Lettie Dorst: ‘Translation programmes change how we interpret the world’
Associate Professor Lettie Dorst has received a Vidi grant to research how machine translation programmes such as Google Translate and ChatGPT translate words and expressions used metaphorically. This still regularly goes wrong, resulting in far too literal, incorrect and sometimes incomprehensible translations. When do these mistranslations of metaphors result in problems for users of translation apps?
Entering a text into Google Translate or ChatGPT will give you back a translation within seconds. While this often leads to convincing-sounding texts, the models continue to struggle with metaphors. Dorst’s aim with this project is to find out which metaphors are difficult to translate for which models and what kinds of errors this leads to. Dorst: 'When ChatGPT translates something from English to Dutch, it looks purely at common language patterns in Dutch, and it may well happen that words and phrases appear in the translation that were not in the source text, so-called hallucinations, or that the meaning changes completely. Most users are completely unaware of this, and at the moment we can't quite predict when exactly it will go wrong.'
Contemporary machine translation programmes like Google Translate, DeepL and Microsoft Bing are trained on existing translations and use deep learning and neural networks. Conversational AI like ChatGPT and Copilot use Large Language Models and are basically not specifically developed or trained to translate but can be used for that purpose. Both translation models and language models look for common language patterns and use them to predict what the translation should be.
Metaphors
‘When things go wrong with these translations, you see it in the often overly literal constructions,’ Dorst explains. 'Then you see, for example, that the English expression “keep your voice down” is translated as “keep your voice low” instead of “don't talk so loud”. We don’t recognise this literal translation as a fixed expression in Dutch, and you would therefore expect the GPT models to avoid such a construction, but almost all machine translation programmes choose it, perhaps because the construction “keep ... low” does occur in many other combinations, both literal and figurative.'
Why those translation programmes sometimes manage to find the correct translation and sometimes do not is still a mystery. Machine translation based on algorithms remains basically a black box; you can only see what comes out, but not how the machine arrives at a particular conclusion. ‘We hope to gain more insight into this by investigating what translations are used with respect to different types of metaphors in different types of texts,' Dorst says. ‘Does Google Translate produce more conservative and ChatGPT more creative translations? What is the reason for that? And how consistent is the programme in applying certain translation strategies? Those are the kinds of questions I'll be looking at in this research project.'
Changing language
Dorst is going to look not only at the machine translation programmes in her research; she will also look at the effects the translated texts have on people. 'In pilots of this study, it emerged that some participants actually liked the mistranslations. For example, “He got under my skin”, which means “he irritated me” was translated literally as “hij kroop onder mijn huid” (he crept under my skin). Respondents felt this was “beautifully visual” and “nice and gross”. They found the mistranslation creative and insightful, even though the expression doesn’t exist with this meaning in Dutch. In the project, I want to find out whether people understand such mistranslations because they know the English metaphor, or because the expression resembles existing Dutch expressions, such as “in iemands huid kruipen” or “iemand op de huid zitten”, or because they resolve the metaphor as a creative puzzle and then actually appreciate it more.'
Journalists and literary translators
'Journalists and literary translators are an important link in the use of metaphors,' Dorst continues. If they adopt a mistranslation - whether unconsciously or not, or because they like the sound of it - it helps normalise that particular metaphor. Because the programme mistranslates something, it potentially changes the way we interpret, describe and understand the world around us. You can already see the effects of this. For example, I saw an ad that said “Geld doet de wereld draaien” (Money makes the world go round). That's just plain wrong in Dutch, but apparently it's already become established as a metaphor via English. The big question is, of course, when do such translations enrich our language and when do they cause misunderstanding and miscommunication?
If we think of medical or legal texts, mistranslated metaphors can have major consequences, which of course nobody wants. But it is precisely those texts where one might not immediately see what the consequences could be that I find really interesting. With this research, I not only want to see how these mistranslated metaphors come about, but also how we can make sure we get better at recognising and avoiding problematic translations.'