Browser-based multilingual translation without the cloud
Automatic web page translation has emerged to meet the increasing need to find or use information on the Internet in a language not one’s own. Most web browsers now have built-in translation tools. The text is sent to a server in the cloud. This means privacy becomes an issue, including data and viewing history privacy. The EU-funded Bergamot project has created free translation software that runs locally on the user’s machine as a browser extension for open-source Mozilla Firefox. Now corporate and government workers can more easily adopt machine translation without compliance issues or the purchase of an on-site server. Bergamot has unlocked safe and secure machine translation for all.
Automatic web page translation: it takes two to communicate
Installing the Bergamot browser extension not only enables web page translations that make information in multiple languages accessible to any user. Its algorithms support bidirectional live translation. It can automatically translate input from a user written in the browser’s language setting to the language detected on the web page, as in filling a form on a page in another language. It will also translate any new text that appears such as a response to the user in a non native language in a chat window. Kenneth Heafield of the University of Edinburgh and Bergamot project coordinator explains: “For example, the form to file a complaint to German bank regulator BaFin is in German. We made an assistant that translates the questions to English, allows one to type responses in English, and translates and enters these into the form in German.”
Local machine translation: neural network efficiency and dynamic domain adaptation
Creating a local translator even better than a cloud-based one in addition to being private required ingenuity. Consumer hardware is much less powerful than cloud servers. “The key was combining many approximations and tuning their application to maximise speed and minimise loss of quality. We wrote some of the software in machine code and simplified the model’s structure. For example, instead of floating-point values, we use integers,” states Heafield. The first time a particular language is translated, the browser extension downloads an optimised neural network model, about 15 32 megabytes compared to 800-3 200 megabytes for a typical unoptimised model. Then the optimised code runs the model locally. Translation is significantly enhanced with dynamic domain adaptation that considers the domain or context of a word. For example, ‘stock’ takes on different meanings in commerce, finance and even agriculture (livestock). As Heafield notes, “some companies will customise a machine translation system for a particular domain if a specialised client is willing to pay for this service. Dynamic domain adaptation performs a lightweight version by looking at the domain of the text being translated.”
Real-time automatic web translation with unprecedented privacy, speed and accuracy
“The trend is to train ever-larger models, but deploying these models is hard because they are so computationally expensive to run. Significant quality can be preserved by distilling a smaller model, followed by optimising it. Then, local translation protects privacy because text never leaves the computer,” notes Heafield. The prototype has 32 004 users and counting. You can see it in action and try it out here
Keywords
Bergamot, translation, web page, browser extension, dynamic domain adaptation, neural network
