The world of research that brings machine translation closer to "true translation"

Machine translation (MT) technology has been studied as "a problem that converts the text of starting language into a target language text."However, in fact, it is impossible to translate just by equivalent to the content of the text.For example, when translating product manuals, patent application documents, corporate press releases, etc., the specifications of punctuation specifications specified according to the published media, how to translate existing terms, how to translate existing terms,It is necessary to properly refer to the information in.

With the improvement of the quality (in the sense of text conversion) of MT technology, the work process has been used in the field of translation as an industry to correct and deliver MT output.In the process of this correction, two types of modifications are made: (1) modification of text conversion mistranslation by MT and (2) corrections in light of the information on the outside text.

I collected MT modification cases, based on the English -Japanese translation of news articles, and analyzed the types of information outside the text to refer to for correct translations.These are roughly divided into the four types shown in the table.

(A) is an example that indicates the necessity of using the text, but it also needs to follow detailed specifications, such as phrases, punctuation methods, and symbols.(B )'s MT translation seemed to be correct, but it was a translation because it was an article about the Pope.(C) was a reference about the pitcher's record of the US League, so the victory or loss is correct.(D) is an example where the ambiguity of the expression in the starting document (brother or brother) could be eliminated.The modified person was able to solve this ambiguity for the first time by obtaining (reliable) documents related to the person in mention.

In order to bring MT closer to a true translation, we are currently conducting research on technologies that appropriately refer to such information.

Nikkan Kogyo Shimbun November 9, 2021