quick thoughts on ai & translation

AI, Chat GPT: these are the buzzwords of the year. The machine mind has gotten smart enough to score in the 90th percentile on the Bar Exam and capable enough of human mimicry to offer plausible self-help advice. There is of course considerable debate as to what its real impact will be: a complete societal revolution followed by the machines taking over the planet and relegating humans to a subaltern role? Or another over-hyped technology, akin to the self-driving cars that were supposed to have put us all in the passenger seat by now?

But there is at least one arena that seems to keep coming up as an accepted point of agreement: translation. Open AI CEO Sam Altman recently mentioned in an interview that he enjoys being able to travel the word and communicate, thanks to technology, with people anywhere. On my favorite podcast, Vox writer Kelsey Piper said yesterday how excited she is for AI in the field of translation, comparing the transition to the shift in the nineteenth century toward machine-made rugs. Putting aside considerations on the quality of communication when traveling with Google or my thoughts on comparing rugs with language and human interaction (perhaps for another blog post), and acknowledging that despite some of my luddite tendencies I am not actually against the idea of technology used in the service of translation (we will all be cyborgs by the decade’s close), I want to take a moment to express my current reservations about these tools.

Out of curiosity, I’ve tried Chat GPT for translation. And in my professional life, agencies and direct clients often ask me to work with specialized translation tools that use both AI and more traditional machine translation methods. I would say that compared to similar systems even a few years ago, the improvements have been massive. Still, given all the hype around these technologies, I am continually surprised at how extensively I have to intervene, often to the point of completely retranslating giant chunks of text. Just this morning, I was working with one such tool that has been trained in industry-specific French to English translation, and to give you a short example of what we translators are dealing with, here are a few of the suggested translations compared to the end translations:

  • “These two models, fitting at the waist” became “These two shoes run small”
  • “Think about changing the price tags” became “Be sure to change the price tags”
  • “Put the product on the reverse side” became “Turn the garment inside out”

I’ve chosen these three examples nearly at random from a list that would be too long to reproduce here. The main criterion for selection was length: I didn’t want to do an in-depth autopsy of a long passage (again, perhaps for another day), preferring to give you a sample that could be grasped quickly and easily. As you can see, the first translation is simply nonsense. The second makes sense but does not convey the proper thrust of the imperative: this is not a suggestion but a command! Finally, while the third translation almost passes in terms of meaning, it lacks flow and naturalness; it sounds like a machine translation.

Ultimately, my point here is this: not so fast! I think there’s a lot to be considered in terms of what will be lost when we start replacing humans with machines for language jobs, from what it will do to the economy to notions about how we communicate with each other and create shared meaning. But for now, my thoughts are very much in the day-to-day work of translating and editing, and I would say that the task of the translator is still very much translation; that is, analyzing and understanding a text in its source (original) language and then, with considerable care and reflection, carrying that message over to the target language so it can be understood by a “foreign” reader.

This is why I am careful about the kinds of ‘post-editing’ (machine translation followed by human proofreading) jobs I take on. Because the language produced by artificial intelligence and machine translation still requires a lot of work. At times, I even ask myself if it takes me longer than a traditional translation, since I have to consider not only the source text but also the suggested translation–twice the reading! Moreover, the power of suggestion is difficult to counter. If the translation provided by the machine is not good–or simply if it could be better–it takes a certain amount of almost willpower to forget it and come up with something else.

My fear is that the promises of AI in this field will serve to drive down prices, even if the “assistance” provided does not save time in a proportionate way.

Why Machine Translation is Not My Bogeyman

File:Gustave Doré - Dante Alighieri - Inferno - Plate 13 (Canto V - Minos).jpg
Gustave Doré – Dante Alighieri – Inferno – Plate 13 (Canto V – Minos)

There are many misconceptions about the job of translators. When I get out from behind my computer and into the world, the people I interact with, from family members to other professionals to parents in my son’s 0-3 play group, inevitably do one of the following: treat me as a human dictionary, express skepticism about my stance on not translating out of my mother tongue, act surprised I don’t speak a gazillion languages, give me a sly look as they ask about machine translation and what they assume is my impending obsolescence. An author once jokingly admitted his astonishment at meeting a real live translator, as if I were a rare specimen, one bound for extinction. (I went on to translate several of his books.) Even though much of what we do as translators happens remotely, and so necessarily through the mediation of machines, our task remains inherently human. I have no interest in reading a novel or poem written by a robot, since how can a machine give me insight into the human soul? Likewise for a translation made entirely through machine translation: how can technology understand and render all the nuance, tone, cultural innuendo, and je ne sais quoi of great literature? Indeed, the fact that some books get retranslated again and again, to better speak to different generations of readers, is telling.

Yet machine translation, particularly together with other features of Computer Assisted Translation tools, is becoming more and more vital to our work, making us more efficient, less prone to error, and better able to collaborate—provided, that is, we know when and how to make use of these tools. For some texts, they can be an asset. Translation software is often a good solution for documents featuring a lot of repetition, since it helps translators be consistent in their vocabulary choices and can free up mental space for more challenging areas of the text. It can also ensure that numbers and dates don’t get distorted through typing errors. Moreover, in my work as a translation proofreader, I have come across many translations with missing sections of the original text. This can happen when translations are made under tight deadlines that don’t give translators the time to review their own work against the source text. But, even under time constraints, it can be avoided with the help of translation software. Finally, for large, on-going projects (as in some legal cases, long-term marketing campaigns, etc.) software can help translators work together and provide consistent language choices, even over periods of many years.

In my experience, machine translation tools are not well suited to literary or even academic translation. In addition to the loftier ideals mentioned above, I attribute this in large part to parsing. Most translation software parses texts into sentence-by-sentence segments and will then provide translation suggestions, recommended vocabulary, and so forth. But a literary translator needs to work not only at the minute level of the sentence or the individual word, they also have to grasp the larger whole, which can sometimes call for work on a more architectural scale as the translator rebuilds a section to better reflect what the source text is up to. Even the suggestions given by machine translation can be a nuisance, cluttering the translator’s mind and preventing them from straying away from the source text’s syntax, for example, to find other, less obvious solutions.

Similar reservations can be expressed for academic translation, though there is one thing I would like to add: academic style. Cultural differences apply not only to the content of what is being expressed but to the form. The ways in which we express ideas, including sentence length, the amount and type of jargon used, what we consider “smart” language, the acceptable use of repetition or passive voice, and so many other factors, can vary dramatically from one academic culture to the next. And this cultural straddling, this refashioning of a text to make it cohere with the target culture’s expectations and norms: this requires human judgment.

I am a translator, and I am not afraid machine translation will put me out of a job. Machine translation is one of many tools at my disposal to provide the best possible translations for my clients. Now, what is my bogeyman? A last-minute babysitter cancellation when I’ve got a deadline coming up. Or the Internet going down …

The American Translators Association has a useful position paper on machine translation: ATA Position Paper on Machine Translation: A Clear Approach to a Complex Topic. Their takeaway: “Professional translators and machine translation engines work together very well. […] If reliable and secure translation is desired, machine translation should not be used without the ongoing involvement of professional translators.”