We make distributed stochastic gradient descent faster by exchanging sparse updates instead of dense updates. Gradient updates are positively skewed as most updates are near zero, so we map the 99% smallest updates (by absolute value) to zero then exchange sparse matrices. This method can be combined with quantization to further improve the compression. We explore different configurations and apply them to neural machine translation and MNIST image classification tasks. Most configurations work on MNIST, whereas different configurations reduce convergence rate on the more complex translation task. Our experiments show that we can achieve up to 49% speed up on MNIST and 22% on NMT without damaging the final accuracy or BLEU.

We present Marian, an efficient and selfcontained 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.

Considerable effort has been made to address the data sparsity problem in neural grammatical error correction. In this work, we propose a simple and surprisingly effective unsupervised synthetic error generation method based on confusion sets extracted from a spellchecker to increase the amount of training data. Synthetic data is used to pre-train a Transformer sequence-to-sequence model, which not only improves over a strong baseline trained on authentic error-annotated data, but also enables the development of a practical GEC system in a scenario where little genuine error-annotated data is available. The developed systems placed first in the BEA19 shared task, achieving 69.47 and 64.24 F 0.5 in the restricted and low-resource tracks respectively, both on the W&I+LOCNESS test set. On the popular CoNLL 2014 test set, we report state-of-theart results of 64.16 M 2 for the submitted system, and 61.30 M 2 for the constrained system trained on the NUCLE and Lang-8 data.

We train a neural machine translation (NMT) system to both translate sourcelanguage text and copy target-language text, thereby exploiting monolingual corpora in the target language. Specifically, we create a bitext from the monolingual text in the target language so that each source sentence is identical to the target sentence. This copied data is then mixed with the parallel corpus and the NMT system is trained like normal, with no metadata to distinguish the two input languages.Our proposed method proves to be an effective way of incorporating monolingual data into low-resource NMT.On Turkish↔English and Romanian↔English translation tasks, we see gains of up to 1.2 BLEU over a strong baseline with back-translation. Further analysis shows that the linguistic phenomena behind these gains are different from and largely orthogonal to back-translation, with our copied corpus method improving accuracy on named entities and other words that should remain identical between the source and target languages.

This paper describes the University of Edinburgh's submissions to the WMT17 shared news translation and biomedical translation tasks. We participated in 12 translation directions for news, translating between English and Czech, German, Latvian, Russian, Turkish and Chinese. For the biomedical task we submitted systems for English to Czech, German, Polish and Romanian. Our systems are neural machine translation systems trained with Nematus, an attentional encoder-decoder. We follow our setup from last year and build BPE-based models with parallel and backtranslated monolingual training data. Novelties this year include the use of deep architectures, layer normalization, and more compact models due to weight tying and improvements in BPE segmentations. We perform extensive ablative experiments, reporting on the effectivenes of layer normalization, deep architectures, and different ensembling techniques.