A comparison of statistical and neural machine translation for Slovene, Serbian and Croatian

Abstract

In this paper we present a comparison of translation quality using of Statistical Machine Translation (SMT) and Neural Machine Translation (NMT), considering translation directions between English, Slovene, Serbian and Croatian. Our experiments show that on a reduced training dataset with around two million sentences, SMT outperforms the NMT neural models. Furthermore, we present experiments with enlarged neural architectures, using 1,000 nodes and 4 hidden layers, which shows improved translation quality in terms of the BLEU metric