Context-Aware Coordination of Cascaded Machine Translations

Recently more and more people communicate through the Internet, and machine translators are used as communication tools between peoples who cannot speak their partners' mother language. Since English is a hub language of development of language resources like machine translators or bilingual dictionaries, combining multiple machine translators via English enables intercultural communication and collaboration between non-English languages. To combine multiple machine translators, problems interfering communication sometimes occur. Cascaded machine translators often yield mistranslations due to inconsistency of word selections, even if all translators are combined correctly and each translation result is correct. This is because each translator considers each input sentence only. This phenomenon is a big problem for both multi-hop translation cascading multiple translators and machine translator-mediated communication. Then, coordination of cascaded machine translators is needed. For resolution of such problems, this research addresses following issues. In order to examine whether the sense of translated sentence is different from the one of the source sentence, equivalent terms of all languages is required. Equivalent terms between two languages is developed as bilingual dictionaries between lots of languages, while that of more than three languages is developed manually among parts of languages. Therefore this research aims to generate multilingual equivalent terms automatically from existing language resources. Coordination of translators, that is, consistent word selections, can be realized by extracting context and propagating it to machine translators. Context is extracted from the source sentence or whole document including the source sentence. The sense of translated sentence is kept consistent by ii selecting translated words which suit propagated context. Methods of extracting context are proposed in previous researches. This research assumes that the context is already extracted, and focuses on the coordination by propagating extracted context. To solve these two issues, this research proposed following solutions. This research proposed the method to obtain multilingual equivalent terms by combining multiple bilingual dictionaries. Relations between words and translated words are represented as a graph, and equivalent terms are obtained by using the structure of it. If simply combining multiple dictionaries, there are some cases where multiple terms which do not share the same sense are also obtained. Such error can be prevented by considering the structure of the graph. This research proposed a method to coordinate cascaded machine translators so as to select translated words based on the context which was propagated using multilingual equivalent terms. Each context extracted from the sentence is represented in …