Realization of Pivot Translation System Using Multilingual Synsets

Recently, internationalization and the spread of the Internet are increasing our chances of participating in intercultural and multilingual collaborations. This trend increases the number of multilingual groups where the native languages of the members differ and increases the demand for machine translation (MT) between non-English languages. However, because of the limited availability of such MTs, pivot translation using English as a hub language is a realistic way. In pivot translation, the second translation result might have the different meaning from the first translation result because of the polysemous nature of words in the intermediate language. As a result, there is a possibility that the final translation result has the completely different meaning from the source sentence. The previous research calls this feature intransitivity of pivot translation. This research tried to resolve the problem of intransitivity by creating the set of trilingual equivalent terms (called triples) with bilingual dictionaries and ensuring consistency of the concept of translation words. This research achieves some progress, but there are still remaining problems as follows. Further deterioration of the translation quality The previous research tried to modify translation errors with the set of triples, but the quantity of triples in the set is not sufficient. Therefore, even if the original result of pivot translation generates correct 3-tuple of translation words, such a translation tuple might not be included in the extracted triple set. As a result, the translation result might be modified by a contextually incorrect triple and its quality might be deteriorated further. Dependency on language resources between non-English languages In the previous research, the set of triples is created by using bilingual dictionaries among three languages. Generally speaking, qualities and quantities of bilingual dictionaries between non-English languages tend to be less enriched because of their low demands than those of dictionaries between English and a non-English language. Utilization of dictionaries between non-English languages undermines ii versatility of the previous research. To resolve the former issue, two types of triples are defined in this research, lower bound and upper bound against the ideal set of appropriate triples. It is ensured that words in triples of lower bound have certain common concepts, on the other hand, upper bound contains all appropriate triples definitely while it might include some errors. Namely, 3-word-tuples which are not included in upper bound do not have any common concept. The author proposes pivot translation algorithm using these features of two …