Traducerea automată a literaturii. O himeră încă vie?

Artificial intelligence is about to reshape the world and redefine the relationship humans maintain with their professional activities. Recent machine translation systems, as well as the prior generation of statistical machine translation, rely on extensive aligned corpora and generate outputs whose quality can compete with human translations. This is particularly the case for technical and even medical texts. The role of the translator is to be questioned. Could algorithms potentially substitute this profession? Certain texts types, such as legal and literary ones, remain resistant to machine translation. The main obstacle for these systems is their lack of direct, tangible interaction with the outer reality from which these texts emerge, hence the systems’ inability to understand and interpret them. The ability of Transformer models is confined to analyzing intra-sentential context. In 2017, a new player entered the machine translation scene, DeepL, whose primary advantage lies in its integration of the Linguee database, an enormous parallel corpus of high-quality human translations. Generative natural language models include translation functionality. These models can be trained to identify or even replicate specific literary styles. Since literary works are replete with wordplay, extratextual references, and stylistic devices, we propose to investigate whether the task of decoding their message, particularly the author’s intent (understood as the addresser) via algorithms remains a chimera. Research efforts within the domain of natural language processing, such as the creation of parallel corpora, text “cleaning” and the morphological and syntactic annotation of texts (pre editing) have undeniably played a significant role in achieving a level of quality superior to that of rule-based or statistical systems. The results achieved in the field of automatic literary translation, however, do not meet the high demand upheld by human translation (biotranslation). How automatic translation systems are created and how do they function? What parallel corpora are? Are parallel corpora sufficient for generating high-quality automatic literary translations? What quality in translation is? Is there any evaluation metrics for human translations? And what about the metrics used to assess machine translations by humans? What is the current state of automatic translation for texts from classical Romanian literature? These are some of the questions we commit to answering in this paper.