Much has been written about the artificial intelligence (AI) revolution happening around us, as new tools like ChatGPT, DALL-E 2, and Github CoPilot radically shift our understanding of what tasks can now be handled or at least assisted by machines.
As an eLearning publisher who specializes in technology training, I’ve been playing around with these new tools for months, hoping to find new tools to help improve our workflow. We’re already using ChatGPT and Bing to help in content research, and Dall-E 2 and Midjourney have occasionally helped with certain image needs. But what about localization?
Everyone in today’s global, multicultural world wants their training materials to be available to as many people as possible. And while AI translation tools have been around since the early 2000s via tools like Babel Fish and Google Translate, their use in any professional context would usually be described as “mixed at best.” Their translations might be vaguely understandable by a native speaker, but it was never something you would want to show to your customers or manager.
Enter ChatGPT and similar large language model AI platforms. The ability to generate almost-human responses and content creation has been making headlines, but ChatGPT also turns out to be a surprisingly effective language translator!
Unlike earlier machine translation tools, ChatGPT does not look to translate each word, but rather tries to convey the same intent and information as the original text, in the target language. It also recognizes (sometimes, not always) when text is referring to a proper noun or a culturally specific idiom and adjusts accordingly.
Take this text from a spanish language soccer news site for example:
“Que Diego Armando Maradona pueda seguir o no al frente de Gimnasia y Esgrima La Plata, conjunto argentino de la máxima categoría al que salvó del descenso gracias en parte a la abrupta finalización de las competiciones por la irrupción de la pandemia, va a depender de que demuestre que no tiene cocodrilos en el bolsillo. O lo que es lo mismo, que es consciente de que es sensible con la nueva realidad económica.”
Google translate will correctly recognize the proper name “Diego Armando Maradona” and also the team name, “Gimniasa y Esgrima La Plata” when translating to English, but will mangle the end of the first sentence and most of the second with this English rendering:
“…will depend on what show that you have no crocodiles in your pocket. Or what is the same, that he is aware that he is sensitive to the new economic reality.”
ChatGPT on the other hand, recognizes “cocodrilos en el bolsillo” as an idiom, an illustrative figure of speech. As such it does not directly translate it but puts the phrase in quotation marks. It also more accurately recognizes that in the second sentence Maradona is not “aware that he is sensitive” but “aware of and sensitive to.”
Here is how it renders the same section — notice it also combines the two sentences into one, while conveying the accurate intent of the original Spanish sentences: “…will depend on whether he can demonstrate that he does not have “cocodrilos en el bolsillo” or, in other words, that he is aware of and sensitive to the new economic reality.”
Go one step further and ask ChatGPT to define “cocodrilos en el bolsillo” and it happily does so: “The Spanish idiom “Cocodrilos en el bolsillo” literally translates to “Crocodiles in the pocket” in English. This idiom is used to describe a situation that is dangerous or risky, particularly in a financial or business context. It implies that someone is taking a big risk that could lead to serious consequences or losses, similar to carrying live crocodiles in one’s pockets, which is obviously a very dangerous and foolish thing to do.”
Content localization is not just a matter of better raw translation — in the example above a literal translation of “crocodiles in the pocket” would be meaningless at best or confusing and distracting at worst to most English audiences.
By recognizing and adjusting for these cultural nuances, AI holds the promise of a future where we localize our training materials to suit new audiences, without needing to reduce or remove the stories, colorful descriptions and humor that might lost in translation today (but which also keeps our training from becoming dull and simplistic).
Other tools which can help us localize our training materials include:
- Synthesia: Create almost-lifelike videos with AI-based avatars reading from your scripted content. With support for 120 languages, it can easily create multiple video versions of your scripted content for each target audience. And a wide range of available avatars means you can choose to make your video host look more like their learners.
- Blanc: Have video content in one language but need it in another? The AI magic behind Blanc’s video translation tools can created dubbed versions of your videos, with multiple AI-based voices, almost instantly.
- Murf: Need to change a voiceover from one language to another? Upload your translated script to Murf for a natural sounding narration.
No tool is perfect, and there is still lots of room for improvement in the AI tools available for localization. But with literally thousands of new apps and AI platforms in the works, we are only at the beginning of this wave of AI solutions.
The best improvement may be for cultures who do not have as much mainstream support from existing translation tools. For example, I used to do some work in Rwanda and I have many friends who speak Kinyarwanda, the regional language in Rwanda. Not many tools support Kinyarwanda as a translation option — but ChatGPT does. And so it may be appropriate to end this article with me asking ChatGPT how to tell a Rwandan friend “The Best is Yet To Come”:
In Kinyarwanda, a Bantu language spoken in Rwanda and parts of the Democratic Republic of Congo, “the best is yet to come” could be translated as:
“Icyizere cyiza kizaza”
This literally means “Good things are still coming”. Note that Kinyarwanda is a tonal language, and correct pronunciation requires attention to pitch and tone.
