Abstracts

Facebook/Meta is a social network that systematically offers the translation into various languages of a variety of original content posted by Internet users, thanks to a convolutional neural network (CNN). Like many of its competitors, Meta highlights the results produced by its model, namely the dissemination of information in real time throughout the world, as well as its accessibility and democratisation, by sharing it since 2022 in open source.

In times of conflict, journalists and citizens have been turning to social media to obtain information on theatres of operation that are not well covered by news agencies. In 2017, Facebook launched the option to read these posts in automatic translation. Although the risks of misinformation are numerous, they are offset by the supposed reliability of the source, which is often picked up a few days later in the traditional media, then translated by a human translator, professional or not. The content often comprises war literature or a form of poetry that has been termed 'poetry of extremity' -- eyewitness accounts, pleas, poetry of despair or resilience. These texts are literary because they express a singular voice - often that of righteous indignation - and imbued with many erudite reference. They demand to be read but seem to resist translation, at least by the Facebook algorithm. This raises the question of whether this device works as well as the social network seems to claim. This qualitative study proposes to examine this phenomenon located at the border of literature and translation. We will look in turn at a testimony from China in 2022 to commemorate a dissident, then at a pamphleteering post condemning non-interventionism, and at poems written at the beginning of the war in Ukraine.

Isabelle Chauveau, Loïc De Faria Pires - Université de Mons

Post-editing NMT in the classroom: the antithesis of feminist literary translation?

This proposal concerns a multidisciplinary study linking post-education teaching and feminist literary translation. Literature can be seen as a vehicle for women's identity construction (Fourgnaud, 2017: 5). This seems to be particularly true for fairy tales (Schanoes, 2014: 1). Therefore, the gendered translation of such works should enable the receivers of the target text to fully benefit from the works in the target language. Gender bias seems to be a frequent MT problem (Rabinovich et al., 2017: 1074; Prates et al., 2019: 1), regardless of language pairs and engines (Stanovsky et al., 2019: 1679). Thus, gender biases, when not properly post-edited, lead to the discrimination of women (Vanmassenhove et al., 2018: 3003).

We conducted a practical study with 60 male and female first-year translation students at the University of Mons. We provided them with a 300-word excerpt from Angela Carter's (1979) feminist rewriting of the Bluebeard tale, entitled The Bloody Chamber. This extract was automatically translated into French by two commercial engines: Google Translate and DeepL. Half of the cohort were asked to post-edit the DeepL version and the other half the Google Translate version. Beforehand, we compiled a list of relevant gender biases produced by the two engines. We then analysed how the students post-edited these items. Finally, the students filled in a questionnaire related to the EP strategies used.

During our presentation, we will present the data obtained from these questionnaires, which we will put into perspective with the post-edited texts, in order to determine whether the content produced by the students corresponds to their declared strategies. We will also determine whether one of the two engines leads students to produce better texts (in terms of genre) than the other. Finally, we will conclude with a series of future perspectives related to gender issues in the field of post-publishing NAT.

Daniel Henkel - Université Paris 8 Vincennes St. Denis - TransCrit EA1569

A Corpus-based Study of Target-language Norms/Deviation in Human and Machine Translation 

This is the first large-scale corpus-based study of both human and machine literary translation.

Original works and translations which were both in the public domain (late 19th-early 20th c., ca.1865-1930) were collected on a 1-work/author-translator basis so as to produce a “bidirectional” (Johansson 2007) corpus consisting of 4 subcorpora: Original English (35 authors, 3.5 m. words), Original French (35 authors, 3.4 m. words), English-translated-from-French (35 translators, 3.3 m. words) and French-translated-from-English (35 translators, 3.7 m. words). All source/target pairs were aligned at sentence-level and tagged for part-of-speech and lemma with TreeTagger. At 140 authors/translators and over 14 m. words total, this is currently the largest bidirectional, aligned, linguistically-annotated corpus of English and French and the largest English/French literary corpus. All original texts were translated a second time in late 2021-early 2022 using the latest neural-MT version of DeepL Pro.

The bi-directional model allows comparisons to be made, not just between source- and target-texts, but between Original English and English-translated-from-French. The statistical comparison between authors and translators in the same language has demonstrated that, in many respects, human-translated texts display a combination of interlinguistic interference or “translationese” and interlinguistic influence or “shining-through” and thus deviate from target-language norms enough to be considered as a distinct “sub-species”.

The inclusion of machine-translated target-texts then allows further comparisons to be made between human translators and MT. Whenever deviation from target-language norms is observed in the corpus of human-translated target-texts, it can be seen that MT likewise has learned to imitate human translators, rather than authors, or is at least subject to the same interlinguistic interference and influence. Moreover, basic metrics such as source/target-length ratio show that MT exhibits less variation than human translation, which may be construed as a rough indicator of the level of creativity or “freedom” with respect to the source-text.

Daniel Henkel is an Associate Professor at Paris 8 University, where he teaches Computer-Assisted Translation and Terminology, and translator of scientific articles. His research in Contrastive Linguistics and Translation makes use of corpora and statistical methods to evaluate interlinguistic influence and interference in translation and assist human translation through electronic resources.

Bartholomew Hulley - Université de Lorraine : EA2338 - Interdisciplinarité dans les Etudes Anglophones - Interdisciplinarity in English Studies

Using NMT to measure bio-translator style

In the past, texts output by machine translation engines (SMT and NMT based) have been evaluated by comparing them with ‘gold standard' bio-translations. This practise has led to the formulation of numerous metrics that purport to reflect the accuracy of the machine translated text, by quantifying the number of disparities, or ‘errors', between the two. Acronyms such as WER, TER, HTER, BLEU and NIST have therefore become familiar to those researching machine translation tools. However, it is worth recognising that such metrics are not wholly objective because they rely upon human subjectivity to define what constitutes an error-free translation. Until a truly objective way of automatically measuring translation quality is devised, any metric involving human judgement will inevitably be imperfect. Any system based on the notion of a singular ‘gold-standard' translation therefore risks classifying perfectly acceptable translations as erroneous.

In this context and given the widely recognised and increasing reliability of NMT, it seems logical that a certain amount of objectivity would be restored into the practise of textual analysis if the subjects concerned were reversed. That is, instead of using a bio-translation to measure a machine translation, to do precisely the opposite: to use a machine translation to measure a human generated TT. Indeed, considering that an NMT engine will translate according to ‘learned' paradigms with complete disregard for extra-linguistic contexts (such as style, purpose and function - meaning it will tackle every ST in precisely the same manner), it thereby offers a dynamically ‘fixed' baseline from which to measure one or many human generated TTs.

In a case study analysing English translations of awards-nominated French comics titles, I demonstrate how applying this rationale can offer the researcher a measure of the translator's approach to any given text and enable the objective comparison of one translator with another.

Waltraud Kolb - University of Vienna

"Quite puzzling when I first read it”: Is reading for literary translation different from reading for literary post-editing? 

Translating and post-editing literary texts are widely considered to be two quite different translatorial activities, not least by literary translators themselves who tend to feel constrained by machine translation output (e.g., Moorkens et al. 2018) or see their “artistic autonomy” at risk when post-editing (Oeser 2020, 23).

One of the most obvious differences is that in the case of post-editing the MT system provides the first draft, while in human translation translators have to invest considerable effort in drawing up a first version. This (human) drafting phase (Mossop 2000) is usually the phase in which the bulk of research occurs and the translator's (creative) engagement with the source text is the most intense. One significant part of this engagement with the source text is the act of reading. Is “reading for translation” different from “reading for post-editing”?

In my presentation, I will use translation process data (think-aloud protocols and keylogs) from an empirical study in which five professional literary translators translated a short story by Ernest Hemingway into German and five different literary translators post-edited a first draft generated by DeepL. I will first outline differences in task organization and workflow patterns between the two groups of participants (some results relating to workflow patterns of the group of translators have been published in Kolb 2017) and then zoom in on the participants' reading processes. Drawing on cognitive stylistics and narratology frameworks (Herman 2003; Boase-Beier 2019), I will show how translators and post-editors develop narrative understanding of the story, the characters, and the worlds evoked by the source text; how they infer meaning from the text and its style and, for example, fill in blanks and gaps during their readings; and how differences between reading processes may be reflected in the final target texts.

Claire Larsonneur - Université Paris 8 – Vincennes – Saint-Denis - TransCrit 

Can an AI package be a literary author? The case of language models

The beauty of artificial intelligence technologies is that they can be applied to all sorts of domains, once they are converted to data. Neural translation is part of a continuum of linguistic tools: text-to-speech and vice versa, predictive typing and so on. All these facets of machine language processing are mobilised in what could be the next stage of linguistic artificial intelligence: language models. Although the names BERT or ELLMO are not well known to the general public, the launch of GPT-3 in the summer of 2020 has made it the talk of the town in the press, because it is a joint project between Microsoft and Elon Musk, but also because of the power of the tool. Language models are distinguished by their versatility. From a few initial elements provided by humans, they can autonomously write all kinds of content: summaries, medical diagnoses, philosophical essays, news, computer programs. This is made possible by the computing power currently reserved for a few very large groups, and by the complexity of the model, which goes as far as processing several hundred million parameters. After a brief review of the history of these tools, I would like to analyse in more detail the type of literary production they generate by looking at several examples in the fields of journalism, online role-playing games, and theatre. I will then look at the reception of this type of work and ask how it is positioned in the economy of culture, between performance and commercialisation.

Bruno Poncharal - Université Sorbonne nouvelle - EA 4398

Testing NMT for discourse coherence

Mehmet Şahin, Tunga Güngör, Ena Hodzik, Sabri Gürses , Zeynep Yirmibeşoğlu, Harun Dallı, Olgun Dursun - Bogazici University

Literary Machine Translation to Produce Translations that Reflect Translators' Style

In our presentation, we will be reporting on a research study* that involves building an English-Turkish machine translation (MT) model to translate literary texts whose style has been defined using corpus tools. We aim to determine the extent to which an MT system can reproduce a translator's style if it is trained on a corpus of previous translations by the same translator. Our presentation will focus on the following questions: 

Does corpus stylistics allow us to better study the style of the translator?  

Can we teach machines to replicate a translator's style?  

To answer the first question, we adapted Leech and Short's (2007) methodology of literary style analysis to Turkish and used AntConc (Anthony, 2022) and WordSmith (Scott, 2021) as corpus visualization tools to analyze the style of a particular English-Turkish translator both qualitatively and quantitatively (Youdale, 2019).  Our preliminary findings suggest consistent use of particular lexical and morphological forms as discourse markers. Our corpus analysis findings will be presented against a reference corpus representing Turkish literary translations between 1946 and 2015 in order to determine whether the stylistic features we identified are exclusive to our translator or general linguistic phenomena.  

We have also applied text classification methods used in authorship attribution to identify different translators. One key question here is how these classifications are affected when two variables, namely author and translator, are involved instead of only the author.  

To replicate a translator's style in machine translation, we built several machine translation models based on both LSTMs and transformers. We are using our own datasets consisting of chosen translators alongside publicly available bilingual corpora. We will conclude our presentation with the initial findings and implications for our future work.

*The study is part of a scientific research project funded by TÜBİTAK (the Scientific and Technological Research Council of Turkey). The title of the project is “Literary Machine Translation to Produce Translations that Reflect Translators' Style and Generate Retranslations” and its grant number is 121K221.

Anna Schewelew - University of California, Santa Barbara

“Pushkin Need Not Shudder”: Machine Translation and Poetic Language

In his 1949 memorandum on Machine Translation, Warren Weaver famously outlined the idea that computers could be used to translate. Yet Weaver explicitly excluded literary texts as a use case for machine translation because literature is the locus of the “alogical elements in language” (Weaver 1949) that cannot be transformed into mathematical terms and thus defy automation. Weaver's “reverence” for literary translation is remarkable because it suggests that understanding literary texts as art, that is beyond their merely propositional content, is out of reach for “intelligent” language processing machines. While machine translation has come a long way since 1949, I argue that it is still ill equipped to tackle poetic language because poetic language transforms linguistic structures into aesthetic objects that cannot simply be reduced to a statistical distribution of words because their meanings are also derived from their sound, graphic appearance, their “mouthfeel”. To understand and translate literature, it is not enough to understand and translate a distribution of words. It is equally important to understand and translate a distribution of words as aesthetic objects. In order to translate literature properly a machine translation system would thus need to be able to switch from a “logic” mode to a “sense” mode and vice versa – something that is very much out of reach for contemporary machine translation systems that can process quantifiable data only. Finally, by looking at contemporary neural machine translation systems through the lens of Schleiermacher's theory of translation, I show that – if literary texts are to encode novel, multilayered ways of speaking about the world, as Schleiermacher suggests – machine translation alone must fail to recreate this novelty because due to its reliance on distribution probabilities within large corpora of texts it is can reproduce what has already been said over and over again.

Samuel Trainor - Université de Lille - Centre d\'Études en Civilisations Langues et Lettres Étrangères (CECILLE) - EA 4074

Mixing Human and Nonhuman Translation: Neurodiversity and the Sociology of Inclusion in Contrapuntal Translation Practice

In Bruno Latour's essay "Mixing Humans and Nonhumans Together: The Sociology of a Door-Closer” (1988), he proposes the term “transcription” to refer to the “translation of any script from one repertoire to a more durable one”, such as the replacement of a police officer with a set of traffic lights.

This paper calls into question this paradigm of social replacement as it relates to NMT. Latour's metaphor derives from an instrumentalising misrepresentation of translation. It is hardly surprising that current debates surrounding machine translation are haunted by its spectre. The majority of NMT platforms are designed to work autonomously, even when they incorporate responsive learning. They are also invariably ‘domesticating'. Their designers want them to pass a Turing Test: producing simulacra of human translations capable of convincing a reader they were written by an accomplished native speaker. This is classic translatorial ‘invisibility'. Anyone with a grounding in translation theory can unpick the basic premise. However, this paper suggests that translation professionals feeling threatened by NMT should avoid abstruse theoretical objections, or appeals to human creativity, and promote the ethical case for a more inclusive, ‘neurodiverse' translation practice.

What separates ‘inclusion' from ‘accessibility' in social policy is the involvement of a variety of target users at an early stage of conception. The growing influence of artificial neural networks in our societies is redefining the concept of neurodiversity, potentially rendering all humans neurodivergent. So the argument for early-stage ‘biotranslation' (not merely in post-editing) has become an ethical question of inclusion. A key contention is that NMT cannot be discussed in isolation from other technologies and their social and ecological impacts. Inclusive working methods are significantly boosted by technological platforms for collaboration. Meanwhile, similar technologies are making literary publishing increasingly multimodal and interactive. But all of this comes at a cost. If NMT could be moved away from current paradigms of replacement, made ecologically more sustainable and methodologically more open and plural – in particular if it could be integrated into technologies that enable interaction, working polyphonically in collaboration with biotranslators – then it could become genuinely valuable in literary translation.

Jean-Louis Vaxelaire - Université de Namur

Translating text or sequences of phrases? Machine translation between Turkish and French

Ray Kurzweil, one of the leading figures in the computer world, announced in 2011 that there would be no need for translators by 2029. However, he made an exception for literary translation, which he said was already difficult for humans to do. Since then, with the advent of neural MT, research teams (at Microsoft and Google for example) claim to have surpassed the quality of human translations. From my linguistic point of view, this kind of triumphant discourse mainly shows a lack of understanding of languages and the translation process.

An analysis of the examples of MT presented in the research literature shows that, while they are sometimes excellent, they always involve English as source or target and, secondly, the textual genre of the translated passage is always close to the training material. As soon as one translates between languages that do not include English and in genres that are far from the original corpus (the range is wide, from Facebook's oral writing to literary texts), the results are much less satisfactory.

We will begin by observing the theoretical problems that explain the difficulties encountered by MT and then secondly, we will analyse the results of some excerpts of Turkish poetry and prose texts (among others Karaosmanoğlu, Hikmet, Dağlarca, Baykurt and Pamuk) in French and compare them with the official translations. The results vary according to the four translation tools used (Google Translation is generally superior to Webtran), but pose different types of problems that can often be explained by the fact that, unlike the human translator, the software does not have a global vision of the text and treats it as a series of unrelated sentences. The question of genre, which is essential in literature, is also never taken into account.