Reading time approx. 5 minutes
von Ljubica Negovec
Work has been underway to develop tools for the automatic translation of texts since the 1950s. Since then, translators have been told that their work is in danger of extinction. But is this really true and what can modern AI translators actually achieve? A report from translation practice.
Who among us has never had Google Translate etc. help us with emails with foreign business partners or while on vacation when translating the menu? Modern AI translation services are fast, easy to use and can be used free of charge for small amounts of text. As an alternative to communicating with hands and feet or in broken English, they have become indispensable. With the explosive publication of ChatGPT at the beginning of 2023, the topic of AI has also reached the minds of the C-suite, because AI promises enormous time and cost savings. However, the use of AI translators also entails risks that can result in incorrect translations with fatal economic consequences.
When AI causes a company to die
āThe ABC Group is an independentā¦ā That seems to be an easy sentence to translate. But for the ABC Group, a large Austrian company in the luxury segment, whose name has of course been anonymized here, this resulted in an embarrassing translation error in several languages āāthat was visible on the company website for weeks. The reason for this is that the service chosen for the translation uses English as the relay language (as most AI translators do, by the way). The German text was first translated into English and then from English into the desired target language. The problem: The AI āāconfused the German article ādieā with the English verb ādieā (āto dieā). This error has been carried over into all target languages. This meant that French, Czech and Dutch visitors to the site had to read the sentence āDie ABC Groupā¦ā. This is not only grammatically and incorrectly incorrect, but also literally fatal for the companyās brand and image.
Also on the ABC Group website was the German headline āEating together at a toque levelā, which praises the above-average quality of the employee canteen. Especially in Austria, someone who has been recognized by the Gault-Millau restaurant guide is referred to as a ātoque chefā or āaward-winning chefā. In the English-speaking world, where Michelin stars are the common award for restaurants, the literal AI translation āEating together at the same levelā is completely incomprehensible. Whatās more, ātoqueā can mean not only a ācookās hat,ā but also a āknitted hat,ā which has no place in a kitchen. It is clear to human translators that the expression ātop levelā in the English translation must be paraphrased, for example as āgourmet levelā. But the AI āācan only translate what is present in the source text. It has no way of responding to cultural peculiarities and adapting them to the expectations of the target audience.
These two examples were chosen because they particularly clearly illustrate the weaknesses of AI. Of course, they werenāt the only errors on the ABC Groupās automatically translated company website. Industry-specific technical terms were translated inaccurately and inconsistently, salutations repeatedly changed from āyouā to āyouā in the middle of a sentence, abbreviations remained untranslated, references were made incorrectly in complex sentence structures and the corporate wording was completely ignored in the foreign languages. For suppliers of cheap goods, the poor quality of AI-translated product descriptions is now just as expected as the poor quality of the products themselves ā and is no longer an obstacle to purchase for most of us. However, if a company that positions itself as high quality or even luxurious obviously cuts corners when it comes to translating the company website, even promises about the quality of its products no longer seem credible.
So do you have to forego the benefits of translation AI entirely?
No, because of course there are ways in the translation industry to take advantage of AIās advantages and mitigate its weaknesses. For example, AI can be trained specifically for a companyās texts. However, this is very resource-intensive and usually only makes economic sense at the scale of Amazon, Facebook or Booking.com. For all other companies and private individuals, the best solution is to rely on the combination of AI and humans. In so-called post-editing, the AIās output is checked and edited by a translator. In this way, it is corrected in terms of content, language and style and industry-specific technical terms or corporate wording are standardized (if the translator is paid extra for it). This is referred to as an AI-assisted or AI-based translation.
Where can AI-supported translations be usefully used?
For widely used languages āāwith a lot of bilingual training data, such as English, German, Spanish, Italian and French, AI translations are usually acceptable as a basis for post-editing. Languages āāwith complex grammar, such as Hungarian or Finnish, are not well suited for AI translation. Here it can easily happen that the AI āādoes not correctly recognize the relationships between parts of a sentence and reproduces the content incorrectly. The AI āācannot translate languages āāāāwith many dialects (e.g. Arabic or Chinese) well because the meaning of words differs greatly from region to region.
The department also makes a big difference. In general, the more specific the specialist area, the less useful an AI-supported translation is. The safety instructions in a toaster instruction manual are a great application example for AI translation with post-editing. The maintenance instructions for a custom-made heavy machine, on the other hand, must definitely be left to a specialist in this area. In contrast to AI, this can understand technical contexts, use illustrations of the machine as assistance and simply ask questions if anything is unclear.
The intended use of a text is also important. Where the aim is purely to provide factual information in another language, AI can score points. But if a text is intended to convince or touch people, human empathy is also necessary during translation. When it comes to advertising or literature, the effort required for high-quality post-editing is often greater than if the translator translates himself and chooses the most appealing and meaningful formulations for the target group.
A turnaround is noticeable
After a year of enthusiasm for AI of all kinds, the wind is beginning to change. New customers report failed attempts with AI translation tools, whose poor results led to criticism from the target groups and internally from native-speaking employees. At least in marketing, the trend is returning to the sensitivity and accuracy of human translation. But AI is no longer indispensable. They have revolutionized the translation industry in a way that only the inventions of the printing press and the Internet had ever revolutionized. But, as ChatGPT itself warns: āChatGPT can make mistakes. Consider checking important information.ā So if you donāt want to miss out on the time and cost savings of AI, this is it combination Recommended by AI translators and human, native speaker expertise.