Crossing Sentence Boundaries in Statistical Machine Translation

50 December 2016 Standard phrase-based statistical machine translation (SMT) systems translate one sentence at a time, completely ignoring discourse dependencies and the wider context of the document. As a consequence, words with multiple senses are often mistranslated when they are ambiguous in the local context. These translation errors decrease the quality of the translation, threatening the cohesion of the text. Research in discourse-aware SMT tackles document-level issues to improve the translation and to ensure that discourse features such as cohesion are maintained in the translation.