Using AI for book translation: an interview with Rossum Press

In an interview with Media Makers Meet (Mx3) last year, Aliya Itzkowitz, Manager at FT Strategies, commented on the potential ability of Large Language Models to translate written content into new languages, enabling publishers to compete for global audiences. Publishers have already seen some success with this approach, but recognise that human review and editing remain necessary.

In this context, Sam Gould, Senior Consultant at FT Strategies, recently interviewed Ryan D.R. Cook, the founder and publishing coordinator at Rossum Press, an independent press which uses computer-assisted literary translation to publish under-the-radar novels in English for the first time. The aim of this interview was to see what news publishers can learn from the processes that Ryan is developing and using for book publishing, as both industries undergo enormous change in the age of AI.

What are the potential benefits of using AI for translation?

The risk to any literary translation, whether it’s computer-assisted or not, is that it won’t faithfully reflect the experience of reading a work in its native language. That means there’s a whole spectrum of “wrongness” in translation, from a total misreading, to a failure of nuance.

We currently use a team translation system, with AI being one member of that team, joined by at least one contributor skilled in the source language, and one contributor skilled in the target language. By the time the process is complete, the outcome, which is the result of substantive human editorial labour, is not practically different from the outcome of a purely traditional translation.

As far as technical challenges, in my experience, using a general purpose text transformer obviates the bulk of the social bias that one sees in other AI contexts, before the manuscript even gets to validation. Training data that reflects human prejudices gets smoothed out by the narrative context.

Honestly, our biggest technical challenge on accuracy is when we run into the well-intentioned content filters that screen out intentional Sex And Violence from our titles. Whenever we introduce a new model into our system, we have to adapt our approach to get around this.

Which models and tools do you use for AI-assisted translation?

Since the start, we’ve used publicly available general purpose text transformers. The flexibility of foundation models like these is key for literary translation, as it automatically, and with decent accuracy, translates figurative language which would escape a purpose-built translation model (since a foundation model is more likely to translate based on meaning and context, rather than direct word-to-word translation).

Beyond that, their performance is broadly comparable for our purposes, and we basically use ’em all: GPT, BERT, Claude... Right now, we generally use a translation from Claude 3 Opus as the base translation, but other models will be used for automatic validation. Claude does the best, by a small margin, by a lot of the metrics important to us.

All of the models currently available have hard and soft character limits, where they’ll either crash or start losing the plot after a while. So the first step is breaking the source text into pieces. Using an available model, we translate the pieces individually, reference the translation against an alternate model, resolve or note substantive differences, and reassemble them. This is generally the limit of our automation. The next step is to review the output, modify the prompts, and iterate until we get as close to perfect as the AI will get us.

Can you describe the typical workflow for AI-assisted translation?

The process is new enough that we make some substantive improvements each time we do it. Broadly, an editor skilled in the source language starts with a translation that has been generated and translated by AI. The source language editor validates the text on a semantic, non-figurative level, and then does a close reading to flag key passages. After this, an editor skilled in the target language edits the manuscript for literary quality, in collaboration with both the source language editor and the author. In practice, we hop around the stages a lot in the interest of publishing the best translation we can, it’s a lot of conversations. We also try to involve the author as much as possible.

How do you see AI transforming the publishing industry more broadly?

I’m worried that the publishing industry is in denial about an inevitable change that’s already in progress. People assume that I’m very pro-AI, because I started a company that centres AI in our business model. Really though, my concern is to create a socially productive application of an inevitable technology.

We will see, and are already seeing, widespread adoption of AI in translation, proofreading, and design. I’m certain that, on the marketing side, AI is going to make it easier than ever to bring readers together with exactly the books that they are looking for. Although there’s a lot of understandable concern about threats to intellectual property, I think we’ll also see new technologies applied to defend copyrights as well.

Reading took a big hit in the 20th century. From the invention of radio, then television, then the internet, then the smartphone, our lives are filled with easier, faster dopamine sources than a well-written novel. (Which, for me at least, is more like an IV dopamine drip.) For those of us who still love to sink into a fictional universe with a good book, I think it’s imperative that we turn towards the 21st century rather than away from it.

For those of us who still love to sink into a fictional universe with a good book, I think it’s imperative that we turn towards the 21st century rather than away from it.