This article examines the point where half-century old machine translation systems freely available to the public reached in 2014. People have different impressions of machine translation as its quality largely varies with the languages, areas, text lengths and so forth the translation deals with. For example, we have no general-purpose highly accurate automatic translation system for Japanese-English readily available yet. On the other hand, we have highly accurate general-purpose automatic translation systems for Japanese-Korean. This article elaborates on statistical machine translation, the core technology, to help recognize the true state of today's automatic translations. Statistical machine translation works in such a way that it derives a statistical model necessary for translation from bilingual data and maximizes probability based on that model. Statistical machine translation offers some unique features including high accuracy in specialized areas and ease of multilingualization. We introduce high accuracy automatic translation by taking, for example, both a speech translation, which applies statistical machine translation to conversations in travel scenes, and a text translation, which applies the same to patent information.
View full abstract