Lokalise introduced their beta AI implementation. I expected CATs to do better with AI. Why was I not impressed?
While I can imagine playing nerdishly with any AI feature, I can’t imagine a language professional fussing about it in their everyday efforts. Honestly, I believe that ChatGPT technology is amazing and that it’s going to be a game changer in every language-oriented industry. What I missed, was any sense of purpose of its Lokalise implementation. Or maybe I have sensed the purpose and I was not at all happy about what I saw. It seems that the Lokalise AI implementation is NOT intended for professional linguists.
IT DOES NOT BOOST PRODUCTIVITY
Dear CAT producers: Each time you say “just a single click of a mouse,” a kitten dies.
Or my hand ends in pain. Mouse is NOT the tool of choice for language professionals. It’s keyboard shortcuts. Professional work environment is not mouse based, as it cripples translator’s hands and forces translators to constantly leave the keyboard, move hands there and back, in result delaying their work.
Now back to Lokalise and the AI feature: You’re working in Lokalise. You open a string and you can immediately translate, or use a translation memory match, or one of three machine translation suggestions. As a professional translator, you’ll have the translation finalized in seconds if the string is short, maybe in one or two minutes.
Now with the new option, you can also click the AI feature (it does not offer it’s suggestion immediately, it probably can’t be that fast). You wait a moment for AI to generate it’s suggestion. Then you can also ask it to generate variations. So you click Variations and wait another moment to enjoy a set of new suggestions.
AI is, unlike MT, context aware. It can even work with your glossary (which is great, no irony). To add context, you click another option, write the context, click to save and have AI generate a new sentence. Click. You can ask AI to make it shorter. Click. You can set a character limit. Click, click. You can ask it to SEO optimize. Click. (No, the keywords suggested and implemented by ChatGPT will NOT be as good as those selected by a SEO optimization expert, tested in person.)
The only thing you can’t do is understand why spend five or ten minutes clicking and fussing about a single string? Well, some clicking will be avoidable, as Lokalise is preparing a method for providing AI with context in batches, so you don’t have to define that you want a button text each time. They are also working on AI batch pretranslation. OK. But otherwise?
Maybe sometimes when your memory tricks you and the good wording for a specific sentence is not coming, you can ask AI to step in. But what about the hype about replacing translators?
But here’s the key: “As a project manager, how can you feel confident about approving a translation suggested by AI if you don’t speak the language?”
That question was indeed asked during the webinar. It felt like CAT producers are trying to sell the tool to unsuspecting customers, pretending to sell a tool producing ready-made translation products. Not computer-aided translation anymore. Just computer translation.
‘No, actually you can’t.’ Well, this honest answer never came.
The actual reply was: “…but with human translations it’s the same. If you don’t speak the language, it’s pretty hard to validate the quality of it.” Miguel Caetano, Lokalise Product Marketing Lead said, and continued about a QA tool Lokalise is yet to create, “so you’ll be very quickly able to understand the quality of the translations.”
Like, REALLY?
Fun fact: The currently non-existent QA tool is to be based on the same AI. So the AI will generate some translations, then check them and return the user some metrics on how well it thinks it was doing, right?
And, finally, the speaker added: “… It’s always good to have a human reviewing the content,” mentioning that it could be anyone from the client’s team speaking the language.
Buy the “ultimate do-it-all” tool, it will perform the task alone and then check its own quality, so that you it can show you a nice report saying that it did a really nice job. You can, of course, use an expensive plugin called the “human”. Doesn’t have to be an expert, “human” is enough of a qualification. Forget about technical experts, proofreaders, copywriters. For my next job, I’ll ask my kids or my neighbour for their input. Go HUMANS!
How are you doing, Hal?
P.S.: The accuracy rate achieved by ChatGPT was 85%+ but it did much better in some specialized tests, e.g. in medicine. Its translation success rate varies by language combination.