Independent Scholar When a 'sport' is a person and other issues for NMT of novels

摘要

We report on a case study in which we as-sess the quality of Google's Neural Ma-chine Translation system on the transla-tion of Agatha Christie's novel The Mys-terious Affair at Styles into Dutch. We an-notated and classified all MT errors in the first chapter of the novel making use of the SCATE error taxonomy, which differ-entiates between fluency (well-formedness of the target language) and accuracy er-rors (correct transfer of source content). We modified the SCATE MT error tax-onomy to be able to annotate text-level phenomena such as textual coherence (e.g. anaphora and coreference) and tex-tual cohesion (e.g. lexical consistency) and literature-specific issues such as cultural references. Apart from annotating the er-rors in the MT output, we investigate how the machine translated version differs from the published human translatcd Dutch ver-sion of the book. We look at stylistic fea-tures such as lexical richness, cohesion, and syntactic equivalence.