Neural machine translation of literary texts from English to Slovene

Abstract

Neural Machine Translation has shown promising performance in literary texts. Since literary machine translation has not yet been researched for the English-toSlovene translation direction, this paper aims to fulfill this gap by presenting a comparison among bespoke NMT models, tailored to novels, and Google Neural Machine Translation. The translation models were evaluated by the BLEU and METEOR metrics, assessment of fluency and adequacy, and measurement of the postediting effort. The findings show that all evaluated approaches resulted in an increase in translation productivity. The translation model tailored to a specific author outperformed the model trained on a more diverse literary corpus, based on all metrics except the scores for fluency. However, the translation model by Google still outperforms all bespoke models. The evaluation reveals a very low inter-rater agreement on fluency and adequacy, based on the kappa coefficient values, and significant discrepancies between posteditors. This suggests that these methods might not be reliable, which should be addressed in future studies.