Pre-trained multilingual representations promise to make the current best NLP model available even for low-resource languages. With a truly language-neutral pre-trained multilingual representation, we could train a task-specific model for English (or another language with available training data) and such a model would work for all languages the representation model can work with. (Except that by doing so, the models might transfer Western values into low-resource language applications.) There are several multilingual contextual embeddings models (such as multilingual BERT or […]

Over the few years when neural models are the state of the art in machine translation, the architectures got quite standardized. There is a vocabulary of several thousand discrete input/output units. As the first step, the inputs are represented by static embeddings which get encoded into a contextualized vector representation. It is used as a sort of working memory by the decoder that typically has a similar architecture as the decoder that generates the output left-to-right. In most cases, the […]

This week I am going to have a look at a paper by my former colleagues from Prague “Transforming machine translation: a deep learning system reaches news translation quality comparable to human professionals” that was published in Nature Communications. The paper systematically studies machine translation quality compared to human translation quality with the main criterion being the human judgment about the translations. Already in 2016, Google announced almost reaching human parity on their internal test sets. However, these results were […]

After a short break, MT weekly is again here, and today I will talk about a paper “CSP: Code-Switching Pre-training for Neural Machine Translation” that will appear at this year’s virtual EMNLP. The paper proposes a new and surprisingly elegant way of monolingual pre-training for both supervised and unsupervised neural machine translation. The idea is quite simple. The model they use is the standard Transformer; all the magic is how the model is trained. First, it is pre-trained on synthetic […]

This week, I will discuss Nearest Neighbor Machine Translation, a paper from this year ICML that takes advantage of overlooked representation learning capabilities of machine translation models. This paper’s idea is pretty simple and is basically the same as in the previous work on nearest neighbor language models. The paper implicitly argues (or at least I think it does) that the final softmax layer of the MT models is too simplifying and thus pose a sort of information bottleneck, even […]