How we brought Swiss German to DeepL

Many businesses assume that operating in Switzerland is a case of localizing into Italian in Lugano, French in Geneva and German in Zurich. In doing so, though, they’re excluding the language variant that the vast majority of Swiss people actually use.

"A lot of companies run their whole DACH operation out of Germany, and localize for Switzerland by changing the flag on the ad," says DeepL’s Account Executive for Switzerland, Tina Maric. "German is German, they figure. But it isn't. Swiss customers see the difference instantly, the same way British would notice when they receive a copy that’s written in American English."

Switzerland might be part of the DACH region that’s defined by a shared German language, but the German that’s spoken and written in the country is very different from Standard (or High) German. Swiss German is distinct, a source of national pride that will be a big part of Swiss National Day this coming weekend. It’s also distinctly varied, comprising both the written form of Swiss Standard German, and a complex linguistic landscape of regional and village dialects that vary from both the standard form, and one another.

The details Swiss German speakers instantly notice

The written form of Swiss Standard German (Schweizer Hochdeutsch, as it’s usually referred to), is used by all Swiss German speakers, about 62% of the population. It’s close enough to Standard German to be easily understood when read by a German speaker, but incorporates distinct characteristics that a Swiss German speaker will instantly pick up on.

There’s no eszett, which Swiss German dropped in the 1930s. Straße becomes Strasse. Weiß is Weiss. Any occurrence of ß in Swiss copy comes across as visibly foreign.

The vocabulary varies significantly, in part due to the close proximity and influence of French and German. Swiss Standard German uses Velo, not Fahrrad, for bicycle. Parkieren, not parken, for to park. Estrich, not Dachboden, for the attic. The distinctive vocabulary is referred to as Helvetisms.

When these elements go missing, it highlights content that’s been written for Germany and then just distributed over the border.

"Swiss people are genuinely proud of Swiss German, and equally of their own dialect within the language,” explains Olivier Raemy, another of our Account Executives for Switzerland. “Someone from Bern will tell you their dialect isn't Zurich's. You can't paper over that with generic German."

Beware shortcuts for similar languages

Insights like these have made launching Swiss German a priority for DeepL. Indeed, it’s one that was requested by the Bundeskanzlei (the Swiss Federal Chancellery) itself, for translating documents and co-ordinating work across federal departments, and across the several different official languages of government. It’s a challenge that our Research Manager Markus Kaiser has been rising to, over the last six months. And it involved making several crucial choices about how to approach the language.

"You could start off thinking that, because written Swiss German is similar to High German, you can get away with translating it into German and then adding customization,” says Markus. “That was the first approach that we tried, but we realized that it doesn’t hold up in practice.” Swiss German has its own distinct identity and evolution. Linking it to another language within the platform compromises this.

Focusing on the written form

Instead, Markus and the team took the decision to treat Swiss Standard German as an independent, written language and focus their efforts on representing it in all of its nuances, distinctive grammar and vocabulary. 

Equally important was the choice to focus on this written form, rather than the varied family of regional dialects (Zurich Swiss, Bern Swiss, Valais Swiss) that people actually speak. Trying to ignore this split would have meant trying to train a model to adapt to infinitely shifting variations that lack a fixed written form, and for which insufficient training data is available. 

Focusing on Swiss Standard German (or Canonical High Swiss German, as it’s sometimes referred to) involved prioritizing the language of emails, contracts and news. However, it also reflected the reality of how Swiss German works in practice.  The common foundation of this language is written. It’s what every Swiss German speaker can relate to.

Building a standalone Swiss German model 

Markus and the team put Swiss German through the same model training pipeline as other standalone languages on DeepL: a purpose-built training set with refined vocabulary and grammar, testing and evaluation with specific language experts, carefully filtered vocabulary and grammar. Internally, DeepL evaluates Swiss German on its own terms, protecting it from degrading when other languages get updated.

Throughout the six-month journey to perfect the language model, research and product teams iterated together. Research optimized for quality. Product stress-tested this against the experience of the language that Swiss German speakers wanted to have.

"Just because research likes the quality of our model doesn't mean customers will", says Markus. "People will always find the small details that don't quite work."

Working in Switzerland has always involved embracing a rich variety of different languages. With Swiss German now available for the DeepL Translator platform, including document translation, and the DeepL API, we’re ensuring that the most commonly used of those languages is experienced and represented the way it should be. Learn more: Decoding Trust in Enterprise Language AI. In conversation with Slator