Naslov
Statistical and Neural Machine Translation: Changes in the MT-Output from Swedish into Croatian

Autori
Ljubas, Sandra

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Translation, Interpreting and Culture: Old Dogmas, New Approaches (?): Book of abstracts / Tyšš, Igor ; Hutkova, Anita ; Höhn, Eva - Banska Bistrica : Belianum, 2018, 75-76

ISBN
978-80-557-1435-6

Skup
Translation, Interpreting and Culture: Old Dogmas, New Approaches (?)

Mjesto i datum
Nitra, Slovačka, 26.09.2018. - 28.09.2018

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
statistical machine translation, neural machine translation, Google Translate, error analysis, the Swedish-Croatian language pair

Sažetak
The aim of the present study is to examine whether Google Translate’s recent shift to a neural model has had a significant impact on the quality of machine-translated texts regarding the Swedish-Croatian language pair. In November 2016, a preliminary study was conducted, in which the author analysed the errors found in translations produced by Google Translate, who was using a statistical translation method at the time (Ljubas 2017). The analysis showed: a) that the produced Croatian translations contained a substantial amount of morphosyntactic errors, and b) that lexical errors had a great impact on the level of (un)intelligibility. Specifically, GT failed to correctly render the Swedish formal subject “det”, phrasal verbs, reflexive pronouns, double negation, abbreviations, compounds etc. Shortly after the study had been completed, however, Google started applying a new neural method capable of deep learning, namely the Neural Machine Translation (NMT) model, in hopes of improving the quality of machine-produced translations. Currently, GT’s NMT model only supports translation from and into English, but GT has traditionally used English as an interlingua for other language pairs and a significant improvement in the output data has indeed been recorded although “the NMT system creates its own cryptic ‘interlingua’” now (Hurskainen 2018). The newly machine-translated texts contain fewer errors, but the results indicate that GT either omits or leaves a considerable number of words untranslated. Generally, the translations from Swedish into Croatian are now morphologically more correct and intelligible, yet less precise and still not acceptable as end-products. This stands in correlation with the expectations and findings from previous studies on the subject (Popović 2015, Wu et al. 2016, Castilho et al. 2017).

Izvorni jezik
Engleski

POVEZANOST RADA