Isch doch mega cool, nöd wahr, jetzt gits KI made in Switzerland.
(Translation: It’s super cool, isn’t it, now there’s AI made in Switzerland.)
A Swiss Startup and spin-off of the ETH in Zurich (that’s where Einstein was professor of theoretical physics) is now challenging DeepL and Google.
It’s called Textshuttle and it can do a few things they famously can’t, such as translate from and into Swiss German (yes, spoken dialect). Currently there are two dialects only – “Züridütsch” and “Berndütsch”, but more will be added depending on popular demand. And it can do Romansh (sometimes referred to as Rheto-Romance), that ancient tongue still spoken by some 60.000 Swiss. Both Schwizerdütsch and Romansh have been neglected by the giants Google and DeepL, and so it came about that a group of brainy language and IA experts based at the ETH felt it was time that multilingual Switzerland had its own translation tool to cover its national languages – plus a few others, with more being added all the time.
Romansh is a descendant of the spoken Latin language, replacing the languages previously spoken in the area of the Grisons, Celtic and Raetic. Earliest written texts date back to the 10th or 11th centuries, and the language declined rapidly during the 19th century – a consequence of industrialization. It became an official national language in Switzerland in 1938 (perhaps not surprisingly given the political context, 92% of Swiss citizens voted in favour…). More recently, efforts have been made to revitalize it (TV, radio, schools), but it is still endangered.
Co-founder Samuel Läubli says “Die Schweiz hat eine Plattform verdient, die unsere Landessprachen spricht.” More likely he actually said: “D Schwiiz het e Plattform verdient wo eusi Landessprache redet.” (“Switzerland deserves a platform that speaks our national languages”).
Until last week its language services were a well-kept secret, used mainly by those top-secret companies that have been at the centre of so much controversy – the Swiss banks (surprisingly, Russian is not yet covered by Textshuttle though). But on 10th May it launched a free translation service for anyone, which means it’s in an open challenge with its large rivals.
Textshuttle is keen to emphasize that they develop their software entirely within the Swiss borders and that it runs on (highly secure) Swiss servers. “When we say AI, we mean ‘real, modern machine learning’ – and you can see that when you compare our tool with other systems, such as Google Translate or DeepL”, says a confident Läubli.
Asked about whether Textshuttle is prone to hallucinations (see my last blog), Läubli says: “We have been able to massively reduce hallucinations in our models over the last few years, so that the problem is hardly significant – perhaps our system will ‘hallucinate’ a punctuation mark in a translation where there is none in the source text. In rare cases, the opposite occurs with us – and with all our competitors in the field of AI translation: Omissions, i.e. information in the source text that is omitted in the translation.”
So why does Textshuttle have such a good command of Swiss German, while other providers seem to cut their teeth on it? – that was the interviewer’s question. Läubli’s response:
“Large companies focus on large languages. This is relatively easy because AI processes for automated language processing are data-hungry. In other words: If I show my AI 100 million sentences that people have translated from English into German, it can quite easily learn to imitate them – it has to generalise less because it finds an example for almost every word in almost every context.
With languages like Romansh or Swiss dialects, it’s different: it is highly unlikely that there are 100 million sentences in the whole world that have ever been translated from English into Bernese Swiss. Not to mention the fact that people from the city of Berne and the Emmental would write the same sentence differently.
In the field of machine learning, this is an exciting challenge for us. Based on basic research at the University of Zurich, we have developed, among other things, methods with which neural networks – on which modern AI applications are based – can learn much more from much fewer translations made by humans. But this is just the beginning: our AI will get better, especially for Swiss dialects, the more people work with it on our platform.”
Here just a tiny experiment. First my input in Swiss German, followed by Textshuttle’s translation, and then DeepL:
Gosch du höt znacht uus go ässe? oder erscht morn obe? -Es chunnt e chli drof a, win i mi füle. Und öb ichs mer überhoupt cha leischte und ou wies mit der arbed vora got. Be mim Chef weiss me jo nje öb mer wider Überschtunde münd leischte. Aber natürli wärs ächt super me chönt mol wider e chli ploudere. Also, bis denn und heb der Sorg.
Are you going to dinner tonight? Or tomorrow night? -It depends a little bit on how I feel. And whether I can afford it, and also how it goes with the work. With my boss, you don’t know if you’re going to have to work overtime again. But of course, it would be really great if you could have a little chat. So, see you then and have the worry.
Are you going out for dinner tonight or first thing tomorrow? -It depends a little on how I feel. And whether I’ll be able to read it over and over again and what’s going on with the work. With my boss, I never know whether we’ll be able to talk about it again. But of course it would be really great if we could have a little chat again. So, see you and take care.
As you can see, Textshuttle does much better than DeepL, which gets lost in 3 places: the phrases market in italics and bold are serious mis-interpretations. Disappointingly though Textshuttle fails to give me the correct wording for the very last phrase (while DeepL in turn interprets correctly) which was used almost like a mantra during the recent pandemic.
Clearly, the Swiss Textshuttlers have more work to do! And I’ll keep testing.