The role of AI in localization - with Olga Beregovaya (VP of AI at Welocalize)

Hello everyone, and we're here with GALA Champion Olga Beregovaya. I hope I pronounced it correctly. Who is Vice President of AI Innovation at Welocalize. Why don't you just start by telling us how you started in translation, in localization more precisely?

Translation, localization, globalization, internationalization, all the, all the Asians. Exactly. AI now and everything in between. Uh, yeah, I, um, well, I grew up in, I grew up in St. Petersburg, Russia, which is a very, um, probably one of the most international cities in Russia, right? Being on the Baltic Sea, it was always exposed to the Western culture. So maybe unlike, unlike other places in Russia, we were always exposed to other languages, for the foreign languages, it was kind of mandatory that you study at the very least English or a couple other languages. As a teenager, I worked as an interpreter for— because it was perestroika time, right? So we started seeing—

let's remind for those who weren't there, it's 1987, Gorbachev.

Yeah, when I was in power, when I was a teenager, like middle school, early high school, it was perestroika time. So I was interpreting for foreign students that were coming to St. Petersburg. So I kind of started working as an interpreter and then I already in school, I was making a little bit of money doing written translations. So it kind of started itself off by itself. And I was studying in a school with specialized, um, specialization in linguistics and foreign languages. And then it kind of was laid out to me. I didn't even need to think. It was like, okay, what am I going to do at the university? Huh? Applied linguistics. Okay, great. What am I going to do for my advanced degree? Okay, how about we go deeper into structural linguistics? I come to the United States. What am I going to do for my graduate school here? Well, okay, let's go deeper into structural linguistics and add a little bit more of computations and a little bit more of, uh, more in-depth linguistic studies. So The choices were kind of made for me. And then as you graduate with a degree in linguistics, you kind of scratch your head like, okay, now, now I've got to make money. And where do you make money with this degree? Aha, there is something called localization. And there is— at the time I was blessed because by pure coincidence I came across people that worked for SYSTRAN, and then I started working with rule-based machine translation. So The path was laid out with very little initiative or thinking on my part, I think. Wow.

Or maybe there was some unconscious choice.

Maybe there were some unconscious choices, divine intervention, whatever.

But it just— Exactly.

The short answer would be it just happened.

Yeah, exactly. And what it is that you do today? Because you're VP of AI Innovation. So what does that entail exactly?

So when I first joined Will Localize, before Will Localize, I was the CEO of the enterprise division of ProMT, or maybe even broader, broader presence of ProMT all over them, all over the West. And when we started talking to Smith, the interest was about, can I help with Will Localize growth and machine translation? Practice. It was 11 years ago, so it started from machine translation, right? Because at the time LSPs were more or less, and I think more, most of the LSPs were at the same place. Okay, well, machine translation is good, we kind of want to implement it, but the feedback from translators was not as excited. So my role at Welocalize at the time was deploying and building machine translation program. And then since then, with the evolution of— so we have machine translation, machine translation gets plugged into CAT tools, right? CAT tools themselves become much more of NLP tools or NLP-enabled tools. So with that, my role at Will Localize and the role of my department at Will Localize has grown to MT to CAT, CAT to broader NLP covering different areas, and then NLP become being a subset of AI, and then broader delivering to the vision of WeLocalize being a fully AI-enabled company. So that's my responsibility.

It's a big one.

Yeah, well, when I say my responsibility, there's a group of extremely talented people behind me. I'm just, I'm just a talking face. For a group of people that actually make it happen together with other departments within the company.

But it's interesting when you say you started with rule-based machine translation, etc., and now today we're with neural machine translation, and there's a lot of talk about this human parity. You know, every year there's someone screaming, we reached human parity. Have we really?

Is it really Uh, you know, if you look at the trend, and, uh, I want to tread very cautiously because I have a lot of friends and colleagues, MT developers, and many of them are the ones that make the statements around human parity. If you look at the trend, usually a statement around human parity is made, right? And then a counter statement is made when the samples or the article that was presented and the statement was made, okay, we reached human parity here, and then it's taken apart, and then it's, okay, but how far away was it from your training corpus, right? Or like, what were the conditions where the human parity was reached, or what was the language pair or the language pairs? So quite often the human parity statement often faces certain counterarguments. And indeed, quite often, if you look at certain sentence lengths, certain content type, you do see that, yes, machine translation can perform on par with human translation. But our best judges If we talk about not machine translation for gisting, but if we talk about machine translation for post-editing, and we are in the localization field, right? Our best judges are our translators, and the fact that we're still post-edit speaks for itself. And the fact that our translators are still giving feedback and they still wish to see higher quality, that still says that we are still are away from human parity. Now, having said that, knowledge-based content, certain user-generated content, shorter sentences, sentences meant for comprehension. Obviously, there are areas where I would say simpler languages, languages that render themselves better for machine translation. There are areas where I can say we are coming fairly close to human parity. And we have a department and we localize program management department. What they do, they measure at a distance, they measure it at location. They measure feedback from translation using human and automated metrics. There's absolutely nothing there that points at us not needing translators. Just the field data, not emotional reaction, but just the very field data.

Exactly. What's your take on this? Is AI really something fabulous? Is it— should we be cautious with it still? What can we do with AI?

I'm extremely just by virtue of, I guess, my personality, I'm mortally afraid of buzzwords in any manifestation thereof. So when, when AI is used so broadly as a buzzword, I'd be extremely cautious, right? When we just say, oh, you know, AI is changing the world, everything is AI, this is AI-enabled, right? When we are saying our business is AI-enabled, you know, AI-enabled business processes, let's look under the hood and let's be very specific what we mean when we say that our processes are AI-enabled. And what do we actually mean when we say AI? I've said it on multiple occasions. If we say that AI is a way for a machine to mimic human behavior, yes, we buy that, right? But when we talk about AI applications, what is it? Does the machine do— does that machine perform human functions in our business processes? Does it really? What do we do? Do we do it for art's sake, or does it really help us solve a business process? If it's a business problem, if it helps us solve a business problem, then I'm sold. Then AI is a quote unquote a good thing. Right. And if we look at our industry, then yes, a lot of things are and have become AI-enabled, right? Content delivery, for instance, content evaluation, content assessment. Like my favorite, one of my favorite examples, for instance, is litigation industry. I'm a strong believer that AI has completely transformed multilingual litigation industry. You do not need— and I speak a lot to patent attorneys about it— you do not need a patent attorney with a red pencil to— sorry, a litigation attorney with a red pencil to sit and skim through 1,000 emails when you can actually use a summarization engine that is going to do the work and render the results right there. You have your perfect AI process. But just to throw out, oh, AI is solving all the world's problems, it's not. It is just like a taxi driver, right? You tell the taxi driver where to take you and you will get from point A to point B. If you don't, your cab driver is gonna drive you around London for, for hours and you will not know where to get. Yeah, I hope I'm making sense, right? Sure.

Absolutely.

AI is great, but you need to know where to get there.

You need to know what to use it for, basically.

Do you— what to use it for? And Set a clear goal. Sometimes if you buy off a PowerPoint with an unclear goal of, oh, I'm buying something flashy, it will deliver something flashy and you will not know what to do with it. Then AI is not going to do you any good. But does it have a great potential and is it solving a lot of problems in the modern world and even more so in the modern, uh, global content transformation world. There are very specific tasks that are being solved nowadays.

Yeah, great. And you also presented at Gala back in March, and you had an interesting project. Do you remember? It had to do with machine learning and language, um, bias language.

Oh yeah, yeah, language.

Do you remember? Do you want to talk about that a bit? It was very interesting.

That's actually the— I mean, the whole— that's actually the bad of AI, right? We spoke about the good of AI, but then there is also a lot of talk about, or is AI actually— what threats and what dangers does AI pose? And one of the dangers is ethics of AI, the bias of AI. Does AI introduce bias, and can AI help us control bias, right? Because if huge trained generative models have been trained on legacy data, right? And AI is only— just like MT, we used to say MT, garbage in, garbage out, right? AI is the same way. If a model, if a language model is trained on trillions of legacy tokens, some of them can be biased and can generate biased output. And then you have your racial bias. Right. And then you have your full vocabulary and then you have your gender bias and then you have your certain non-inclusivity. Right. And non-inclusivity can be an issue. Right. Generative models can generate non-inclusive concepts because they don't know any better.

Mm-hmm.

Because this is how we have operated for years. So what I spoke about, the project that I presented is a localized in-house proof of concept, machine learning. I'm also a little bit dubious around machine learning, AI, AI, machine learning, because AI is a way machine can reason like a human. But machine learning is basically a set of algorithms that help machine do it. So it's our machine learning project that actually helps catch non-inclusive and offensive content. Flag it and propose substitutions. I'm not going to go into, into examples. Actually, it was really fun when I was sitting with our NLP team. We were looking for examples that we could present and we're like, okay, no, this is not presentable. We cannot— no, we're not going to show this on screen. This is even worse. So we went to the like the ones that usually show on screen like white label blacklist, those you can actually catch them in the generative model output or in software UI or in help or in legal like redlines. Even I didn't even know that redline is actually a term that originates from non-inclusive background. A lot of things surface that we didn't even think about. You can flag them. And you can propose a substitution that would be recommended. So yes, this is a project that again, it's still in the lab in a proof of concept, but we've piloted with a couple of customers and to great satisfaction. There are a lot of interesting things, like even things like why do male names dominate? Why do we have domination of male pronouns in certain contexts and lack of female pronouns? And so it's very, very interesting. And that's the reality.

Mm-hmm.

So that's the flip side of AI for you. It is there because it's trained on legacy information. It doesn't know any better.

Okay. So let's, um, just switch, uh, topic for, for a second. You're a Gala champion. We invited you also because we want to know how Gala fits in your work, basically.

How Gala fits in my work. Uh, so the— I've always been, I've always been a strong proponent of Gala and, uh, Willokalize has been a member of Gala for a long time and even pre-Willokalize time when I was on the client side, I was always attending Gala events and during my Primacy time I was always attending Gala events and attending. Now, I mean, the world is virtual and as you said, I was presenting at a virtual— what was— you guys don't call it technology showcase, you call it something else, but whatever that was, Gala Connected. Gallagher, but the technology portion of it.

Yes.

What I really like about Gala and what I think is very helpful for even more so LSPs and maybe even individual freelancers is this is a great open platform for service providers, technology providers and buyers to connect in an open forum and facilitates and fosters an open discussion. So rather than maybe being less in the presentational forum but more being in a discussion forum, if you look at the demographics at Gala, what usually strikes one is how many LSPs how many suppliers and how many technology providers you usually have. If you look at the breakdown of the room, whether it's in an exhibition room, whether it's breakout sessions, whether it's social events, usually you would have demographically— you would have probably significantly more. And I attend a lot of other conferences and active in other associations. But the first thing that strikes you is you usually have significantly higher representation of providers and providers feel that we have an open forum to talk about things that really bother us and things that we really want to talk about and things that we want to connect on and things where we want to give direct feedback to technology providers like, hey guys, this works for us, this does not work for us. And the buyers and the suppliers in the room hear from us firsthand. So I like the openness and that's always been extremely helpful. Ability for the service providers to actually give direct feedback and that open forum and that open idea exchange. So I think that's been extremely helpful in our work to be able to, to be able to have that open forum and open platform. Where I learned a lot and what was a great learning experience for me is being a part of the organizing committee for a couple of GALA conferences. That was really great. That was great learning experience to be in the room and actually be behind the scenes and actually see how gala conferences are shaping up and how presentations are selected and what matters and what people really want to hear about. So that was a, that was a great learning experience. But otherwise, just following the publications, just following, attending, I'm pretty religiously attending all the webinars. So that's obviously, that's, that's, that's a— and the content, the content is never repetitive. And what I really like And I've given this feedback to pretty much most associations that I've worked with. What I really like about Gala content, it's never salesy. It's always informative. There are never— is infomercial the right word? Probably not.

Yes. Promotional.

Promotional. Yeah. Promotion. Like, okay, show nurses. It's early.

Yes, exactly.

Early California morning and my first cup of coffee. But yes, it's always— there's rarely commercial content. There is rarely commercial content. It's really content that you learn from. So I really— I see a lot of value in association. And as I said, if I were to sum it up, always informational, valuable informational content, always the platform where service providers get to express their opinion. Share what matters, get to provide feedback directly to technology providers, and hear directly from the buyers. So fostering this open dialogue and give the platform to service providers and very carefully crafted content for the events. So all in all, all in all, definitely see great value. And it's not— it's very open. It's a very open platform. It's a very open platform and a very open, very open organization. So big fan, big proponent.

Thank you. And would you recommend then volunteering for GALA? Because you said that you were in a program committee, and it is a lot of work, right? We've heard it from other GALA Champions like you that they said it's, it's a lot of work. Would you recommend it?

I would recommend it. And any volunteering, I think anybody who gets involved with a professional association should be very realistic about— you take on an additional responsibility on top of your regular work. But I think, I think it's rewarding. I think it's an association that really benefits the industry. So I definitely recommend it. But, but just like any volunteering, assessing your bandwidth, because there is nothing worse than saying yes, I'll do it, and then basically just decline, decline, decline when it comes to, when it comes to meetings, when it comes to planning sessions. So If the bandwidth allows, I would definitely recommend volunteering.

Thank you. What accomplishments are you most proud of?

Holy Lord, I'm just the worst person to talk about, talk about myself and my—

Would you like another question?

No, I can take this one. I can take this one. Well, I mean, the first accomplishment, uh, I mean, I came to this country very young and very young and as a single mom. And I guess getting, getting an advanced degree from UC Berkeley was probably, I'd say, qualifies as an accomplishment.

Definitely. Absolutely.

That would probably, that would probably, yeah, that would, that would work. Learning to drive a car. Getting my driver's license from the 4th attempt. And I had to promise to the guy that I would never drive in the part of town where he lives. I mean, he just looked at me. He's like, OK, I understand you need a job. I understand you need to get around San Diego. The public transportation here is horrible. But swear you'll never drive where I live. So finally getting my driver's license definitely qualifies as an accomplishment. And 20 years— 27 years fast forward, I haven't improved much. But getting the driver's license is an accomplishment.

Absolutely.

So, uh, starting and running my own music promotion company as a side business for many years alongside my localization career, and I'd say pretty successful music business, is definitely an accomplishment.

And absolutely. What kind of music? Now I'm curious.

Actually bringing bands from the former Soviet Union and taking them on tour around North America is something myself and my partner, who also has a full-time job on the client side, that's— I think that is an accomplishment.

And considering your career then, what advice would you give to someone who's just starting out in localization?

At this point, at this point in the world and where we are, I mean, first of all, I think that we live in the most exciting times for our industry. And never, ever, ever listen to whoever says that the translator's profession, or maybe language profession, is going away. If anything, it is evolving. And I would always say that, and I'm always saying that AI is not a killer of our profession. AI is an enabler of our profession, opening infinite opportunities. So my advice to somebody starting in localization or language industry professional: absorb as much tech as you can. I'm not saying everybody should wake up a computational linguist tomorrow, but learn how to interact with NLP applications, learn how to interact with AI, look at it with an open mind, realize that AI is not going anywhere. It's only going to be evolving and think, OK, what am I going to be doing tomorrow? Am I going to be validating output of AI applications? Am I still going to be doing some transcreation and The creative part is not going anywhere. I'll still be writing marketing copy. Am I going to be dealing with generative models? Are my CAD tools going to be evolving completely? Will I recognize them tomorrow? But they will still evolve. So basically looking at it with a modern mind and follow the evolution of AI technology as it applies to our industry. If you are a PM, learn how to interpret data. If you want to be in the financial field in our industry, realize that finance are impacted by our— our entire ecosystem is applied by next-gen technology. So approach it with an open mind. Realize that it's a new profession, but it's still an old profession of the language tech. So that's my advice.

Yeah, and we are talking about skills, so it's clear that skills are evolving very rapidly. So Do we need to change the collaboration between academia and industry? And if so, what kind of shape or form should this collaboration—

Absolutely. I think, and you probably would agree, what was happening for years, there were a couple of two separate islands, right? And that was most visible at conferences. Like if you go to MT Summit or ACL, or you go to like say AMTA, And you see, like, there is an academia island, right? And then there is the industry island. And even the tracks and the selection of paper for tracks would be different. And I think that the convergence is absolutely necessary. And it's great to see that, like, American Translators Association or their European counterparts are doing so much to bridge that gap. But the convergence is absolutely necessary. And it's great to see that, like, say, in Monterrey and other places, So many courses are being modified to introduce technology into the translator degree or into project management degree. Localization Institute is offering so much now. So, but the convergence is absolutely inevitable. Otherwise, we will not be able— people will not be able to graduate with relevant degrees.

Yeah, absolutely. Thank you for that. Um, another question: where do you find inspiration? Huh, I don't know, while driving probably.

No, no, no, no, no, no, no, no, that's more, that's more like where do you find sheer panic. No, that's no, but I mean, first of all, I absolutely love what we do. I really passionately love our field, like I wake up happy. It's one of those— I, I, and I'm, I'm being absolutely honest. Like, I wake up, I don't go to work. I wake up to the list of interesting things that I know myself and the team. And I hope that my team is— and people around me and people I work with in the industry are as inspired. So I just think, as I said, I think that we live in such an exciting time. It is just the excitement of what's going on around. So that'd be one. I have this ambition of writing a book about what's going on around. So it's also like this little worm in my head, like, oh, that would be cool to write about. Oh, and that's not to forget. And that would be cool to remember. So there is also this little ambition. Oh, and by the way, one day I'll be a published author. So that's my other source of inspiration. And then just little things around. I don't know. I'm blessed to live on the beach, so small things like beach walks and doing yoga and, um, I don't know, spending time with my son and having long philosophical debates with him. Just, I don't know, a lot of things. Just the beauty of day to day combined with the excitement of what I do. And as I said, this little ambition of, and one day I'm just gonna be damn famous. I guess all those things, all those things combined.

Yeah, absolutely. Thank you, Olga Beregovaya.

Thank you. [MUSIC] [MUSIC]