Using a multilingual literary parallel corpus to traint NMT Systems

Abstract

This article presents an application of a multilingual and multidirectional parallel corpus composed of literary texts in five Romance languages (Spanish, French, Italian, Portuguese, Romanian) and a Slavic language (Croatian), with a total of 142,000 segments and 15.7 million words. After combining it with very large freely available parallel corpora, this resource is used to train NMT systems tailored to literature. A total of five NMT systems have been trained: Spanish-French, Spanish-Italian, Spanish-Portuguese, Spanish-Romanian and Spanish-Croatian. The trained systems were evaluated using automatic metrics (BLEU, chrF2 and TER) and a comparison with a rule-based MT system (Apertium) and a neural system (Google Translate) is presented. As a main conclusion, we can highlight that the use of this literary corpus has been very productive, as the majority of the trained systems achieve comparable, and in some cases even better, values of the automatic quality metrics than a widely used commercial NMT system.