Testunterlagen schnell und billig per KI übersetzen. Fäddisch! Oder?

00:05:58: Petra Kemmerzell: Thank you for your willingness to explain to our listeners what to watch for when using AI in translations. Let me start with a sentence of yours from the conversation we had at the bar that really stuck with me. You said: ‘AI saves me the typing but not the thinking.’ So, Peter, what is the biggest time gain achieved when using AI-based translation tools?

00:06:21: Peter Ganslmayr: Well, Petra, that’s quite simple. Everything you actually never want to do – namely that stupid repeating of always the same formulations, always the same sentences, you can do those things with AI. If I use AI here then I don’t use it in the actual sense of “Artificial Intelligence”, but also for machine-translation tools that are more or less AI-based. So, for such stupid things, AI is absolutely perfect. For example, the Canadian Weather Service has no more human translators for the translation from English weather forecast to French and vice versa. This is all because there are only so and so many different sorts of weather and you get along with that very wonderfully by using machine translation.

00:07:04: Petra Kemmerzell: What do I need to know when I talk about AI-based translations? You just mentioned a keyword – machine-translation tools. How is this different from AI-based translation? Can you please elaborate a little bit?

00:07:23: Peter Ganslmayr: Yes, of course. First, you can differentiate actual machine translation. It has organically grown in history, some parts are rule-based, some parts are based on statistics and then there also are neural networks. That started in the 1980s and that has nothing to do with AI at all. I’m talking about programs like Google Translate or better, since slightly more sophisticated, DeepL. You know those, of course. With DeepL, there is an AI component. But you can’t say that DeepL is something like ChatGPT for example. That’s something completely different. And ChatGPT is not really intended to be used for translations because for this, you need a machine-translation program, because, as I already said, a machine-translation program works on different levels: statistics, rule-based, and neural. So, what does a machine-translation program do? On the basis of the rules, the language rules that were programmed to the machine are used in a neural network. That means training the algorithm to work as in a human brain. And the emphasis here it’s trying to do it. Because for many, many reasons it doesn’t. One reason for this: language is actually something profoundly illogical, because it is used by humans.

00:08:43: Petra Kemmerzell: And when will it actually be legally dangerous if I rely solely on machine translation? Can you give me some examples?

00:08:58: Peter Ganslmayr: Oh, there are plenty. I always like to describe the problem using the example of a scalpel. In the hands of a neurosurgeon who performs open-brain surgery, a scalpel is a blessing. The same scalpel is more problematic to see at the supermarket in the hands of a raving-mad 3-year-old whose mother doesn’t want him to buy his beloved Mars bar. This means that translation can always be relatively safe, if afterwards, in the so-called post-editing process, an expert is in charge. And such experts are made of flesh and blood, and they are called translators. Another example: you must have noticed at the end of last year or at the beginning of this year, a 36-year-old passenger, born in Iraq, was on a train to be separated along its route and wanted to know from another passenger if the section of the train he was in was the right part to his destination. Now, he couldn’t speak German. So, he used a translation app, in this case Google Translate, and entered his question in Arabic. And the translation app, for whatever reason, translated that the train would explode en route. Now you can imagine what the passenger who read the German text did. He panicked, informed the train attendant and shortly after that the train was evacuated. 18 more trains were affected, and the result was a cumulative delay of 1,000 minutes. This is a classic example of how machine translation without human post-editing becomes the scalpel in the hands of the 3-year-old.

00:10:11: Petra Kemmerzell: Wow! That’s a wonderful story and I think it’s very drastic. Now, you are not only a professional translator and interpreter yourself, but you also train our future translators at university. How do you teach your students the use of AI-based translation tools? What are the limits or stumbling blocks these young talents have to know?

00:10:32: Peter Ganslmayr: First, you must know that no translator needs to fear machine translation, unless they are bad translators. Bad translators are of course in danger, because every machine tool can do a bad translation without post-editing. The students need to know how a machine tool works, what algorithms are processed but, above all, what problems machine translation faces when analyzing a German or an English sentence. Because you must not forget, for a machine translation tool, the end of the sentence usually is the limit. It doesn’t look any further! That means after the sentence a new sentence begins and in English, for example, it begins with the beautiful pronoun “it” which is not very rare in the English language. And in this case, the machine translation would need to know what this “it” refers to! Well, this could be he, she, it, a name, an object and your guess is as good mine. Sometimes it works and the program gets it right, sometimes it doesn’t. And then it doesn’t always work the same way in each of the following sentences! Students must be aware of these things, and they must understand the whole process and also, I repeat, learn proper post-editing. And this is a relatively complex procedure because post-editing doesn’t mean to re-write everything you get from the machine the way you want it to be. You need to take the results of machine translation and only correct obvious mistakes and errors. I can give you an example. You have the word “evaluation” in English. In German, this can mean “evaluation”, too, but it could also mean “judgement”, “survey”, “assessment” or “review”. So, if the word “evaluation” is used several times in one text, the reader needs to know if there is the same concept or idea behind this word each time or if we need different translations for this one word since it refers to different things. So, if terminology is not consistent, which means that not the same terminology is used throughout, we get onto the slippery slope of communication. We miss the goal of the translation, namely a clear, targeted – or target-group oriented, and functionally equivalent translation. That means, first, a translation must not be recognizable as a translation and, second, the translation must have exactly the same effect as the original text.

00:12:52: Petra Kemmerzell: And this, Peter, leads me to my last question which I wanted to focus more towards qualitative market research. Here, we always have so many examples where the machines can help with translations and where they are actually getting in our way. Strictly speaking, documents that need to be written for the target groups of our interviews. So how can I use AI and where are the limits? I think questionnaires and discussion guides is a good start and you can use AI here quite well. But what about any other test material?

00:13:49: Peter Ganslmayr: Well, at the receiving end of a discussion guide you have someone like you: a moderator who knows what they are doing. It will not throw you off track if in sentence 4 the word evaluation is translated as “judgement” and in sentence 7 the same word is translated as “examination”. You won't lose any sleep over it. With questionnaires, it may be different. If you lead the interviewee through the questionnaire, there’s no problem, because you can interfere if they get anything wrong. Let’s assume you don’t post-edit the machine-generated translation. That might actually be the case with guides and questionnaires. But now you need to be very brave. You know your target group, you know the participants and their population, respectively, their interests, their attitude, their opinion, their level of knowledge etc. The machine-translation program, however, has no idea, and you can’t tell the machine, hey, listen, this text is for a nurse, for a specialist, for a patient or for a general practitioner. You can’t tell the machine because machine-language tools are not listening. So, everything that addresses a “consumer” can be pre-translated by a machine in the best case. But then it must be post-edited by an expert for translation. And it gets very bad with everything that has to do with advertising, verbalized commercials etc. Because even as a human translator you’re leaving the area of “translation”, and you move over to the area of “transcreation”. I render wordplay, I go away from the actual words and build some beautiful idioms in German and so on. And I think it is an absolute no-brainer that a more or less static computer program cannot deliver this.

00:15:04: Petra Kemmerzell: Brilliant, Peter! Another example from my side comes to mind. You remember the last ads we both translated? We did the translations before we had any first reactions and before we had spoken to the global team and learned about their findings in the US market. We translated the word “excitement” more or less directly into “Aufregung”. But it became clear very quickly, that the direct translation of “excitement” into “Aufregung” was not what the ad wanted to communicate. “Aufregung” implies a kind of happy nervousness when or before you start something new. Yet this context of the ad, “excitement” meant “celebration”, “fun and joy of success” and, therefore, it needed to be translated with “Begeisterung”.

00:15:45: Peter Ganslmayr: Absolutely correct. And another important point comes in here. Ads are always provided with some illustrations, pictures. And here’s the thing, machine translation works without any visual context. And that underlines what I said before: machine translation is completely unsuitable for this.

00:16:19: Petra Kemmerzell: Thank you so much, Peter! Hope you have a wonderful day.

00:16:35: Peter Ganslmayr: Thank you, Petra, the same to you!