It is possible to improve your translation performance with AI. But it does require a bit of care: the models aren’t perfect, and effective quality control procedures need to be put in place. I’d like to share my own experience with you.

You may have come across those Facebook pages or articles that make fun – sometimes rather harshly, in fact – of mistranslations. Some are frankly ridiculous, others downright pathetic. If you don’t want to end up on one of these mocking pages – and ruin your brand image for good – here are a few tips based on my own experience.

Two software programmes to improve your translation performance with AI

Ever since we set up the Dys MOOC on learning difficulties, I use the DeepL app to help me with the translations. I do mean ‘help me’. As you will see in the following paragraphs, translating with AI presents a few problems. It therefore requires human supervision at every stage.

I also use the programme (plug-in) Translate Press to help me translate articles and pages on our websites. There are two ways to use this programme: either let it translate the content automatically, or use it to translate the content manually. The advantage is that it translates everything, including what the website’s readers cannot see:

• the URL
• alternative text for images
• menu links
• etc.

Personally, I prefer to use the manual version of the programme in conjunction with DeepL. As I’ve added the DeepL extension in my Chrome browser, I can use it on a wide range of websites and also in my Gmail account to translate emails. The extension appears as an icon (the DeepL logo) next to the text I want to translate, as shown in the photo below:

I copy the French text into each field and use the DeepL translator for the relevant language by clicking on the icon. I repeat this process for each language. It may seem time-consuming, but in reality, it saves a huge amount of time.

How to improve your translation performance with AI: keeping an eye on common mistakes

These two programmes, DeepL and Translate Press, are among the best in their category. However, as always, artificial intelligence is not perfect and the results must be checked by a human.

Language detection errors

Take the verb “argumenter”, for example. DeepL consistently interprets this verb as a norwegian plural noun. Whereas for me, it’s all about the French verb in the infinitive form. The translation differs only slightly: “arguments” instead of “to argue”. But this distorts the meaning of the sentence.

Coordination between the various parts of a sentence

When I translate articles from our blog into other languages, the sentences are split into ‘strings’. A sentence may be split into several strings, particularly if it contains a hyperlink. In this case, because it is a linguistic model, it cannot align the strings with one another. They are treated as sequences of signals that are independent of one another.

For example, I recently translated an article on the ABC Learning Design method. At one point, a string appears which, in French, ends with ‘je vous ai présenté la’. The next string begins with the word ‘méthode’. In Italian, Spanish and Portuguese, ‘méthode’ is masculine (metodo). The phrase ‘je vous ai présenté la’ becomes ‘vi ho presentato la’ in Italian, but is followed by ‘metodo’, which is masculine. This means I have to correct every sentence containing feminine articles and masculine nouns. It’s no big deal, but it would undermine our credibility in the eyes of Italian, Spanish and Portuguese readers.

You can download this Mindomo mind map on this page from our partner Biggerplate.

If you do not have the Mindomo software, download it for free here.

You can also download the PDF format by clicking on the button below.

Issues relating to the generative AI predictive model

Large language models are probabilistic systems. When they encounter a string of characters, they estimate the probability of what comes next. For example, in a sentence that reads “la population française en”, the model will calculate that there is a 90 per cent chance that the next word will be “général”. In some character strings, the probability is so high that the model adds the word “général” of its own accord. However, as my context is academic and professional, the next word is likely to be different. But if I do not proofread this text, I will end up with an Italian or Portuguese translation containing the word ‘général’, even though it does not appear in the original sentence and has no meaning or function in the intended translation.

AI does not take context into account

A single word can have several meanings, which vary depending on the context. For example, I am using the word ‘qualification’ in a professional context. It therefore means ‘aptitude, a set of skills relating to a trade or profession’. However, DeepL systematically translates ‘qualification’ as ‘selection’ because, in the world of sport, this word means ‘to be ranked or selected’. Fortunately, there is a feature that allows you to ‘set’ the desired type of translation. However, this example, like others, clearly shows that translations must be constantly proofread and checked. Even though it saves time, this does not mean that they should not be reviewed by a human. That is why European legislation on AI transparency exempts content reviewed by a human from the labelling requirements applicable to content managed by generative AI.

Correcting contextual errors: using a glossary

Both DeepL and Translate Press have a feature that allows you to force the software to translate a word in a particular way. This is the ‘glossary’. It is a sort of dictionary or thesaurus containing a list of words and their designated translations. For example, when we created the Dys MOOC, we explained in the first part of the course that learning difficulties, such as dyslexia, dysphasia, etc., were difficulties arising from brain development, but not illnesses. However, on several occasions, the software translated ‘disorders’ as ‘illnesses’. MOOC participants were quick to point this out to us. Using the ‘glossary’ feature, we were able to review all the translations, ensuring that this error would not happen again.

This glossary also ensures that we do not translate the company name in our articles: in one version of the title (the OG title, for search engine optimisation), the name ‘Formation 3.0’ appears and was systematically translated. This is prohibited by law, as the name is registered in French only.

To improve your translation performance with AI: the counter-example of simultaneous translation

YouTube – just like Prime Video and other streaming platforms – has recently introduced an automated, LLM-based instant translation feature. The results are often hilarious. Words are mistranslated, or even completely made up by the model. So much so that the original video becomes practically

Conclusion: the ongoing need for human supervision to improve your translation performance with AI

The errors highlighted above are not a major issue in themselves. However, they underscore the fact that any content generated by generative artificial intelligence should be subject to human review. Furthermore, when you translate a text into another language using AI, it is best to have a basic understanding of that language and how it works. Otherwise, you risk losing all credibility with your audience.

Apart from the issues mentioned, AI-assisted translation saves us a great deal of time. This enables us to publish more content and attract new readers to our website. Without AI, we would not be able to publish in five languages and offer resources in six different languages.

Did you enjoy this article? Subscribe to our newsletter Enjoy this article? and don't miss out on any more publications! It's free and guaranteed spam-free!