If you follow translation news, you may have noticed a few recent articles that claim AI can help with localization. The idea is intriguing, but the more you look into it, does it really hold up?
What is localization?
Localization involves adapting content, including marketing material, awareness campaigns, patient education material, and entertainment, to specific markets.
For instance, a company might choose to use different celebrities to be the “face” of their product, depending on their popularity in a specific market.
Localization involves a deep knowledge of the culture and, often, language, of a specific target group.
Transcreation is localization at another level. In addition to creating content that will resonate with a particular audience, the language being used has to, as well. It’s not enough, for instance, to have a standard Spanish translation of an ad. Transcreators have to be able to adapt things like slogans, slang, and idioms so that they’ll be understood by different cultures. They also have to have an understanding of everyday life in a particular place to be sure that material is relevant.
This is understandably a complex task, and leaving it to someone who is inexperienced or unqualified can result in newsworthy fails.
With this in mind, using artificial intelligence, which learns and outputs language based on things like print sources, databases, and algorithms, seems like a very bad idea when it comes to the subtleties of adapting materials to different audiences.
Why are some sources saying that AI can be used for localization?
Despite this, several sources claim that AI is becoming intelligent enough to be used for localization.
If you read these articles closely, they offer little - and usually no - proof that machine translations are effective in generating localized or transcreated material. Instead, most point to some ways that AI could help localization and transcreation experts do their job.
For instance, machines can now handle things like billing. Programs can be set up to automatically convert currency or other standard features of different markets. And of course, translators can use CAT (Computer-Assisted Translation) tools to help with things like using established jargon or translations for specific terms.
But little of this directly relates to the complex art of localization or transcreation.
How can AI be used for localization?
In recent years, AI has become a way for many popular apps and websites to adapt content to their users. One often cited example is Spotify, which creates music playlists based on what a user likes to listen to. Some experts claim that AI can do the same thing for localization. For instance, by studying trends in different markets, AI can create suggestions for consumers.
A good example of how this has already been implemented is Google Search results, which are adapted to individual countries and regions. So, AI expert Ilia Shifrin explains, when someone types “car” in the United States, the top suggestions will be related to car models that AI finds to be popular with US consumers. A Russia-based user would get results based on what algorithms show are popular car choices in the Russian market.
To some extent, these specialized content and search results could be considered localization. But they don’t cover language. As much as AI can follow algorithms to help you find your new favorite song, or even provide basic, not always accurate, translations, it’s a known fact that it has trouble with figurative language, slang, and lesser known languages and dialects.
One source that I came upon in my research claimed that Google Translate has dramatically improved in its ability to translate slang words. The example used to illustrate this impressive claim is two side-by-side images showing an improvement in accurately translating the French phrase Un sourire coûte moins cher que l’électricité, mais donne autant de lumière (A smile costs less than electricity, but gives just as much light).
The phrase uses figurative language, which AI can have difficulties with, but it contains no slang and in fact uses extremely simple, standard French vocabulary that shouldn’t have posed much difficulty to translate in the first place.
So, take the improved slang translation claims with a grain of salt.
Can AI localize images?
Another popular claim is that AI can now localize images. One of the applications of this is to take text photographed in one language and translate it into another. This could be useful for anything from adapting stock photos, to ad campaigns…if it worked.
Several articles I’ve come across illustrate this idea with a side-by-side comparison of a photograph and its “localized” version. The first is a photo of shelves in a Francophone supermarket’s produce section. The image beside it is the same photo, with the labels shows the labels below each veggie translated into English.
It’s certainly impressive, at least at a glance. Unfortunately, while many of these produce terms are, again, basic French words, their translations aren’t always accurate. For instance, the translation for patate is misspelled “potatoe”. More egregiously, p. de terre, another term for “potato”, hasn’t been translated at all.
But the most notable errors lie in the limits of what the software can do. For instance, the bin at the top right is labeled piment, which the AI has translated as “chili”. But piment is a notoriously tricky French word to translate, since it can be used to mean a number of different peppers, or even simply, “spice”. How can a machine know which option to choose?
Add to this the fact that the vegetables in the bin look nothing like chili peppers, an issue that’s found throughout the image; this seems to be a very poorly organized produce section, with products that don’t always match up to their labels. If AI is programmed to simply translate labels, it of course won’t be able to catch this issue. But a human translator, localization specialist, or transcreator would.
So, can AI help with localization and transcreation?
Many articles and websites that seem to promise that AI can or will be able to create coherent, correct, and relatable content adapted to all markets, but as of now, that’s not exactly true.
What AI can do is help with workflow and use data and algorithms to create recommendations adapted to customers in different markets. Not a negligible contribution to localization, by far.
AI may also be able to handle translating standard materials like forms and standard symbols like currency.
But before we start thinking AI will soon be transcreating entire ad campaigns or adapting text in social media images to foreign markets, let’s remember “potatoe”, the inability to recognize produce, and the promises of translating slang that really only comes down to translating expressions that use basic vocabulary…and still takes multiple tries get right.
When it comes to localization and transcreation’s goal to adapt content so that audiences with different languages and cultures can connect with it, it’s hard to entirely replace a human touch.