Mozilla has released a toolkit for self-contained machine translation from one language to another, running on the user's local system without recourse to external services. The project is being developed as part of the Bergamot initiative, together with researchers from several universities in the UK, Estonia and the Czech Republic, with financial support from the European Union. Developments are distributed under the MPL 2.0 license.
The project includes the bergamot-translator engine, self-training machine learning tools, and ready-made models for 14 languages, including experimental models for translating from English to Russian and vice versa. The level of translation can be assessed on the online demo.
The engine is written in C++ and is a wrapper around the Marian machine translation framework, which uses a recurrent neural network (RNN) and transformer-based language models. GPU can be used to speed up learning and translation. The Marian framework is also used to power the Microsoft Translator translation service and is developed primarily by Microsoft engineers in collaboration with researchers from the Universities of Edinburgh and PoznaΕ.
For Firefox users, an add-on for translating web pages has been prepared, which translates on the browser side without resorting to cloud services. Previously, the add-on could only be installed on beta releases and nightly builds, but it is now available for Firefox releases as well. In the browser add-on, an engine originally written in C++ is compiled into an intermediate WebAssembly binary representation using the Emscripten compiler. Among the new features of the add-on, the possibility of translating while filling out web forms is noted (the user enters text in their native language and it is translated into the language of the current site on the fly) and translation quality assessment with automatic marking of dubious translations to inform the user about potential errors.
Source: opennet.ru