The Empty Briefcase
Lola Rogers on translation and the rapid rise of language machines
First published in The Author, the journal of the Society of Authors
My husband once bought a second-hand briefcase with a three-digit combination lock, and was trying to find tips online about how to open it – was there a mechanical solution, or a method of examining the case or the lock to deduce what the combination might be? I suggested that he just try every number from 000 to 999. He thought that would take too long and be too tedious.
‘It’s like knitting,’ I told him. ‘You knit one stitch, and another stitch, and after you’ve done that hundreds of times, you have a sock.’
The combination turned out to be 249, and it took him about half an hour.
My husband is a thinker and a poet at heart, but the briefcase lock was a simple machine, and even if there were a way to figure out how the mechanism worked, just trying every number was a quicker way to get it open. Sometimes the quickest way to solve a problem is the least interesting one.
I majored in linguistics as an undergraduate, in the first half of the 1990s. Having been one of those strange children who enjoyed diagramming sentences in English class, my favorite undergrad subject was syntax, the study of the grammatical rules of language. Every class session was like a knotty puzzle we were solving as a group, uncovering the intricate, systematic patterns we follow when we speak, and learning to distinguish them from the rules of “good grammar” we may have been taught in school. I remember hearing a story, attributed to E.B. White, illustrating the difference – about a father who goes upstairs to read his child a bedtime story. “What did you bring that book I don’t want to be read to out of up for?” the child asks, ending a sentence with multiple prepositions, as we do in English.
I also heard a lecture on a new course being offered in Computational Linguistics, studying ways to program computers to recognize and reproduce human language. From my brief studies of the complexity and diversity of language, I was skeptical that this would be possible any time soon. How could anyone program a computer to use human grammar when humans themselves hadn’t yet fully understood it?
The speaker explained that most computational linguists were no longer focused on programming the mechanisms of human grammar into their machines. Instead, the goal was to simply collect large quantities of language data and use it to find the word or phrase most likely to accompany another in a given context – an approach now known as a large language model. This, too, seemed like it couldn’t possibly work. What use was it to know what words or phrases have usually been used? Don’t you need to know what words or phrases out of all the words or phrases in all the world could be meaningfully used? What about creating new words, or new grammar? How would you ever come up with anything interesting using a program like that? And where would you find enough data to do it? It would take decades. It sounded like a very tedious thing to study.
Thirty years later, I am now a regular user of a variety of computer translation tools built by those very methods. My translations have been added – without my permission, or that of the authors of my source texts – to a vast pool of human-generated data. It turns out that programming machines to reproduce something like human language is like opening that briefcase – trying every possible option is quicker than understanding how it all works. Provided you have vast quantities of ‘free’ data, enormous arrays of machines, and an endless supply of energy.
The resulting language models are useful. I regularly use a translation memory tool for translating prose. It provides instant synonym suggestions and keeps a record of how I’ve translated words or phrases in the past. If I’m translating a book by an author I’ve translated before, I can include that work in my translation memory to remind me which words I used so that I can maintain consistency, or diverge from it. It’s not much use for poetry, though. It’s not much help with any text when I’m trying to spark unexpected connections, or create natural dialogue, or rhyme or rhythm or delicious sounds. I have to do that myself.
The new so-called AI models are useful in other fields as well. They can provide instantaneous medical diagnoses, or offer legal advice in layman’s language. The information they provide can’t be relied upon for any but the most basic questions, however, without being vetted by a human expert.
Machine translation is similarly unreliable. I was recently approached and asked to use a new computer translation system to translate a text and edit it for publication, on a much tighter deadline than my translations usually require. Based on my experience with machine translation, I was pretty sure the project would take as long or longer than it would take to translate the book myself, but they assured me that their system was much more advanced than the widely available versions I had used. And I wouldn’t simply edit the machine’s translation. I could read it and prompt the program to improve the translation, and keep prompting it until the text was acceptable. They sent me examples of this process, from the first draft through two ‘improvement’ prompts.
The first go was a lifeless translation with multiple errors, but the ‘improved’ versions were worse. They just used longer and more obscure words. It was clear that the project would take forever. The only way I could think of to manage it would be to narrow the program’s data search to the works of experienced human translators working with very similar texts. But the result would no doubt still need a lot of editing. I would basically have to translate the book myself while simultaneously providing more human data to improve their machine, using the ‘free’ accumulated human labor of my colleagues, on an impossible deadline. It would be a needlessly tedious addition to an inherently creative project.
I declined the offer. Machines can provide ideas for ways to translate something, but they are as yet unable to translate literature. A machine doesn’t have a human’s intuitive, ever-changing system of rules upon rules and meanings upon meanings to be followed or diverged from based on experience of the world. It can shuffle words and offer them up as options, but it has no idea what they mean.
I’m still skeptical that computer systems will ever create literature. But will they create a facsimile of it sufficient for many purposes, in ways I can’t anticipate? Like my husband opening his second-hand briefcase, I over-estimated how long it would take for a simple, try-everything approach to begin to predict the patterns of language. The tedious method worked. But knitting is a creative act, too, and opening the briefcase wasn’t exactly like knitting a sock. He didn’t have a sock when he was done. He had an empty briefcase. It was up to him to decide what to put in it.
First published, in edited form, in The Author, the journal of the Society of Authors