Looking Around: Challenges and Advances in Contextual MT
Over the last years, machine translation (MT) has been on the spotlight as one of the key problems in natural language processing (NLP) that deep learning has made tremendous progress. We have moved from the incoherent ramblings of statistical MT to the suprisingly fluent and adequate translations produced by neural models. These systems (TODO: capabilities of neural systems)
However there is key aspect where this models are still suprisingly miopic: the use of surrouding context to improve translations. Humans rarely translate sentences in a stand-alone fashion. When translating a book, translators use the sentences from the same paragraph, chapter or even the whole boook to help disambiguate words that could otherwise have multiple translations (TODO: example) Despite this, state-of-the-art MT systems are, in general, sentence-level: for example, to translate a document, the system would split the document into sentences, translate each sentence, and concatenate the translations to produce the translated document.