Human-Adapted MT for Literary Texts: Reality or Fantasy?

This article presents the evaluation of a bespoke machine translation (MT) system adapted to the work of a literary translator who specializes in fiction. Actualizing a research avenue for the English–French pair and complementing a growing number of studies for other language combinations, our work on literary machine translation shows that it is possible to fine-tune a base model trained with publicly available tools and datasets and obtain interesting results with a much smaller subset of custom training data. Although the raw unedited output remains nowhere near human production, metrics show that the adapted system produces translations that are much closer to the desired human reference, while further automatic evaluations reflect improvements regarding important textual features for literary texts. Finally, a manual evaluation details the type of errors produced by our system, indicating that these primarily involve adequacy errors that are already well documented for MT and typically harder to spot, rather than purely literary aspects. This annotation process also revealed unsuspected
observations and suggests new perspectives for the use of MT. Namely, that of an individually personalized and interactive tool better suited to the translation of creative texts, although its application to the literary domain opens many ethical and practical questions.