The release of the OpenNMT Machine Translation System 0.28.0 (Open Neural Machine Translation) has been published, utilizing machine learning methods. The project employs TensorFlow's deep learning library for building the neural network. The code for the modules developed by OpenNMT is written in Python and is distributed under the MIT license. Ready-made models are available for English, German, and Catalan, while models for other languages can be created based on datasets from the OPUS project (for training, the system receives two files — one with sentences in the source language and the other with a quality translation of those sentences into the target language).
The project is being developed with the participation of SYSTRAN, a company specializing in machine translation tools, and a group of researchers from Harvard, who are creating human language models for machine learning systems. The user interface is maximally simplified and only requires specifying the input file with text and the file for saving the translation result. The extension system allows for implementing additional functionality based on OpenNMT, such as self-referencing, text classification, and subtitle generation.
The use of TensorFlow allows for the capabilities of GPUs to be utilized (to accelerate the training process of the neural network). To simplify product distribution, the project is also developing a self-sufficient version of the translator in C++ — CTranslate2, which uses pre-trained models without dependency on additional requirements.
The new version has added the initial_learning_rate parameter and implemented several new arguments (mha_bias and output_layer_bias) for adjusting the Transformer model generator. Other than that, it includes bug fixes.
Source: opennet.ru
