Author-Tailored neural machine translation systems for literary works

Abstract

In this chapter, the process of training and the evaluation of neural machine translation (NMT) systems tailored for a given author for the English–Spanish language pair is described. The goal of the system is to create bilingual e-books for one work of a given author. These bilingual e-books allow the user to go from one sentence in the original to the translated version and come back to the original version via a simple click on the text. These e-books are intended for students of the source language willing to read the original version of the work. To create the training corpus, a novel technique allowing the selection from a very large parallel corpus of the most similar parallel segments to the source segments of the novel being translated was used. With the selected segments, an NMT system is trained and evaluated using three strategies. Several automatic metrics were calculated, and the trained system is compared with Google Translate and DeepL. Then, some fragments of the selected work are post-edited by several translators and the post-editing effort is calculated. Lastly, the opinions of readers of the bilingual edition are obtained via a questionnaire.