor is a doctor, but when you translate into German, it’s often the case that you specify a gender for that doctor, and that introduces the potential for bias in translations. However, if we can signal to the user the assumptions that a translation is making, it turns this potential problem into something that’s actually helpful, by asking people which approach is correct for them.”
Danielle Saunders, Research Scientist: "Our in-house language experts collected examples of questions that Clarify could ask, and we clustered our data to identify general categories. We saw that a lot were about gender, some were about formality, others were about words with multiple meanings. This actually tracked really well to what Freddie, Simon, and the Product team were interested in, and so it was nice to be able to jointly solve what we were each working on. Throughout this project, we’ve had weekly meetings as a regular touchstone between the Engineering, Product and Research teams — and that collaboration has been really critical to making Clarify a success.”
New ways of working for a pioneering feature
Simon Lenz, Front-end Engineer: "For our teams, the Clarify project represented a different way of working. Historically, Research comes up with an idea, they work with Product to develop it, and then Front-end Engineering gets involved once the logic has all been settled. We turned that process around, creating prototypes that Freddie could use for user experience testing internally, and then starting again with more knowledge. I developed two prototypes that I threw away before we got to the feature we would implement, and the learnings from those prototypes have really helped to deliver a higher-quality result.”
Danielle Saunders, Research Scientist: "In the training, we are collecting a lot of examples of text and using those to help the models recognize when to ask a question. The feedback on the prototypes really helped us identify when we were asking questions we shouldn’t be asking, and if we were missing questions that we should be asking.”
Balancing quality and the user experience
Freddie Sukprasong, Product Designer: "Clarify is very powerful and can come up with so many different potential questions. It can detect specific names — for example, a local newspaper — which might not be familiar in the target language, and ask if the user wants to add more explanation. Abbreviations often don’t translate easily from one language into another, which is another thing that Clarify can determine. There are idioms and words with multiple meanings. Then there’s something as simple as date formatting — whether the user wants to list the day or month first. That’s crucial in business operations because it avoids so many misunderstandings.
One of the big challenges in building the solution was making sure that the user doesn’t feel overwhelmed by all this, and that we aren’t asking them unnecessary questions. We want people to feel that they understand what Clarify is trying to do, and that this feature is genuinely helpful and easy to use.”
Simon Lenz, Front-end Engineer: "In a way, the translation and the questions are two separate processes that are independent of one another. We needed to combine these in a way that works for users. The easy option would have been to translate the entire text first, and then ask all of the questions in one go. However, that’s not what people want, especially if they’re translating a long piece of text. We group questions in a way that avoids having to ask the same question too often, but still delivers accuracy. There’s a balancing act in how many times you ask whether the lawyer being mentioned is male or female, for example.
A big issue that we were always aware of was the time this takes. Our first prototypes on test infrastructure had waiting times of 20 seconds for Clarify to ask questions. Now, in production, we are closer to two seconds, and that is a world of difference.”
Freddie Sukprasong, Product Designer: "We are still in the first iteration of Clarify. We know there are many things we’ll be able to improve, including adding more languages, and we’ve got a long list of ideas that we will be testing.”
Danielle Saunders, Research Scientist: "The exciting thing is that the hardest part of this process is now complete. Launching Clarify for our first pair of languages, and getting it to the stage where we were happy to share it with users, has been a really transformative step, and we are excited to continue building upon that.”
The confidence that comes from Clarify
Freddie Sukprasong, Product Designer: "It’s great to hear what customers are saying about Clarify! What I’m hearing most is that it’s giving them confidence and control, especially if they don’t have a multilingual person on hand that they can ask to check a translation. Having Clarify do that makes them feel much more confident.”
Simon Lenz, Front-end Engineer: "I think it’s a change in the paradigm for a solution to ask for your input in this way. I can feel the extra confidence myself when I use it. I can be confident that there are no ambiguities because I can see the questions that Clarify has proactively asked me. I’m a native German speaker, and even so, Clarify shows me potential issues with gender in language that even I wouldn’t have thought of. It’s an awesome feeling to know that those things are being spotted and taken care of.”
Ready to try Clarify for your team? If you have a DeepL Pro subscription, it’s waiting for you to use: https://www.deepl.com/en/translator/l/en/de