From "Crime and Punishment" to CLARIN:EL

To celebrate the International Translation Day (30 September 2023), we dedicate this month to resources which are hosted in the CLARIN:EL Infrastructure and constitute members of the Literary Translation Resource Family. One of these resources is the translation of Dostoievski's Crime and Punishment by Papadiamantis. We conducted a very small experiment, by comparing a small abstract as it was translated by Papadiamantis and as translated by ChatPGT.

We asked Titika Dimitroulia, Professor of Translation Theory and Practice at the Department of French Language and Literature of the Aristotle University of Thessaloniki, translator and literary critic, to comment on the above experiment and to talk about literary translation as a process, the skills that the translator should have, but also the possibilities and limits of artificial intelligence:

"Translating literary texts is a complex process. George Steiner may have said that the basic logic of translation is the same whether somebody is translating Dante's Paradise or a bill of lading, but in reality each type of translation has its own principles and norms. So, in literary translation, the point is not to imitate a work in another language. The translator has to understand the text in depth within its context and the challenges that arise in order to integrate it into a completely different spatio-temporal, linguistic and cultural context. They should understand precisely the linguistic, stylistic and other choices of the author and select from the range of options available in their own language in order to recreate them. Literary translation therefore requires a wide range of skills, goes far beyond mere linguistic and cultural transfer and is situated in the realm of aesthetic creation.

Literary translation experiments with machine translation systems in the context of translation studies research have completely failed. The process of editing a literary work that has been automatically translated, for example, is not only a laborious and time-consuming process, but also obstructs the creativity of the translator-editor. Modern machine translation systems already incorporate artificial intelligence (AI), but the rapid developments marked by the availability to the general public of AI systems such as chatGPT that, among other things, translate, bring back, from a different perspective, the question of the relation between AI and literary translation: can AI systems take on this complex task with worthwhile results? And if so, are there any translators whose work is of unique value and will remain through time? 

A first answer to this question could be given by the two translations of Dostoievski's Crime and Punishment quoted above. The abstract of Papadiamantis' translation comes from the resource Alexandros Papadiamantis: 4 translations which contains the following translations of this great Greek writer: Crime and Punishment (F. Dostoievski), The Maxman (Hall Caine), Doctor Rameau (Georges Ohnet) and Tartarin of Tarascon (Alphonse Daudet). The resource was originally created as part of a research project conducted jointly with my colleagues Eleni Politou-Marmarinou and George Mikros to test whether and to what extent certain anonymous translations in newspapers could be attributed to Papadiamantis (attribution of paternity) using hyphometric techniques. In this context, a system had to be trained to identify Papadiamantis' style, through various text corpora, works and translations, and to distinguish it from that of his contemporaries - that is, to discern which elements were part of the collective organization of the discourse of the period and which were his personal choices. You can access the results of this research here.

Beyond this research, however, as in Greece we do not have a corpus of translated texts, such as the Translational English Corpus (TEC), it was clear that this limited corpus could be used for research and education in a variety of ways. It is of particular interest, for example, that the translation of Dostoievski, unlike the other three, is indirect, mediated, i.e. it is from French and not Russian. On the one hand, two subcorpora, Dostoievski's translation and all the others, can be studied comparatively, with the aim of identifying different approaches to indirect and direct translation by the translator, if any. It may also be possible to study this particular translation in relation to other indirect translations of the period, given that they were particularly widespread and the intermediate language was most commonly French. But it is also possible to study this indirect translation in comparison with other direct translations in terms of aesthetic effect, in the context of the discussion of bad indirect and good direct translations, conducted in terms of stereotypes. Finally, in addition to any other hyphometric research, the relation between original and translated speech can also be studied. All this, insofar as someone does not create a parallel corpus of texts, including the original texts and the translations - and in Dostoievski's case it could be a triple corpus, i.e. including the French translation, assuming we manage to locate the translation from which the Greek translation was made. In this case, many other issues can be studied, both research-wise and in the classroom.

To give you an example, as in the context of various courses, my students and I have seen abstracts of this translation compared to other Russian translations, one of which is the well-known, excellent translation of Aris Alexandrou, it is striking how the Papadiamantis' text remains unsurpassed in its level of recreation of Dostoievski's text compared to direct translations, on the level of the ethos and style of the Russian work - thus debunking the myth of direct and indirect translation to which I referred above.

At the same time, the parallel text corpora, which today the general public, through the systematic use of systems such as Linguee, considers to be something like dictionaries, are in fact a set of examples of how a translator has faced specific difficulties: students have much more to learn from these examples than from any dictionary, since they see the translation process in practice. We hope to post a parallel corpus of Papadiamantis' translations in the near future, while integrating the French translations into the FREL parallel corpus of texts, which we are currently working on again, creating a new interface and full-text accessibility for works which have no copyright restrictions.

I will sum up with a reference to the translation provided by ChatGPT. I don't think one needs to think too hard to choose between the two translations. AI is trained on large language models, and with literary texts, but it can in no way manage meaning, let alone style. In reality, AI correlates and regurgitates data. So it makes sense to remind ourselves that at a time when we have not decoded human intelligence it is a joke to talk about simulating it. And the definition given by the mathematician and cognitive scientist Daniel Andler in his latest book on human intelligence (Intelligence artificielle, intelligence humaine: la double énigme, 2023), that it is the ability to respond to a specific situation, in which there are infinite variables, which a human being organizes, mobilizing all the necessary means, to finally arrive at the problem and choose its solution, fully illuminates the process of literary translation in its relation to artificial intelligence."