Mozilla has released a toolkit for self-contained machine translation from one language to another, operating on the user's local system without relying on external services. The project is being developed under the Bergamot initiative in collaboration with researchers from several universities in the UK, Estonia, and the Czech Republic, with financial support from the European Union. The developments are distributed under the MPL 2.0 license.
The project includes the bergamot-translator engine, tools for self-training machine learning systems, and ready-made models for 14 languages, including experimental models for translating between English and Russian and vice versa. The quality of translation can be evaluated in an online demonstration.
The engine is written in C++ and serves as a wrapper over the Marian machine translation framework, which employs recurrent neural networks (RNN) and transformer-based language models. For faster training and translation, a GPU can be used. The Marian framework is also used to power the Microsoft Translator service and is primarily developed by engineers at Microsoft in collaboration with researchers from the University of Edinburgh and the University of Poznań.
For Firefox users, an extension has been prepared for translating web pages, performing translations on the browser side without relying on cloud services. Previously, the extension was only available in beta builds and nightly versions, but now it is accessible for releases of Firefox as well. In the browser extension, the engine, originally written in C++, is compiled to an intermediate WebAssembly binary representation using the Emscripten compiler. New features of the extension include the ability to translate while filling out web forms (the user enters text in their native language, which is then translated on the fly to the current site’s language) and a quality assessment of translations with automatic marking of questionable translations to inform users of potential errors.
Source: opennet.ru
