Marian: Fast Neural Machine Translation in C++

By Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu T. Hoang, Kenneth Heafield, Tom Neckermann, Frank Torsten Bernd Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev and 2 more. 2018

797 citations recorded by the source corpus.

We present Marian, an efficient and self-contained Neural Machine Translation framework with an integrated automatic differentiation engine based on dynamic computation graphs. Marian is written entirely in C++. We describe the design of the encoder-decoder framework and demonstrate that a research-friendly toolkit can achieve high training and translation speed.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

Citations in the corpus

fairseq: A Fast, Extensible Toolkit for Sequence Modeling (2019, background): “MarianNMT (Junczys-Dowmunt et al., 2018) focuses on performance and the backend is written in C++ for fast automatic differentiation.”
Neural Machine Translation for English–Kazakh with Morphological Segmentation and Synthetic Data (2019, methodology): “We then tried different length-penalty values for the decoder (parameter normalize in Marian), using 0.9 (instead of the default 0.6) we reach 23.47.”
Parallel Corpus Filtering via Pre-trained Language Models (2020, methodology): “The quality of the resulting subsets is determined by training a neural machine translation system (Marian) 6 (Junczys-Dowmunt et al., 2018) on this data.”
Editing Factual Knowledge in Language Models (2021, methodology): “We employ English-to-German and German-to-English Transformer models from Marian Neural Machine Translation (MarianNMT; Junczys-Dowmunt et al., 2018) provided by Hug-gingface Transformers (Wolf et al., 2020).”
Findings of the 2018 Conference on Machine Translation (WMT18) (2018, background): “2.4.32 UEDIN (Haddow et al., 2018) For Estonian↔English and Finnish↔English, the UEDIN systems are an ensemble of four left-toright systems, reranked with four right-to-left systems, built using Marian.”
Why Self-Attention? A Targeted Evaluation of Neural Machine Translation Architectures (2018, methodology): “We use the bi-deep RNN-based model (RNN-bideep) implemented in Marian (Junczys-Dowmunt et al., 2018).”
Findings of the Third Shared Task on Multimodal Machine Translation (2018, methodology): “…types of models were submitted: AFRL-OHIO-STATE 1 2IMPROVE U is a system combination of the VMT system and an instantiation of a Marian NMT model (Junczys-Dowmunt et al., 2018), and AFRL-OHIO-STATE 1 4COMBO U is a systems combination of the VMT system along with instantiations of Marian,…”
Findings of the WMT 2018 Shared Task on Parallel Corpus Filtering (2018, methodology): “NMT For neural machine translation, we used Marian (Junczys-Dowmunt et al., 2018).”
FUDGE: Controlled Text Generation With Future Discriminators (2021, methodology): “G is a pretrained Marian (Junczys-Dowmunt et al., 2018)”
Code-Switching for Enhancing NMT with Pre-Specified Translation (2019, methodology): “Adam (Kingma and Ba, 2014) is used for training; warmup step is 16000; the learning rate is 0.0003.”
Gradient-based Constrained Sampling from Language Models (2022, methodology): “Terminology Constrained Translation We follow the setup in Dinu et al. (2019) and use an off-the-shelf English to German translation model by MarianMT (Junczys-Dowmunt et al., 2018) to translate a subset of WMT17 en-de test set (Bojar et al., 2017).”
One Question Answering Model for Many Languages with Cross-lingual Dense Passage Retrieval (2021, methodology): “Our translation models (to English and from English) are the pre-trained MarianMT (Junczys-Dowmunt et al., 2018) style OPUS-MT (Tiedemann and Thottingal, 2020) models available in Transformers library (Wolf et al., 2020) that are trained on the OPUS corpus (Tiedemann and Nygaard, 2004).”
Microsoft’s Submission to the WMT2018 News Translation Task: How I Learned to Stop Worrying and Love the Data (2018, methodology): “…marian-examples and perform the following steps: • preprocessing of training data, tokenization, true-casing1, vocabulary reduction to 36,000 joint BPE subword units (Sennrich et al., 2016) with a separate tool.2 1Proprocessing was performed using scripts from Moses (Koehn et al., 2007).”