NOTE: This post was originally written on 1 November 2021, but it’s published now, months later because of the cyberattack that UAB suffered then and that caused the temporary suspension of this blog
I am not following Netflix’s South Korean mega-hit Squid Game, being currently off the platform, but I have noticed that the series has attracted much controversy about an issue few people really care about: the quality of the subtitles. What is remarkable is how different the controversies are depending on the linguistic area. For English speakers the problem seems to be whether the subtitles are truly accurate and how much is lost in translation. For Spanish speakers that was also the issue until the audiovisual professional translators called attention, a couple of weeks ago, to the use of automatic translation for the subtitles. I find the problem of accuracy far less urgent right now than the matter of automatic translation, which is not being properly addressed and will have huge consequences in the near future. See how.
As we all know, English is the dominant audiovisual language but the immense popularity of some foreign-language series on the streaming platforms, and the generally low quality of dubbing into English, has forced many spectators to use subtitles. I’m following a CNET article, “Still watching Squid Game on Netflix? Change this subtitle setting immediately” by Jennifer Bisset (https://www.cnet.com/news/watching-squid-game-on-netflix-change-this-subtitle-setting-asap/) to present the problem affecting the spectators using English subtitles to follow the series’ original Korean-language dialogue. All was more or less well until Korean Tik-Tok and Twitter users started protesting about how much the English subtitles missed, from glaring errors to matters of nuance. Bisset warns, like others have done, that although perfect accuracy cannot be achieved, the English subtitles option works much better than the English Closed Captions option for the deaf and hard of hearing, which most spectators use. The English CC subtitles, she explains, are “often autogenerated” and, in Squid Game apparently “a closer match to the English dub than the English subtitles”. The English subtitles which she recommends are not, unlike the English dub, forced to adapt translation to lip-synching, and are, thus, more accurate, though not necessarily error-free, as many Korean speakers have also pointed out. It’s, then, a case, of choosing between the bad and the worse, but not an experience exclusive to this series or to the English-speaking world. As I know from having watched thousands of English-language movies and series with Spanish subtitles the errors are many. They often show that this kind of translation is done in a hurry by underpaid translators lacking sufficient experience (apologies to the ones who are experienced but anyway underpaid).
For the Spanish case, I’m following the article by Héctor Llanos Martínez, “Los traductores españoles protestan por los ‘mediocres’ subtítulos de El juego del calamar, hechos por una máquina” [Spanish translators protest against the machine-made ‘mediocre’ subtitles of Squid Game] for El País (https://elpais.com/television/2021-10-14/los-traductores-espanoles-protestan-por-los-mediocres-subtitulos-de-el-juego-del-calamar-hechos-por-una-maquina.html). Llanos reports that ATRAE, the Asociación de Traducción y Adaptación Audiovisual de España, has complained that Netflix employs a multinational company specializing in automatic translation, Iyuno (https://iyuno-sdi.com/), which produces subtitles later edited by a person working at a third of the translators’ habitual rate. This technique, called post-edition, is what we all use when automatically translating a text that we later revise. According to ATRAE, a translator receives 60 to 100 euros for supervising a 100-minute film, which is awfully low, though getting 300 euros for translating the whole movie doesn’t look too good, either. The ATRAE spokespersons have noted that the AIs generating automatic translation do not understand context, subtext, or wordplay and miss many nuances that a human translator would not miss (though, as I have noted, subtitling is not at all the most accurate type of translation). ATRAE suggests that Netflix may have been unaware of the controversial methods used to translate Squid Game, apparently the first series using post-edition, in view of the care it put into the correct translation of La casa de papel (Money Heist). Llanos comments that Audiovisual Translators Europe (AVTE) had already blacklisted Iyuno in 2020. He also notes that Netflix has declined to make any comments.
AVTE, precisely, released last September an 18-page long Manifesto on Machine Translation decrying the practice (https://avteurope.eu/2021/09/13/press-release-avte-manifesto-on-machine-translation/). The 10 points of the summary include the following declarations: “We do not believe that fully automated localization processes are likely to happen anytime soon”, “Although proponents of MT claim that efficiency gains are guaranteed, fixing a poor translation can take longer than translating the same text from scratch”, “To reinforce sustainability, translators’ working conditions need to be improved”, and “Translators are often not aware that their work is used to train MT engines, nor are they remunerated for this”. This only shows how desperate the situation is beginning to be for professional human translators. We all know, having the experience of using Google Translate, Word’s own translation feature, or other automatic translators such as DeepL that MT has improved enormously in the last five years. In fact, we have all contributed to that, for the AIs are learning from the texts we ask them to translate, constantly improving with practice. I have no idea what I am saying here but Google Translate, which launched in 2006, switched in 2016 to Google Neural Machine Translation (GNMT), MT which uses an AI-powered neural machine translation algorithm capable of processing contextual meaning (much closer to a human brain, then). That explains the dramatic improvement.
My own use of MT means that what would take me 90 minutes to translate from scratch can be ready for uploading in 20 minutes, or less, of revision, which is a great advantage. So, sorry ATRAE and AVTE but in five more years, MT might be as accurate as any human translator, if not more, being already incredibly faster. “Fixing a poor translation” might be by then a concept entirely of the past. I have nothing against translators, quite the opposite: they are professionals I admire profoundly. Yet, it would be naïve to think that any manifestos can stop the march of technology and, above all, the march of greedy, mean capitalism in its search for the lowest-priced acceptable translation. Squid Game’s post-edition methods are just the first sign of what is soon coming. I am not sounding a death knell for professional translation but being realistic.
To my immense surprise, two friends who work as professional translators in public institutions (not as literary or film translations), acknowledged to me recently that the use of MT is common, and that their job consists now of revising rather than translating from scratch. I assume this is the same in many other professional and business environments, and I also assume that many academics with little money to spare from their research projects might choose post-edition translation over the far more expensive translation from scratch. One thing we must be clear about is that MT cannot work without revision: you may take Google Translate and have your academic article translated into Mandarin Chinese believing it to be accurate, but only a native speaker of the target language can determine accuracy. It might well be, then, that translators are sought in the future mainly as revisers. This is the part that scares me very much, not just because the wages of professional translators might be drastically cut and their extremely important task undermined but because if the profession is so badly hit that no young person wants to train as a translator, we run the risk of losing translation altogether as a human pursuit. The vision of a world in which all translators are AI is a frightening dystopia, as it would put a most important tool of human communication outside human reach. Many people believe that a native bilingual person, or someone who learns a second language, can easily translate but becoming a translator requires serious professional training. Who, however, would think of investing long years in that kind of training to compete with super-efficient AIs? And how many people really understand the long-term danger of trusting all translation to AIs?
On the other hand, MT opens up new possibilities worth considering that might enrich the cultural field. Suppose you are an author seeking international publication but failing to find interested foreign publishers. You are being told that the cost of translating and revising your book is too steep, and that expected sales make taking risks of this kind pure gamble. Well, you could self-translate using MT, pay for a professional revision and market your book directly in, say, five foreign languages, through Amazon, or similar platforms, or your own website. Just to give an example, this week I will be interviewing for the new Festival 42 British author Richard Morgan, a relatively well-known SF author whose novels Altered Carbon, Broken Angels and Woken Furies have been adapted by Netflix using the title of the first book. Both Woken Furies and Black Man (known as Th3rteen in the USA), Morgan’s own favourite novel among his production, remain untranslated into Spanish because his publisher lacks the resources. Why shouldn’t Morgan pay for MT plus revision (by a professional translator or an academic like yours truly) and have the novels published as he chooses? He owns copyright, after all. I am well aware that this may sound as anathema to professional translators, but I am contemplating the same process to translate into Spanish my own book Masculinity and Patriarchal Villainy in British Fiction: From Hitler to Voldemort, in view of the dozen Spanish publishers that have rejected it. And, yes, MT has a downside for authors, as you will see if you check MT and copyright on Google: the generation of illegal translations of foreign-language work which do not respect author’s copyright. You might find your own book on Amazon translated into another language, but let it be known that this is illegal, as copyright always belongs to you. So, get there before others do…
A last matter worries me: if I use a translation tool to translate this post, the copyright is still mine, in the same way the copyright of the original text belongs to me and not to Microsoft’s Word, which I use to write it. The software to write and translate is a tool, and not an entity which can own copyright. However, being an avid reader of SF I am familiar with the trope of the AI which becomes sentient and demands to be seen as a full person. If AIs are eventually granted a legal status as persons (as some animals are being granted now), this means that whatever they do, including translation, will be subjected to copyright laws (human translators retain copyright over their translations). The singularity so often announced might happen in 2099 rather than 2022, but it will certainly happen, unless of course climate change wipes us all out. So, brace yourselves for a very strange world in which literary translations, to name the ones closest to my heart, will be signed by AIs bearing personal names. Brave new world… though not for professional translators and, as much as like the idea of AIs, for human cross-language communication.
I publish a post once a week (follow @SaraMartinUAB). Comments are very welcome! Download the yearly volumes from https://ddd.uab.cat/record/116328. Visit my website https://gent.uab.cat/saramartinalegre/. The Spanish version of the posts is available from https://blogs.uab.cat/saramartinalegre/es/