The Art of Translation reduced to Data

by Tadhg Mac Eoghain

In order to survive as a freelance translator in today’s hostile climate, I’ve had to take a Jekyll and Hyde approach to my work. On the one hand, I very happily collaborate with a team of other like-minded, caring and politically driven translators as part of Guerrilla Media Collective (GMC), a feminist, commons-oriented, worker-owned translation and communications cooperative. On the other hand, in order to make ends meet, I also participate in the predatory, machine-led and profit-driven world of corporate agency translation, where I reluctantly contribute to cultivating AI technology that is specifically designed to make me and my colleagues obsolete. While navigating between these two very opposite worlds, I manage to eke out a living, but no matter which end of the spectrum I find myself in, precarity seems to prevail.

With all of the MTs and evaluations that we provided for free, we were simply digging our own graves. With every machine translation that we edit, perfect and humanise, we are strengthening our artificially intelligent oppressor. This is the hostile industry environment endured by most freelance translators these days.

I get to be Dr. Jekyll when working with GMC, where we explicitly refuse to use CAT (computer assisted translation) tools, preferring to exercise our craft of translation in a more traditional way, giving every text the custom-tailored attention it deserves. We also work on every translation as a pair – one translator and one editor – which means that the translation process takes the form of a collaborative conversation. This generally costs more than typical computer-assisted agency translation, but our clients are happy to pay a bit more for a higher quality end product, as well as to show solidarity with us as artists and activist workers with a social mission.

GMC also enjoys a curated clientele, built up over years of cultivating relationships with clients that produce content aligned with our ideals and values. This includes foundations, public institutions, universities and large cooperatives that have a budget for translation, but also smaller cooperatives, individuals and academics who benefit from our tiered pricing scheme and periodic pro-bono translation work. In short, we work with inspiring people and content that interests us. It sounds ideal, except that we very gravely feel the pressure of competition with translation agencies operating under more extractive corporate business models.

I know something about this other world as well, as it is where I am forced to seek employment – wearing my ill-fitting Mr. Hyde mask – in order to supplement the much more satisfying feminist cooperative translation work. Luckily I hold a rather privileged position, as far as European freelance translation goes, as an Irish-language translator. Irish became fully recognised as an official EU language in 2022, meaning that suddenly there was not only a new daily avalanche of institutional documents to translate, but also a years-long backlog of material to work through at a slower pace. There are now probably close to 200 Irish translators employed across all EU institutions, but freelancers like myself are still needed to help keep up with the massive workload. Typically, agencies will tender an agreement with the European Commission or European Parliament to procure large work contracts and then scramble to find translators and editors to do the work. For me this means regular work with minimal competition and a window into the world of agency translation.

This also brings along the advantage of getting to work with a protected minority language and being a part of the Irish-language revitalisation movement. To a certain degree, I see this as decolonial work, bringing more visibility, opportunity, employment and status to a language that was largely outlawed in its native land for centuries by a violent occupying regime. It is a matter of personal pride for me to see Irish recognised and valued at this level of government, and it contributes greatly to the modernisation of the language, particularly in terms of vocabulary. Irish now enjoys top-quality terminological and translation resources because of the vast corpus of translated legislation and other texts available. A far cry from its former reputation as a rural “poor tongue”, Irish is now innovative, versatile and modernised thanks, in part, to the support and funding it has received from the European Community.

EU translation, however, is entirely unlike the kind of activist-oriented, personalised translation work I do with GMC. The language is very dry, formulaic and often governed by strict document templates, standardised translations of specific terms and the conventions of ‘legalese’.

To handle this level of inflexibility and demand for consistency, many AI-powered tools are used, including CAT tools, machine translation (MT), term bases (TB) and translation memories (TM). All of these work together to extract segments of language – ranging from individual words to entire paragraphs – from previous translations and repurpose them to translate identical or similar segments in new documents.

You often find that a brand new assignment is already 45% translated for you, based on previous TMs. Sounds great, except that this marks the beginning of a slippery slope that ultimately ends up hurting rather than helping the translator. It may seem that the translator simply has 45% less work to do in this example, and as translation is usually paid according to a per-word rate, one should really only be responsible for the words they translate anyway. However, the reality is that the translator is generally responsible for reading, proofing and ensuring consistency within the entire document, including the already translated text coming from the TM. This is simply expected as a matter of course and is not compensated in any way. The unpaid work does not stop here though. The translator is also generally expected to update TBs with any new terms that appear frequently but are missing from the database. This takes time, mental load and will also mean that future instances of such terms will likely be counted by the CAT tool as translated and therefore not fully compensated. Many agencies will also ask editors to complete evaluation reports of the translated texts they review. This is understandable from a quality control perspective, but the time spent on these assessments is not compensated. They are simply expected.

The amount of unpaid care-work that is expected of a freelance translator, who is explicitly paid by the word only, is alarming enough, but it is not nearly the biggest injustice involved in this type of translation work. By now you will have noticed that all of the above-mentioned tasks involve the collection of data in one form or another, and beyond helping to improve the end product, this data is being harvested in order to streamline the work process and allow the agency to make as much profit as possible while spending as little as possible.

TBs and TMs grow with every translation completed, and are shared across EU institutions. As the corpus of translated work grows, it is fed directly into AI-based algorithms that are being trained to do the translation work themselves. This brings us to the inevitable topic of machine translation.

Perhaps precisely due to the large bilingual body of both national and European legislative texts and other bilingual data being used to fine-tune AI-based translation tools, machine translation between Irish and English is alarmingly accurate. It is, naturally, no replacement for a human translator – especially considering some of the grammatical and lexical idiosyncrasies of the Irish language – but it nonetheless renders fairly accurate, largely intelligible results. This, of course, was the goal all along, as it creates a very profitable situation for the translation agencies investing in this very technology.

Supposing that a translator charges €0.09/word for translation – note that per-word rates have, at best, stagnated in the past decade and are, for many language pairs, much lower than 20 years ago – an agency will see no point in paying them this rate when they could simply run the text through machine translation software and then pay them half (or less!) to edit the machine translation. The client still pays the agency the same price, but the agency pays the freelancer less money for a faster turnaround. This service is called machine translation post-editing (MTPE), and it is becoming increasingly common.

TBs and TMs grow with every translation completed, and are shared across EU institutions. As the corpus of translated work grows, it is fed directly into AI-based algorithms that are being trained to do the translation work themselves. This brings us to the inevitable topic of machine translation.

In certain translation sectors, MTPE will soon be standard practice, and some agencies are shamelessly forthright about this. For one large agency client, I have been – as far as I can tell – the only German-to-Irish translator they were able to find to handle the needs of an important client of theirs. I recently received an email stating that all translation jobs for that client will now be MTPE jobs, an automatic 50% pay cut for me despite the fact that German-to-Irish machine translation does not yield very good results and the ensuing post-editing process is often more of a headache than simply translating the text from scratch. When signing on with a new agency, I am now generally asked to provide rates for translation, editing and MTPE. To me these requests read like a threat: if you dare ask for a fair translation rate, we will simply machine translate everything and pay you even less.

Those with a permanent translation position are not spared either. The European Language Industry Survey 2023 reports that around 30% of professional translation is now done via machine translation, which is expected to take over most professional translation work by 2030 at the latest. Brussels is hiring fewer and fewer in-house translators while the workload only grows, a trend which has resulted in a reported increase in burnout and mental health issues among permanent staff in the European Commission’s translation unit.

The worst part is that every translator who has ever worked with CAT tools (most of us) knows that the data extracted from our own labour has fine-tuned this weapon of capitalism, which is now being used to render us completely obsolete.

With all of the MTs and evaluations that we provided for free, we were simply digging our own graves. With every machine translation that we edit, perfect and humanise, we are strengthening our artificially intelligent oppressor. This is the hostile industry environment endured by most freelance translators these days.

The market has become seemingly saturated with translators looking for work, and employers take a predatory approach, asking us to lower our already deflated rates. When lowering our rates is no longer a viable option for us, agencies simply feed the work to the machine and find someone who is happy to get any work at all, post-editing included. According to the European Language Industry Survey 2020, 40% of freelance translators reported that they were unable to earn enough income from translation work, a clear sign that something is rotten in the industry. Indeed, the 2023 Survey confirmed that human translation was the area of the entire language industry to suffer the most in the previous year. So how do we push back against this?

In my case, the Irish language itself accords me a uniquely privileged position. Translating into a protected minority language with a legally stipulated workload and a limited pool of professional translators definitely elevates my status to a certain degree and puts me in a position to call the shots more than someone translating between English and French, for example. In the instance of that one agency switching all German-to-Irish translation work to MTPE, I simply informed them that I do not do MTPE. I saw several posts online the next day seeking German-to-Irish translators and am waiting to see if they find someone or if they will come crawling back. Perhaps I’ve lost that client for good, but I’m lucky enough to be able to choose my dignity instead – for the time being, anyway.

Recently, I learned from a freelance Spanish translator that many of her colleagues doing externally contracted work for the European Commission were abruptly informed that they had been replaced by eTranslation, the EC’s own machine translation tool. Curt emails were received stating that it was a final, non-negotiable decision and they would not be receiving further work. Instead, the following statement can now be found on the EC’s website:

This translation is generated by eTranslation, a machine translation tool provided by the European Commission. Machine translation can give you a basic idea of the content in a language you understand. It is fully automated and involves no human intervention. The quality and accuracy of machine translation can vary significantly from one text to another and between different language pairs. The European Commission does not guarantee the accuracy and accepts no liability for possible errors.