#34: AI in Translations

Ian Bowie
Hello and welcome to AI unfiltered with me Ian Bowie and our resident expert, Michael Stormbom, where we will be talking about everything to do with AI in our modern digital society and what the future holds for all of us.

Michael Stormbom
Today we’re talking about AI in the context of translations.

Ian Bowie
You were saying the other day that the stage that machine translation is at the moment, we’re still going to require human input.

Michael Stormbom
Well, I mean, that definitely depends on the use case. There are a number of use cases where machine translation works just fine. So many, especially if you don’t need 100% accuracy, getting the gist of it is sufficient.

Ian Bowie
So the good enough translation.

Michael Stormbom
Yeah, so to speak, and that will probably be of course, the majority of translations in the future, the good enough varieties, I mean, for example, quite a few webstores already just have an automated translation of their product descriptions, for example, when Amazon entered the Swedish market. There were a lot in the news that some of the translations were a little bit so-and-so, that’s right. But of course, with all that news about how the translations were very funny or inadequate. That of course, also simultaneously served the purpose of getting people used to that being the case. So in a way, lowering people’s expectations when it comes to the quality of the translation as well.

Ian Bowie
But there are going to be instances where that level of translation is not good enough. I mean, medical translations, for example.

Michael Stormbom
Indeed not, so I mean, for example, if I want my patient journals, translated from Finnish to Swedish, and well, the neverending problem there is that there is a lack of Swedish translators. So if you want a translator then it takes forever, and then I’ve sometimes when I’ve gotten those translated to Swedish then, the translation might leave a little something to be desired. So I remember some instances where the wrong preposition was used, which resulted in like, shall we say anatomically problematic, right, okay. Yeah.

Ian Bowie
Yeah. But also, I mean, for example, at the moment machine translation, doesn’t really do humor. Probably doesn’t do nuance.

Michael Stormbom
It depends because you don’t need it to be in that sense word for word translation. So I mean, you can have if you have a, let’s say, an idiomatic expression or, or a joke or anything of the sort, and of course you could have, if you have any training data equivalent in the other language, you know, then you could of course, create a model. So that’s certainly possible.

Ian Bowie
That’s quite a good example about this actually is I’m a big Guy Ritchie, film fan. And one of the films that I really love that he’s done, it’s called the Gentleman. Yeah. And there’s a scene in there which I don’t know who’s done the subtitling, but they just missed it because probably I don’t know, maybe not even a way of translating it into Finnish but it’s when the Hugh Grant character is talking about he’s trying to persuade another character to listen to him and pay him money for information that he has. And it gets to the stage where the person he’s trying to get to pay says, Look, you haven’t got enough information? I’m not interested. And the Hugh Grant character says no, no, hang on a minute. I’m only lubing you up. Right. All right now, you know, I know what that means. You know what that means. But no, I mean, at all missed it completely. And that’s the sort of expression that I would also imagine.

Michael Stormbom
Well, I mean, to the point about the machine, so yeah. If, if it had been if shall we say bad translation is there in the training data? And that’s the only only translation available for them to be sure. The machine, right. Yeah. Of course, it’s human translation that forms the basis of what the machine does.

Ian Bowie
I mean, you know, English particularly is so full of, you know, this wide range of expressions. I mean, another perfect one is the fact that London can be referred to as the smoke.

Michael Stormbom
Yeah, you mentioned this in a previous episode, I think.

Ian Bowie
Yeah, yeah. Because I mean, I remember, a colleague of mine, kept getting phone calls from some lady friend of his, who worked for one of these subtitling companies, always asking him you know, can you just explain what this really means so that she could try and translate it into Finnish. And a machine doesn’t have that option. It can’t ring somebody up and say, oh, what does this mean?

Michael Stormbom
For now? Hang on. I need to make a phone call.

Ian Bowie
Yes. That would be actually quite funny, wouldn’t it? Yeah.

Michael Stormbom
I’ve been watching one of these sitcoms on on HBO Max. And I think just by default it has the Swedish subtitles on. I mean, not that I needed sub titles. But of course, we want to have the volume a little bit lower. And you might have the subtitles that was on but but you can definitely tell that there are all of these idiomatic American expressions that whoever has done the subtitles they just didn’t get yeah and it’s not just about well, of course, you have constraints when it comes to how much content you can put into it obviously, but so you have to make some you have to make some trade offs there in terms of representing what is being said, but nevertheless, like completely missing.

Ian Bowie
Because it goes back to what you’ve said before about the can be as good as the data that they’re given. Yes. But I mean when we think about you know natural language processing, I mean English, particularly such a huge language. I mean, I wonder if it’s even possible to input all the possible variations of what could be said.

Michael Stormbom
Definitely it is. I mean, that’s what these large language models as they’re called LLMs, which I mean, they put most of the internet into those models that’s perfectly possible not not for mere mortals as us that don’t have access to the machine power but of course, the big, big companies are doing all that. Infinite machine power.

Ian Bowie
So they’re gonna miss something.

Michael Stormbom
I mean, I don’t think it will ever be perfect.

Ian Bowie
Some bloke in an obscure village in North Yorkshire, for example, will come up with something that’s not in there.

Michael Stormbom
Or and another thing is that language changes, as well. So I mean, that has to be reflected. So it’s not that you can just create one. Here’s our machine translation model that will work perfectly for the next 100 years. But of course, of course, that would work you need to continuously update. Yeah.

Ian Bowie
I think imagining that we’re still going to be here in another 100 years is rather optimistic. You and me personally, I’m just talking about humanity in general.

Michael Stormbom
So the machine translation model might be a little bit moot at that point, but yes, aliens will pick up the machines and aha!

Ian Bowie
My God, these people were backward

Michael Stormbom
Yeah, I just recently saw an article by about our good friends at Meta we seem to fall into this pattern where every second episode we bash them. But specifically within the context of machine translation for languages to do not necessarily have a standard written form. So Meta, they’ve developed a translation system for Hokkien, a language primarily spoken by the Chinese diaspora and which primarily is an oral language. A very large number of languages spoken today, they don’t actually have much in ways of or a standard written, so a large number of languages that only exists as spoken languages. Yes. The idea being in speech to speech translation,

Ian Bowie
I was gonna say, I mean, that’s going to put an awful lot of people out of work, isn’t it? I mean, think about the European Parliament. Is one very small example. So, but actually United Nations, there’s another one slightly bigger example. You’ve got all of these simultaneous interpreters working away.

Michael Stormbom
Granted, but how many simultaneous interpreters are there really, on a global scale? I don’t know.

Ian Bowie
Well, I don’t know. I mean, there must be well, okay. Of course not. Not. It wasn’t.

Michael Stormbom
I mean, it’s definitely a profession that’s endangered. But is it..

Ian Bowie
You know, it’s not it’s not going to cause mass unemployment across the globe. Is it, no.

Michael Stormbom
A challenge for Hokkien and for many, many languages across the globe is lack of training data. It’s a low resource language as it’s called. Because we have spoken about just that in the context of

Ian Bowie
But what about for example, this idea of, you know, getting an algorithm to write articles for you. What if you want an article writing? Alright, let’s take an easy example. We want a travel article writing about Turku. But we want it writing in Italian and French and German and Spanish.

Michael Stormbom
Well, I mean, that’s an interesting topic because of course, all of this stuff that we’ve been talking about on this podcast I mean, it’s been mainly done for English so for example GPT. English and quite a lot of these models are done for English and trained on a material which is also due to the fact that there’s so much material available and so that that may come back to the language equality question or be spoken about previously that for those languages that there is material so they have an easier time getting something. But that’s what also was so interesting about what Meta did with this English to Hokkien translation, because this was specifically done to test it on languages for which there isn’t so much training data. I think that’s quite an interesting one, because then that might make it easier in the future for for languages without so much material available.

Ian Bowie
I mean, I was thinking purely from a commercial perspective, you know that. Well, yeah. You want to attract tourists to your destination, but no, I mean, the basic principle is, you know, if you want to sell to somebody, then you know, you’re going to sell a lot more and and a lot better if you are communicating with them in their own language.

Michael Stormbom
Yeah, I mean, one trend that I’ve noticed on a completely different topic, but in, in in the field, working with… there are a lot of startups, so they, you know, they start out working with Finnish or, or Swedish, but then of course, you know, we will be wanting to go global, we want to internationalize and then it’s always okay. Let’s do English next. But I mean, always it comes down to okay, we should do stuff in English and then all of the other languages are forgotten. So I think that’s…

Ian Bowie
Yeah but actually the interesting thing here is that everybody’s doing something in English. Yes. But there’s a lot of you know, not very many people doing it in these minority languages. Exactly. So you’re gonna carve out a very lucrative nich for yourself. Even even smaller European languages. Welsh, perhaps there’s another because there’s a huge amount of money available. Welsh, Irish, Basque… Of course, if you start working with those languages, you’ve got to have their some native speakers involved as well..

Michael Stormbom
Well. Ideally, yes. Otherwise, I think it wouldn’t really work. Yeah. I mean, you can do the basics in a sense, but you need to obviously have you need to verify that you actually doing that correctly.

Ian Bowie
Yeah. It’s quite an interesting idea in it.

Michael Stormbom
The problem has been for those smaller languages, so it hasn’t been commercially viable. Right. Because it’s been so expensive to develop technologies. Yeah.

Ian Bowie
So now that threshold come down as well.

Michael Stormbom
that he that is the point yesterday, there are starting to be these approaches for languages so called under resourced languages, there’s an opportunity to be able to do so without without so large cost that it will no longer be economically viable. So I think that’s a

Ian Bowie
Is it possible that the market is getting swamped? We know who worked well with translation companies I mean, you know, for example, I just did a search on Google AI in translation. And then just pages and pages of translation companies, all saying, oh, yeah, you know, we use machine translation. We use AI. It’s brilliant. It makes everything faster. And I just sort of thought, well, well, I have goodness. Yeah, there are just so many of them.

Michael Stormbom
Well, I mean, it’s a little bit I think the same thing as when translation memories were introduced in so I mean, either use machine translation or you go out of business.

Ian Bowie
Yeah, I know that. I mean, what I’m thinking is, the market seems to be flooded with translation companies. And perhaps you know, machine translation has made the threshold for entry much lower.

Michael Stormbom
My impression is more that the market has consolidated considerably. As there are a few of these big ones really large multiple there are big ones as well. Yeah. And they’ve been buying up other companies like crazy and then there’s, of course, a lot of these were like, shall we call them boutique? Translation? Companies, but I don’t know if there are any more than there used to be. But yeah, I don’t have any data on that.

Ian Bowie
No. But these these big companies are becoming behemoths, aren’t they? I mean,

Michael Stormbom
They are indeed and that might be a risk, of course to the smaller languages as well that they start focusing on the most lucrative ones, but

Ian Bowie
Well, of course they will naturally. Yeah. But is it actually good for the client? That suddenly you’ve got maybe three or four very big companies and not a lot of other choice?

Michael Stormbom
Just going to another market. Is it a good thing for the movie market or three to four conglomerates that are basically the entirety of Hollywoo.

Ian Bowie
No, actually, no, of course there’s no no less choice. Yes. Because polarized you get what we give you all you don’t watch TV, you know, I

Michael Stormbom
Indeed, is it a good thing that well, of course, Spotify and so forth has changed the game a little bit but that used to be the big three or big four record companies that control the entire market and considering how quickly that went away when we don’t have Spotify and so forth. I think we can well argue that that was not a good thing it wasn’t it was not for sure. Yeah.

Ian Bowie
Of course it does a little bit okay, I translations a little bit different but I mean, if you think about music, film, literature, it’s also now become a little bit of a wild west.

Michael Stormbom
Yeah. So in that sense, probably the entry to start something up with machine translation is probably quite low. But yeah, absolutely. But on the other hand, is also quite difficult to gain a significant hold of the market.

Ian Bowie
Well, that’s what I’m thinking, Is it even worth it? Is it worth the effort? Or should you go and find something else? Yeah.

Michael Stormbom
Yeah, of course. I mean, from a customer perspective, if you can go to just one translation company, that can handle all your language needs. Of course that’s, that’s convenient.

Ian Bowie
It is. Yeah. I mean, it’s the way language training went as well, you know, they wanted one provider. Although, of course, with language training, it’s very, very much about the teacher. So even though you’re dealing with the same company, it doesn’t mean to say that you’re going to get the same level of service across the board because you’re using humans.

Michael Stormbom
That’s an actually an interesting aspect when it comes to machine translation. Because of course, I mean, there might be the individual quality of the translator, the human translator. If you have machine translation there. You might take away these variations that depend, that all these individual variations depend on the person because a fairly big chunk that is automatically run by the machine. So that’s sort of standardized translations, in a way.

Ian Bowie
Yeah. Yeah. I mean, I’m sure that’s very good for for companies. So basically, there might be maybe a greater need for people like me to polish what a machine has translated.

Michael Stormbom
Well, yeah, I mean, that’s basically how much of the translation industry already works, that you take machine translation and you have a human being who post-edits and corrects it, but yeah, definitely the the translation business is fundamentally changing.

Ian Bowie
Yes. So it’s all going to go basically into sort of post editing.

Michael Stormbom
Oh, yes. Yeah, it’s not the case that it’s all about profiting already. The other day I was in in the bookstore here. So happened seeing in the Swedish section, this book by a Finnish author and and I was quite interested in reading the Finnish version. So I took a look at the book to see what was the original title and then it transpires that the Swedish translation was actually a translation of a translation. So it had not been translated directly from Finnish it had been translated from the English translation..

Ian Bowie
Wow, okay. Yeah. That was make sense. Yeah. Yeah. So.

Michael Stormbom
I’ve never seen that for you know Finnish to Swedish before today. I think that’s quite indicative of the lack of Finnish to Swedish translator.

Ian Bowie
Yeah, that doesn’t surprise me. Yeah. That something gets translated into English and then from English into another language.

Michael Stormbom
Yeah. No, no, I mean, as a general concept, of course, but specifically Finnish to Swedish. That’s a quite an interesting one. And in this bilingual country.

Ian Bowie
There’s obviously in sort of, you know, the immediate sort of future massive potential for Finnish the Swedish machine translation software.

Michael Stormbom
Absolutely. And I mean, I see that as really a way to, I mean, not just to maintain but to actually improve the amount of information that’s available in Swedish because, I mean, of course, it reduces the cost significantly in producing the material in the first place. And, in many cases, you don’t need it to be like 100% perfect it is. Yeah, it’s good enough if it’s, if it’s in the in the ballpark, so to speak. Yeah. And of course, in many instances, it does need to be very accurate, but then better to free up to human beings to work with those particular types of translations.

Ian Bowie
You’ve been listening to me Ian Bowie, and my colleague Michael Stormbom. On AI Unfiltered and for more episodes, please, go to aiunfiltered.com Thank you.

Transcribed by https://otter.ai