Automated language translation is an example of how machine learning is poised to revolutionize our day-to-day lives. Yet, current systems have yet to deliver on the promise of error-free translation of idiomatic expressions or more nuanced phrasing, and do not easily allow for incorporation of domain specific terminology. Machine learning approaches, and in particular neural networks with sequence-to-sequence architectures and attention mechanisms, are posed to bring about a radical shift in the status quo.
To achieve this end, it is important to have a very rich dataset of text in many languages so as to be able to train the models. The model learns how to code input sentences in numbers that represent semantic meaning and context, and later on decode them in the target languages. Attention mechanisms make the model more capable in focusing on relevant parts of sentences thus enhancing its accuracy and context retention.
Training requires tuning a wide variety of model parameters and evaluating performance on non-trivial metrics for translation quality. After building a translation model, many hours of engineering are still needed to construct a translation system that will automatically translate sentences between human languages with high performance over a wide range of topics and sentence types.11-13 However, once built, such a translation system would be able to massively and quickly produce high-quality translations from one language to any other language without needing any other form of help or data, such as parallel texts (e.g., machine-translated government documents), bilingual dictionaries (e.g., Wik­ tionary), comparable corpora (e.g., search query logs), information about the world (e.g., Wikipedia), or even monolingual text in either the source or target languages.