Working with AI in Language: What Can Go Wrong?

In episode 2 of the Welocalize podcast, Louise Law is joined by Olga Beregovaya, VP of AI Innovation at Welocalize and Board Member and lead tech mentor at Women in Localization. They chat about AI in language, what can go wrong and also how linguists can make the move into AI and develop their skills.

On this podcast episode, we cover:

1. How can global brands incorporate AI to transform their multilingual communications?
2. What can go wrong with AI programs?
3. How can linguists get into AI and enhance their skills? 

Transcript

 Louise Law

Well, hello and welcome to the Welocalize Podcast. We'll be talking all things multilingual localization, translation technology, AI, machine learning, NLP and much more, helped along by a wide variety of guests. I'm Louise Law, your host, and in this second podcast episode, I'm chatting with Olga Beregovaya, who is Welocalize’s VP of AI Innovation. Olga is a well-known figure in the world of language technology, AI and data development, having worked in the industry for over 20 years. She has a master's degree in Linguistics from St. Petersburg in Russia, and UC Berkeley, and has been specializing in NLP, which is natural language processing, and language technology ever since. She's currently on the board of directors at Women in Localization, leading their technology initiatives, and she's also the former president of the Association of Machine Translation in the Americas, and speaks regularly at leading conferences on topics like ethical AI, and bias and the metaverse. I think it's fair to say she's pretty visionary with that long and glittering introduction, and she's also mentored loads of people going into the field of machine learning and AI and has helped them develop their careers. So, welcome, Olga, and thanks for joining us today from the West Coast of America.

Olga Beregovaya

Thank you. Thanks, Louise, for having me. 

Louise Law

And it's really nice to speak to you. We know that AI and overall tech enablement are no strangers to the language industry, and AI, it's a hot topic that's talked about a lot, and we chat regularly, Olga and I, on the various ways we can use AI to help our global clients to grow globally and reach international audiences. And we know that AI is becoming really important in the overall localization and translation workflows. It can make workflows quicker, it can improve LQA, and we can also use AI to develop training data to power multilingual applications like chatbots, and voice assistants, helping automated voice and text applications deal with people all over the world. But there's still a lack of knowledge, and maybe direction, in the field and how it's actually used in a commercial setting. So, could you, to kick us off, do you think you could tell us how global brands and businesses can incorporate AI to transform their global and multilingual communication?

Olga Beregovaya

Well, first of all, yeah, I completely agree with you on the buzzword subject. There's a lot of talk, every conversation, every conference, inevitably talks about AI, and it is very important to unpack the buzzword and separate it from, okay, what can actually be done and where your organization will actually benefit from AI? The organization will not benefit from implementing AI randomly, “Oh, this looks shiny, this looks interesting, this will optimize our processes by 25%. Let's plug it in.” This half-baked chatbot will help us with our customer interactions. First and foremost, it starts from strategy. Just like a couple of years ago, that was the same for machine translation. Machine translation is no silver bullet. Just plugging in, throwing in machine translation is not going for all content types, post-editing or even more so, throwing in raw machine translation is not going to get you anywhere. You need to have a clear content strategy, and in this case, clear process strategy. And the same applies for AI. Sit down, identify your process gaps, and see this is where human labor is too costly, and this is-- run some pilots, do some tests and see, indeed, implementing AI-based processes do streamline our processes, save us money, and this is where human work can be important, not replaced, but augmented by AI applications. 

Louise Law

Yeah, I mean, it's like the brands need to look at the AI headlines that we see in all the online news, but then they need to look and think, “Oh, what's the strategy for my company? How's it going to benefit my business?” 

Olga Beregovaya

Right, “What will benefit my business?” And then when we say AI, there is AI as it applies to data analysis, AI as it applies to workflow optimization, AI as it applies to natural language processing. So, let's not blend it all into a single bucket. Let us clearly separate the messages that we are being overwhelmed with. Now, at some point, you will be able to recognize the more AI information you're overwhelmed with, the easier it will be for you to separate them into buckets. Which bucket is, am I being told about, and see where it actually applies. 

Louise Law

Easier. 

Olga Beregovaya

The easiest is to maybe engage an external consultant who will help you with SEO strategy. But most important strategy, understand clearly what the promise is, unpack this promise, and potentially start with trying it in your workflow and understand what the benefits will be. Start from benefits and ROI. That's my advice. And then there is this whole side of, “And as I implement it, how do I remain ethical, how do I preserve my workers?” [crosstalk] ...and how do I preserve? If it's AI automation of a physical process, how do I promote and preserve my workers’ and end users’ emotional and physical well-being? So, that's definitely a huge part of the deployment strategy. 

Louise Law

That's a big topic, and I'd probably like to invite you back on another episode where we can really, really talk a little bit more about ethical AI and how that's really affecting today's world. And that's really important, in addition to all the noise, and people love to project what AI could do, but you're working, got your sleeves rolled up, you're really deep in the weeds, working on a number of like quite significant AI implementations. With these AI applications and all the data development and deployment that there's quite a few things that can go wrong, aren't there? It's not all just plain sailing. Can you give us some insight on what could go wrong with some of these activities, or what can go wrong, not what goes right? 

Olga Beregovaya

I was just speaking about it at one of AI conferences, and the team, the team was the AI, in this case, the AI data development team, they were our AI services team, to narrow down a little bit. Our AI enablement team were extremely helpful, pulling together actual real-life examples of what can go wrong. And my opening line for the presentation was, “You guys do it in English, you know how to break it in one language. We can actually work in between 525 translation directions, we can break it for you in 525 languages. But the good news is, we can also fix it for you in 525 languages.” But what can go wrong is, a lot of times, it's very similar to localization challenges when it comes to engineering. It's very similar and you can actually learn from internationalization practices. And that's where we add a lot of value because we come with 25 years of internationalization and globalization, and global content engineering experiences and everything else, like bi-directional languages. Do we know how to model in bi-directional languages? Yeah, we can guess, but it's kind of double trouble because it's both data development and all the concerns around bi-directional languages, Japanese with four scripts, Japanese with four alphabets. 

Louise Law

Of course, yeah. 

Olga Beregovaya

Everything else like Slavic languages, UTF-8 compliance, so all the engineering joys of languages other than English, so that now if you develop content for spoken modality, there is, “You want your bot to understand your speakers, right?" So that's right there. Your bot may not understand your speakers, if you don't write for spoken language. If you don't introduce slang, if you don't, like what if the user says, “Mm-hmm, haha,” how do you-- you need to be able to capture it in your data. So that can easily go wrong, if not done. And a lot of other things we didn't even speak about until we got into AI data development. Now, if you go to the other side of our AI capabilities, for instance, us automating LQA process, and us automating decision making within Pantheon based on AI features. We rely on the prediction, but what if the prediction is not accurate? And we already made a decision that this content should go into a certain workflow based on the prediction. So it's important that we fine-tune the predictions until we can completely trust our model, what content goes to a sub-optimal workflow? So, AI is, I wouldn't say it's in its infancy, but AI is definitely a young field. [Anything] can go wrong when you rely on the model, it’s not necessarily completely mature. You can go wrong when you have not, in the data world, when things are not taken into consideration, which are inherent for languages other than English. And yes, I have now thought about them. 

Louise Law

Communication and languages, they're constantly changing, constantly evolving. As you said before, people use slang, people use different vocabulary. And so, the kind of the AI applications need to be continually keeping up-to-speed with that, which it's almost like a moving target really, isn't it? 

Olga Beregovaya

It is a moving target, and it's also a lot of knowledge one needs to have around all the languages, for instance, that are being commissioned for us for both data development and languages for which we want to optimize the LQA process and reduce the LQA process where it's not necessary, or work, for instance, with edit distance, where we can shrink different steps or eliminate different steps where we see that edit distance does not-- a different process step just do not add any value. But again, we rely on automation, and automation can sometimes be wrong. So, we can say it's 75% accurate and 25% we just put our trust into it, not necessarily knowing whether it's right or wrong. So, there are a lot of things that can go wrong, but there are a lot of things can go right, so it's our decision, where do we set the threshold and where do we trust automation, we trust AI? But that was the same with early adoption of any kind of NLP, any kind of application. What can go wrong, I think, is one of our favorite topics, because you have to go backwards from, “Okay, this went wrong. We are not about talk, we're about action. This can go wrong, let's work backwards from it. We can fix it.” 

Louise Law

It's a good topic to talk about what can go wrong and you move forward and learn from those mistakes, don’t you? I know one of the areas that I want us to talk about is you're really keen on mentoring new talent and talent that feel that they have a calling towards like NLP and AI and I know your team, you're really good at identifying talent and helping them move into these areas and mentor them. And I love the fact that Welocalize, your Welocalize team has really evolved in houses, growing number of machine learning engineers, and AI, and NLP specialists. And I think this reflects just how we're working with global brands, and the fact that they're using technology more, particularly in their language and localization programs, just to kind of manage better, do things better, faster and smarter. They're looking to technology and data management to help create better brand experiences. I think if there's one thing that companies can do to get started, what would be your one piece of advice? So that's my first question. And then my second question would be, if there was anybody who would be looking to get into the field of NLP and AI, what do they need to do? So, what's the one thing companies can do to get started? 

Olga Beregovaya

I think it kind of goes back to our initial question, how can the company best implement AI in their process? What can the companies do to best get started? And I'm pretty confident that again, the company is just going in and saying, “Hey, tomorrow, I am going to plug in AI in our processes.” I think the best thing that companies can do is actually engage with trusted advisors, like ourselves, trusted advisors and partners, identify, do discovery, share their gaps in their processes, and say, “This is where we spend too much money. This is where the manual part, the manual work, is too high, too much, and what can we do to automate it?” Rather than finger in the sky, “We've heard AI is good here, and we've seen animals like this.” That would not work, but really bring in a team of SMEs, solution architects, just for a single discovery session, and just offload their, “Okay, here are our griefs and concerns and, and gaps. And here, this is where we believe AI can help.” And just hear from that like, okay, and just dissect the areas, “Okay, this is what's going on within our business. This is what our--" Bring in different stakeholders, like, “This is what we've done recently, here is HR,” for instance. Who would have known that digital HR is actually looking for automation, and HR-enabled automation? So, here is HR, here is sales, here is support, here is Dev. There wouldn't even be that many people in the room, like the starting point, that many people in the room. And just interview them and talk to them about the gaps and what can potentially be optimized, which would not be very different from any other automation discovery, and just listen to the SMEs and see, “Hey, guys, this is what we could actually do for you.” Yeah, listen to their problems, listen to their spend and see, okay, look at the current tech stack and see, “Okay, this is where AI would be of great help.” And maybe they would already, like they've heard it somewhere, they've listened to webinars and podcasts, listened to their customer and proposed our solutions and maybe they already have ideas, and possibly, they already have some AI implementation onsite and we can compare notes and see, “Okay, here is, say, your monolingual AI and this is where we can help you make it multilingual.” Like for instance, it is quite often that they have a monolingual chatbot. “Hey, here is your monolingual chatbot, but you guys are present in multiple markets. Let us help you go multilingual.” It depends on the company's level of maturity and starting point, but what we have learned over the years, actually, the best conversation is had when we're all in the room, and we're having a discovery session. 

Louise Law

So Olga, you just on to the second part of that question, if there are people with talent out there who were quite keen to get into the whole AI, NLP and machine learning field, what advice would you do to help them boost their careers? 

Olga Beregovaya

We at Welocalize have seen quite a few success cases where a translator entering, at the time, a machine translation team was actually just learning at the job, entering at a more junior level as a coordinator, and eventually growing into a program manager, trainer, knowing all the intricacies of machine translator as a user. So my advice, no matter where you are, like say you are a project manager, say you are in a translator career, and we have had people coming from all ways of life and localization. The easiest way, as we have learned it, is just being in the field, reaching out to the team that is working in NLP, AI, machine translation, and just applying for a more junior role. Yes, you are taking a step back in your career, but you're joining a new team with an expectation of applying what you already know from the industry, and learning a new skill from the team that you're joining. So, that would be my first advice, say you are already in the industry. Obviously, if you're coming with a degree in computational linguistics, data science, if you already have some knowledge of the industry, again, what we have seen and what I have seen that has worked miracles is, join as an intern. Join as an intern while you're still at school, say you are in graduate school. In graduate school, see if you join as an intern, and see if your team collaborates with you on your master's thesis, or on your PhD thesis, and if you take your work material as a material for your dissertation, you are already a part of the team. The data from the team and the materials from your day-to-day are already a part of your learning experience, a part of your dissertation. So you're already an organic part of the team, and the team is very keen on having you around because you have the knowledge and you have the academic knowledge. So you're entering from the academic background into the commercial enterprise field. So that's another way of entering the field. You are in computer science, all the NLP, R&D teams are generally, in my experience, hungry for solid back-end engineers who actually stitch all the NLP applications together. So, being a software engineer, you actually enter the field with software knowledge without NLP knowledge. And then sort of taking classes through day-to-day, you actually build up your NLP, machine learning, AI expertise. So there are multiple ways, multiple angles, you always bring something to the table, and the team is always willing to teach you because there's always something the team will see in you, whether it's multilingual expertise, post editing expertise, computer science expertise, or you are in school, but you have recent academic knowledge. What's also beautiful about recent graduates, whether it's bachelors or masters, you guys know something we don't know, because we've been in the field for too long, and you know all the latest and we don't. 

Louise Law

There's a lot of opportunity out there, so I think it's a healthy field to get into, isn't it? 

Olga Beregovaya

It is. It absolutely is. And when my friends whose kids are starting college now, when they ask, "My child is interested in computer science, or my child is interested in languages and what should they study?" I always say, “Well, hey, how about we go into machine learning and AI?” And obviously, what I recommend is, "And if at some point, they decide to specialize in NLP, that's definitely the hottest field right now." 

Louise Law

That's the future though? 

Olga Beregovaya

Yeah, it is. It is. It is. 

Louise Law

Brilliant. Oh, well, Olga, it's been really, really good as always, to chat to you, and thank you so much for joining us, and I'm sure I'll be inviting you back to kind of do more of a deep dive on some of the topics that we touched on today. So, thanks very much for joining us. 

Olga Beregovaya

And thanks again for having me.