Title: Machine Translation of Novels in the Age of Transformer

Abstract: In this chapter we build a machine translation (MT) system tailored to the literary domain, specifically to novels, based on the state-of-the-art architecture in neural MT (NMT), the Transformer (Vaswani et al., 2017), for the translation direction English-to-Catalan. Subsequently, we assess to what extent such a system can be useful by evaluating its translations, by comparing this MT system against three other systems (two domain-specific systems under the recurrent and phrase-based paradigms and a popular generic on-line system) on three evaluations. The first evaluation is automatic and uses the most-widely used automatic evaluation metric, BLEU. The two remaining evaluations are manual and they assess, respectively, preference and amount of post-editing required to make the translation error-free. As expected, the domain-specific Transformer-based system outperformed the three other systems in all the three evaluations conducted, in all cases by a large margin.