Using AI to help humans translate Swedish labour laws
Over on 666a I provide English translations of Swedish labour laws. For the most part these are based on some translated PDFs that the Swedish government sometimes publish. The government update the laws more frequently than they update the translations though. Providing an up to date resource here meant updating the translations myself, which is a challenging problem.
I used a little bit of AI as part of the solution. But that doesn't mean these are "AI translations". Christ no.
What's interesting about translating legal text in particular is that you really want lots of uniformity in the outcome. If a particular phrase has been translated a particular way nine times before then the tenth translation should follow the pattern too. This means it's not enough to speak Swedish.
Familiarity with the source text and the existing translations is key. In practice, translating any given sentence becomes a long process of searching the already-translated text for similar words and then reading those existing translations to build an understanding of the prior context. Where I found a good role for AI was in optimising that process, enabling me to spend less time on the manual searching and more time on the reading and understanding.
Take the second paragraph of section 13 of the Working Hours Act for example.
| Swedish | English |
|---|---|
| I den dygnsvila som alla arbetstagare har rätt till skall tiden mellan midnatt och klockan 5 ingå. Avvikelse får göras, om arbetet med hänsyn till dess art, allmänhetens behov eller andra särskilda omständigheter måste bedrivas mellan midnatt och klockan 5. Lag (2005:165). | The daily rest period that all employees are entitled to shall include the hours between midnight and 05.00. Derogations may be made if the work, in view of its nature, the needs of the general public or other special circumstances, must be carried out between midnight and 05.00. Act (2005:165). |
First, I use ChatGPT to break this up into sentence pairs.
| Swedish | English |
|---|---|
| I den dygnsvila som alla arbetstagare har rätt till skall tiden mellan midnatt och klockan 5 ingå. | The daily rest period that all employees are entitled to shall include the hours between midnight and 05.00. |
| Avvikelse får göras, om arbetet med hänsyn till dess art, allmänhetens behov eller andra särskilda omständigheter måste bedrivas mellan midnatt och klockan 5. | Derogations may be made if the work, in view of its nature, the needs of the general public or other special circumstances, must be carried out between midnight and 05.00. |
Then a little more ChatGPT to mulch those sentence pairs down further into pairs of words and phrases.
| Swedish | English |
|---|---|
| dygnsvila | daily rest period |
| som alla arbetstagare har rätt till | that all employees are entitled to |
| Avvikelse får göras | Derogations may be made |
| allmänhetens behov | the needs of the general public |
Once we have those word/phrase pairs generated for all existing translations, we can have a page that automatically lists matches for each word in a given paragraph of source text.
The results aren't perfect, and what's great is that they don't need to be. Sometimes the AI fucks up the word/phrase pairs, and it doesn't matter. Since we're not removing humans from the process – but rather optimising how the available time is used – it's okay for this to be a little bit sketchy.
What this enables is to pick out any new paragraph that needs translating and instantaneously scan the entire corpus of existing translations for the bits that might be of interest to the human about to produce the new translation. Even after visually scanning the results one time you immediately start to feel a bit confident that you know how you're going to translate some of the legalistic glue phrases. It frees up your remaining attention and time to focus on translating anything that's actually novel about the new paragraph.
In the end I'm not sure if I saved time from this or just spent the same amount of time more impactfully. With the correctness of legal text having such an obvious importance, you end up wanting to invest a certain amount of love into the work just for a sense of hygiene about the outcome. Like if it's too quick, it just doesn't feel done. So in this case I think it's more likely that using AI just made the work more enjoyable and improved the outcome rather than bringing it to an end sooner. Perhaps that's not the standard capitalist narrative about what this technology's benefits are for society, but I'll take it.
Really happy with this. The level of hype around AI is nauseatingly intense and that's leading to a lot of overapplicaton of the tech. There's a decent amount of people using it to fully communicate on their behalf, which I really don't get. It has such a distinctive tone of voice with all those extra adjectives and adverbs, and that uncanny beehive-looking regularity in the structure of the sentences and paragraphs. It's amazing to me that people either don't see how obvious it is or don't feel how unengaging it is once you twig that you're reading something that wasn't important enough to write. Negative experiences like those leave the door wide open to outright cynicism about the tech in general. I'm glad I have a nice example like this to keep my head balanced.