I let AI translate my novel…: Approaches towards the automatic translation of literary fiction and the suitability of quality assessment methods

Abstract

While automatic translation technology is advancing rapidly and can be useful in certain contexts, it seems to fall behind when it comes to the translation of literary texts. This thesis explores to what extent three approaches to fully automated literary translation – commercial machine translation models, literature-specific translation models and large language models – are suitable for the translation of literary texts, and how existing translation and literary quality assessment methods compare when assessing machine-translated literary texts. In the experiment, excerpts from eight literary texts are translated by a commercial machine translation model (DeepL), a literary-adapted model (Combo) and a large-language model (GPT-3) and assessed using an automated assessment metric (BLEU), human assessment (PIE), corpus-based evaluation (lexical density and type-token ratio) and a literariness assessment (literary vector space). After analysing the results of these assessments, the literary-adapted machine translation model seems to provide the most promising results. None of the employed assessment metrics appear to be ideal for the assessment of machine translated literary texts, and further research into other types of assessment metrics is desirable.