Machine Translation

Machine translation (MT) is the automatic translation of texts from one human language into another. MT methods have evolved from explicit modeling of linguistic knowledge to increasingly data-driven approaches entirely based on machine learning, producing ever more fluent output while relinquishing detailed insight in the linguistic processes involved in translation. At the time of this article, the most successful MT methods are based on deep learning and are characterized by deep hierarchies of vector space embeddings, the use of neural attention mechanisms to propagate information and of subword decomposition to handle derivational morphology. These methods have high data requirements, which hampers their adoption in low-resource languages, and carry a risk of generating output not licensed by the input (hallucination).