Exploring Contextual Information in Neural Machine Translation

This works explores means of utilizing extra-sentential context in neural machine translation (NMT). Traditionally, NMT systems translate one source sentence into one target sentence, without any notion of the surrounding text. This is clearly insufficient and different from how humans translate text. For many high-resource language pairs, translations produced by NMT may be under certain, strict conditions, nearly indistinguishable from human produced translations. One of these conditions is that evaluators score the sentences separately. When evaluating whole documents, even the best NMT systems still fall short of human translators. This motivates the research of employing document level context in NMT, since there might not be much more space left to improve translations on the sentence level, at least for high resource languages and domains. This work summarizes recent state-of-the art approaches to context utilization, implements several of them, evaluates them both in terms of general translation quality and on specific context related phenomena, and analyzes their advantages and shortcomings. A hand-made context phenomena test set for English to Czech translation was created for this task.

NMT, neural machine translation, context, recurrent neural networks, transformer, document level translation, discourse, cohesion, coherence