Neural Machine Translation and Finnish Case-Inflections: Translation Problems and Pre-editing Possibilities

Abstract

Machine translation has slowly been becoming more prevalent from the 1950s onwards with the most recent leap in development being neural machine translation, which has quickly been de-ployed in professional translation settings with human translators post-editing machine translated text. Neural machine translation is still not a perfect form of machine translation. One of the problems when translating between English and Finnish is the Finnish case system, as the Finn-ish case-inflections can be problematic for machine translation engines. This thesis presents an evaluation of the four most prominent neural machine translation en-gines’ capability of translating case-inflections accurately from English to Finnish and proposes pre-editing or pre-translation being a solution for the problem. The four neural machine transla-tion engines that are featured are Google translate, Microsoft translator, Amazon translate, and DeepL translator. Samples were gathered from news media, a fictional book, and a scientific book to account for differences in textual style. These samples were then given as input to neural machine translation engines and the samples’ translations were given a numerical score on a seven-levelled evalua-tion scale. The samples that were assigned scores which were deemed unsatisfactory were select-ed and the possible reasons for their inaccuracy was examined. The further examination showed that the frequency in which a case-inflection occurs predicts it’s accurate usage by a machine translator to some degree. Unconventional syntax structure and longer sentence length were also shown to be predictors for accuracy problems when translating with a machine translator. The results suggested that pre-editing would not be more efficient than post-editing.