Mozilla has added an official translation tool to Firefox that doesn’t use cloud processing to do its job, instead running a machine learning-based process right on your computer. This is a huge step forward for the popular service, which is closely associated with giants such as Google and Microsoft.
Translation tool called Firefox translations, can be added to the browser here. It will need to download some resources the first time it translates a language, and presumably it can download improved models if needed, but the actual translation work is done on your computer, not in a data center a couple of hundred miles away.
This is not important because many people need to translate in their browsers offline – like a screen door for a submarine, it’s not exactly a smart use case. Rather, the goal is to reduce the end dependency on cloud providers with ulterior motives for a task that no longer requires their resources.
This is the result of the EU-funded Bergamot project., in which Mozilla collaborated with several universities on a set of machine learning tools that will make offline translation possible. Usually such work is done by clusters of GPUs in data centers where large language models (in gigabytes and with billions of parameters) will be deployed to translate the user’s request.
But while Google and Microsoft’s cloud tools (not to mention DeepL and other upstart competitors) are accurate and (due to almost unlimited computing power) fast, there is a fundamental privacy and security risk when sending your data to third parties. batches for analysis and sending back. For some, this risk is acceptable – others would prefer not to involve the Internet giants if it is not necessary.
If I translate the menu in a tapas bar using Google, will I start advertising sausages? More importantly, if someone is translating immigration or medical documents with a known device ID and location, will ICE knock? Doing it all offline makes sense for anyone worried about the privacy implications of using a cloud provider to translate in any situation.
I quickly checked the quality of the translation and found it to be more than satisfactory. Here is an excerpt from the front page of the Spanish-language news agency El Pais:
Quite normal! Of course, El Pais translates to “Paris” in the tab title, and there were plenty of other dubious phrases (although he translated each | as “Oh, that’s good” – pretty hilarious). But little interfered with the understanding of the essence.
And, ultimately, that’s what most machine translation is designed to do: communicate the underlying meaning. For any nuance or subtlety, even a large language model may not be able to reproduce the idiom, so a real bilingual person would be the best choice.
The main limitation is probably the lack of languages. Google Translate supports over a hundred – Firefox Translations even supports a dozen: Spanish, Bulgarian, Czech, Estonian, German, Icelandic, Italian, Norwegian Bokmal and Nynorsk, Persian, Portuguese and Russian. This overlooks quite a bit, but remember that this is just the first release of a project by a non-profit organization and a group of scientists, and not the preeminent product of a multi-billion dollar internet empire spanning the world.
In fact, the creators are actively asking for help from disclosure of the training pipeline allow “enthusiasts” to train new models. And they also ask for feedback to improve existing models. This is a useful product, but far from finished!
Credit: techcrunch.com /