EP 203: Translation in the World of AI - Will we have a job tomorrow?

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

If you want your company to compete on a global stage, you need to be able to speak to global customers and global clients. I think a lot of us don't understand how that works, but there is an industry that makes sure that we can all speak together and that's the translation. Right? So I'm very excited today to talk about how AI is going to impact that translation industry. Are there still going to be jobs there tomorrow? All right, we're going to be talking about that today and more on Everyday AI. Welcome. My name is Jordan Wilson and I am the host of Everyday AI. And this is for you. We do this literally every single day, bringing you a live stream podcast and our free daily newsletter helping everyday people learn generative AI and how we can all leverage it to grow our companies and to grow our careers. So, yeah, maybe your company is ready to compete on a global stage and you're worried about how can we take our language, how can we take our product and our services and better cater to a different country, to different regions? So I'm excited to talk about that. But first, before we do, let's go ahead and go over the AI news as we do every day. And if you are listening on the podcast, thank you. Make sure to go to your everyday AI dot com. Sign up for that free daily newsletter and if you're joining us live like we have Tara joining us from Nashville, Woozy from Kansas City, Brian from Minnesota. We got everyone joining. Jason, thanks for joining us from Florida. Make sure to get your questions in. What questions do you have? Like maybe Mabrit here who's, you know, recently worked as a freelance translator. What questions do you have about AI in the world of translation? Make sure to get them in. All right, let's go over what's going on in the world of AI news. There's actually a lot today, y'. All. So first, LinkedIn is bringing an AI chatbot to job seekers with LinkedIn Premium. So LinkedIn is rolling out its new AI powered features to help job seekers navigate the modern job application process and gain insights into job opportunities. So the new features include an AI powered chatbot for job listings and personalized career advice. Based on a user's LinkedIn feed, these new tools offer advice on how to improve your profile and stand out to potential employers. So right now these features are just being rolled out and it's only to premium users for now. Speaking of new updates, we have new AI image editing features rolling out on Microsoft copilot. So Microsoft is updating its Copilot platform with a new design and enhanced image editing capabilities, bringing sustained growth to their Edge browser and their Bing search engine. So the focus is on productivity and creativity, rather than one competing with Google for market dominance. So so far, 5 billion images have been generated through the Copilot platform since its launch a year ago. It obviously uses the Dall E image generating powers from its partnership with OpenAI. All right, last but not least, Google Bard is officially gone. Well, it's just kind of changed its name. Google Bard is gone. Google Gemini is here, and as is Google Gemini Advanced. All right, so Google has released its new updated AI chatbots, Gemini Advanced, as part of their Google One premium subscription package. So it is kind of Comparable to OpenAI's ChatGPT plus and promises enhanced capabilities for understanding and completing tasks. So access to Gemini Advanced is available through Google One's new AI Premium plan, which costs $20 per month right now, though. Hey, speaking of translation, Google Advanced is currently only available in English, but may expand to other languages in the near future. So, yeah, if you log on probably sometime today, if you use Google Bard, you're not going to see it anymore. You're going to see Google Gemini. So that is the base model and then the Gemini Advance, $20 a month, bringing you a more powerful version in using Gemini Ultra 1.0 versus the free version that just uses Gemini Pro. All right, so many buzzwords, so many words. Speaking of words, we got to learn how to translate them. All y' all like and talk about specifically, how is AI changing how we communicate on a global scale stage with global audiences? All right, so I'm excited for this and let Please help me. Welcome to our show. There we go. We got her. All right, so Olga Baragava is the VP of AI at Smartling. Olga, thank you so much for joining the everyday AI show.

Thanks for having me.

All right, and tell me a little bit what you do as the VP of AI at Smartling.

I was actually listening to a description of another company that defined themselves as the trust layer between everything that's happening in the wild west of the AI world and the actual end users. I would think about Smartling and what I do in the exact same way. There's so much happening in the world of AI. Large language models are called language models for a reason. They're built of language. So what we do, we actually help our customers navigate the waters of language and global language and help plug in AI and make sure that they harvest the benefits. Right. So actually one of the taglines for recent industry presentation was from Frenzy to Trust. And that's where I would probably put us. So that's in short, I mean, the shorter answer is we provide AI powered translations.

Okay, I love that. And maybe tell us a little bit for those of us who maybe don't understand how this industry works. Like what in in general is, you know, translation, you know, because I think sometimes people think of, you know, oh, is this just Google Translate? Oh, is it just when you, you know, go to a website in your browser and it's, you know, all of a sudden it's automatically translated into a new language? Like, give us a little bit of background of, of how the industry in general works. Is it all just humans like, like yourself who are reading something in one language and making sure everything is kind of perfect in many other languages?

So what is translation? I mean, first things first, there is absolutely nothing wrong with Google Translate. And there are a lot of times where Google Translate or Microsoft Translate or Amazon Translate or DeepL or any other engine are perfectly fit for purpose. And I'll touch on that a little bit later. But in general, there is much more to this industry. It is not just words, Right. If you think about all your content, it resides in content management system dam somewhere else, right? It lives. Your website is also populated from somewhere. Possibly you could stitch those things together manually in one language, that being English. Now go try doing it for 105 languages if your company is present in 105 countries. Right? And that's basically what the industry is about. It is definitely, again, by virtue of we handle words and there are a lot of words words in this world. By virtue of that, we're definitely on the forefront of the digitization and the digital revolution. Just for the sake of necessity, if you take your company global, your content is multiplied by 105 or as many markets where you operate. So there is technology to it. There is linguists, and I was very happy to see fellow linguists in the audience. There are linguists, there are internationalization engineers that make sure your code is actually global from the get go. There are subject matter experts that make sure that things are factually relevant. So it's actually a huge production to get to your website in Estonia. There are many people working on it and AI to our rescue right now, things have become significantly easier.

Yeah, and let's talk about that. So with, you know, this kind of recent Boom of generative AI. Large language models now being incorporated into our everyday lives, whether we even fully realize it or not. Right. How specifically is, is this impacting the translation industry?

The impact on translation industry can definitely cannot even be overstated or overestimated. Again, I'll go back to the concept of words and linguistics. Right. Large language models started from words. True. Now we have multimodal model modality language models, but in principle they help us handle worlds. You mentioned something in the opening when you were talking about the news, right. That a lot of alums are now multilingual. Right. So that alone helps us because AI can now handle translation. The question will have jobs tomorrow will be covered later. AI can handle translation. If you don't want to translate, you can actually even generate source content. And we see a lot of writing copilots, writing assistance, but equally you can actually even generate the target content. So the whole paradigm of how we do business shifted tremendously with the advancements of AI. It's using technology in the translation industry is no news. I think the paradigm shift is before it was computer assisted translation. Now we can actually trust AI with doing the heavy lifting and we do what's interesting and we do where the human reasoning is absolutely necessary. So I would say that's probably, that's the main change, that's the most dramatic change. AI is a copilot and enabler of the translation process.

Sure. And you know, I'm curious, Olga, because I sit down and like I think about the first time that I saw or used the GPT technology, which I think for us and our team was, you know, 2020 when it came out in some third party products. Right. Do you remember like the first time that, you know, you kind of saw whether it's, you know, the GPT technology or a large language model. Like I'd love to hear, you know, as someone that works with words every day and works with, you know, making sure words make sense together, what was your first reaction, you know, kind of when you saw large language models kind of bust onto the scene?

I would say that, I mean, first things first. Transformer models in general are no news. You just mentioned Google Bard and then there were birds and there, I mean there are so many. Well, Bards is more recent, but let's touch a couple other transform models that have been around for a while. Even neural machine translation is based on transformer models. So in principle that has been around for a while. What happens, what drives the breakthrough is the trillion of parameters and tons and tons of data points that the later generation of large language models has been trained on. So my immediate reaction, I think the first real large language model, like true large language model I saw that was probably GPT2 and we're talking what, two generations of GPT ago. My immediate reaction was holy Lord, this thing can do everything. How am I even staying employed? But yeah, so I mean the first thing was like, oh wow, what do we do now? What do we do now when it's so capable, like write me Shakespeare style poem, Voila. Give me a summary of such and such. Voila. So there was first this wave of fear, meeting excitement, but then you start unpacking it. And as you start unpacking it, you actually realize that along with massive capabilities come massive shortcomings.

You know, like I'm almost thinking that maybe your industry is ahead of so many industries, right? Just because I feel that anyone working in the translation space is always, you know, like you just said, keeping up on, you know, transformer models. You know, I, I hadn't even used GPT2, I only used GPT3 when it first came out. So like, how would you say the translation industry has so far successfully used, you know, these large language models? And then maybe what are some of also the red flags that you've seen or the industry has seen by using these models?

Okay, so there is, you said initially there was this sentiment about let's just deploy Google translator across the board. The initial sentiment and I, I would imagine there are certain, certain group of maybe executive and business leaders in our audience. The initial sentiment, like from me running balance sheet or running the P and L sheet, the initial reaction was let's just plug it in. GPT is awesome. Let's just plug it in. It does translations, it is multilingual. So the initial reaction was again, it will do everything and successfully to a great extent, large language models have been successfully deployed for translation cases. As I said, generative AI is deployed quite a bit. When it comes to digital marketing, Digital content summarization is another very successful case. Right. When you are an attorney, say you're running an ediscovery case and that's a prominent part of our industry clientele is actually attorneys doing a discovery. You are dealing with massive, massive volumes of emails and they happen to be in Japanese. Do you have any time whatsoever to read through terabytes of emails and do you equally have sufficient knowledge of Japanese? I don't. So this is where translation and summarization combined services, multiple industries, another application. And I want to pause a little bit on the Limitations. Think about machine translation and neural machine translation. It was built for purpose and for a single task of translating, right? And you can also train it. You can customize to your company's tone and voice. There are tremendous capabilities around neural machine translation, which by the way happens to be NLP application, natural language processing application, and subsequently equally happens to be AI. So think about neural machine translation. Now compare it to a model that was built to perform variety of tasks, starting from math problems to hey, write my college essay, except my teacher friends are telling me that now they're getting identical essays like 17 of those. So there is a bit of a question mark there. But again, if you deploy LLM that is built for multiple purposes and then you deploy something that was built for purpose, all often neural machine translation still wins. So there are quite a few cases where you still want to default to machine translation and all that customization capabilities. Having said that, machine translation is flawed when it comes to fluency. And those of you having used like I tried to use Google Translate, I spent quite a bit of time in Mexico and Google Translate is not always my friend when it comes to understandability and fluency of translation. LLMs are tremendously fluent. So the best successes our industry sees is combining the best of both worlds and the convergence of Gen AI and other NLP technologies is where we get best results. Okay, let in translation, do what it's good at, which is factual accuracy and then you apply gen AI on top of it and you get factually accurate fluid translation.

Are you still running in circles trying to figure out how to actually grow.

Your business with AI?

Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Gen AI. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use Gen AI. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you can partner with us too. Just like some of the biggest companies in the world do. Go to your everydayai.com partner to get in contact with our team or you can can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a Straight path to ROI on gen.

Explain it so well there. So, okay, we actually have two questions that are pretty much the same here, Olga. So Raul, thank you for the question. And Tara are both just kind of asking, so how, how can AI accurately translate idioms in context and clarify this to the reader? Right, yeah, when I think especially in the English language, but I'm sure it holds true across the board in other languages. Sometimes we have weird sayings or words that mean multiple things or descriptive ways to say something that is like, oh, that's like word for word, that doesn't really make a lot of sense. So how does the AI or the translation industry kind of tackle idioms?

I know that at some point about like five years ago, US government was pouring a lot of money into getting machine translation. And I'll start with machine translation to understand idioms, euphemisms and metaphors. And that was not the most successful project because you're absolutely right. The way machine translation is designed is not as much word to word, but it's still the context window is very limited. The attention mechanism of machine translation only covers this much. Now the beauty of applying Genai is that it does indeed read and translate your content in context. And the later generation gen AI, as we know, different LLMs have different context windows, but you are still able to feed your context. You still go beyond a sentence, beyond a paragraph. So what we see is because of the context, you can disambiguate terminology. It's accurate, right? It's not going to be the bed for like flower bed. It's not going to be translated as a sleeping bed. If Genai sees that. Actually, hey, we're talking about gardening here. So in terms of context, that's what allows Genai to better capture idioms, metaphors, and in general understand more just because it's presented with more. Just one word of warning, don't give it too much because then attention mechanism gets confused and just translates the beginning and the end and that complete gets completely lost in the middle. But if you've nailed just the right amount of context you need to give it, that's when the user will benefit from understanding idioms. I hope I answered the question.

No. Yeah, you did. And it's funny you talked about nailed. Like I always think of like an idiom, like hitting the nail on the head and I'm like, yeah, how would that translate in so many other languages, you know, 10 years ago versus I'm guessing now today because of AI. It's much better and it's much More accurate out of the box. Right. Like you don't as in your, in, in your industry, you don't have to spend as much time, you know, checking those because presumably the idioms, similes, metaphors translate better now.

They would again based on the context and Also because Gen AI generally generate, generally LLMs are much better at NLU natural language understanding that NLG, you don't have much room for hallucination. So if you're provided enough context, it would actually predict that, hey, I'm dealing with an idiom here. I don't really take a hammer and nail, hammer the poor nail on the head. So because I'd say the context window and also ability to identify what text I'm dealing with that I'm actually dealing with more idiomatic text as opposed to user manual to a washing machine. Nothing against user manuals, by the way.

Yeah, we need them every once in a while, right?

Have you read one ever? Okay, anyways, yeah, let's, let's talk about.

The big question here, Olga, because this is the title of the episode how is this going to impact jobs? Because from someone who isn't in the industry and you see large language models in their flexibility, their power, even like what you were just saying, their ability to handle idioms, you know, similes, metaphors across multiple languages. So what does this mean for the many, many humans who are working in this industry now that large language models are, you know, doing this fairly well? What does it mean?

I'm going to have an English as a second language model moment here and ask you, can you say in English we cannot play ostrich in as we cannot stick our head in the sand and pretend that nothing is happening. Now I'm translating from my mother tongue. We would be very naive to say that jobs are not going to change in the translation industry. That would be a very, very naive approach. Of course they will. AI translations. AI powered translation, machine translation, what's called smoothing, which is basically post processing output of other applications with AI, it hasn't reached human parity yet, but it's on its safe way there. It is eventually going to reach some somewhat something that resembles human parody. Now what's going to happen? We do know that models hallucinate, right? And we do know. And it's funny, there is a fun fact. Majority of us built models are built on English phenomena and local phenomena. Right? So what, what would happen is really funny. It would have enough words in Italian to express the context, but concept. But it's not going to have enough anthropological Knowledge to actually reflect Italian phenomena. Job one, fact checking and validation. You are an Italian native and you are told that, I don't know, the best Italian football team, soccer team is Barcelona. Something is off here. Right? So fact checking and validation are definitely jobs that are going to be around for quite some time until the models have enough data and enough world knowledge to produce accurate output across languages. 1, 2. There is something that's called false fluency when the sentence is fluent, again, a little bit on the topic of hallucination, but means absolutely nothing and models are still able to produce. I don't know, whatever. I'm going to the store. The crocodile went to the river. Zero relationship between the two parts. So absolutely, post editing, human, human validation, fact checking. Translators will probably gravitate more and more towards specialized subject matter expertise. So that would be one for you.

Okay.

Project managers, AI based predictions. Do I send it to translation? Do I push it straight to publishing? Project managers in our industry are becoming more and more of data analysts validating AI based decisions. So here is two for you from business. From project management, like hands on moving files around to actual business analysis.

You know, Olga, I'm curious because I'm guessing that there's certain companies in the translation space that are doing this the right and proper way and taking ethics into consideration and then there's maybe those that aren't. Is there actually a more like an increased responsibility on the humans in this process to be even more vigilant? Right. Because maybe now, you know, your company is able to do 50% more or double. Right. And. But maybe things might slip through, you know, a little, a little faster or a little more often. So maybe is the role of, of humans in this space maybe much more important than normal? Because it's much more likely that, you know, errors couldn't theory slip through.

Absolutely. And hence, I mean, first of all, you're spot on the topic of ethical. AI is extremely prominent and extremely visible in our industry. Exactly. Because it's so human driven that you absolutely want to deploy your AI responsibly, making sure that you do have humans in the loop to mitigate potential risks. There is a term that I love which is toxic. Again, it's funny because I'm supposed to be a proponent of AI, but I also know of all the shortcomings and the role for the human. So you want to mitigate bias, you want to mitigate toxicity. What do you do? You actually become a part of the process and you help curate the training data and validate the algorithms of large Language models, making sure the bias and potential toxicity is mitigated. So ethical AI deployment is definitely a huge topic for us. It was a recent case when I don't know if it was GPT or machine translation was used to make decisions, immigration decisions on the border based on the input, based on translation. And hey, there could be better applications than actually make decisions about human lives based on what a model can produce. So, I mean, there are things. And it was, it was very, very, it was very, very prominent case that obviously raised awareness of deployed when it's fit for purpose, when nobody's life or nobody's livelihood is at risk.

Yeah, yeah. I mean, yeah, in, in that case, the stakes are extremely high, you know, to have accurate translation.

Yeah, I know we're running out of time. So I want to mention one other thing where translators and linguists are becoming extremely, extremely important, and that's prompt engineering. The model is like a taxi driver, right? I mean, the model will take you only where you tell the model to take you. And it's all about prompting. So there are only this many data scientists in the world, but there are a lot of linguists in the world. So a linguist can actually help design the right prompt that would produce the right output for their target language. Formality wise, gender wise, tone and voice wise. Right. So there is a huge role and I see a lot of linguists developing prompt engineering skills because it's an absolute necessity. And again, I don't speak, I don't know Bahasa Indonesia. I cannot get by without a local expert helping build the prompt and validating that it's producing a relevant output for their language. So it's not that the jobs are going anywhere. Tons of jobs surfacing and emerging.

That's okay. That's such a great point. I was actually having a conversation at a tech event last night. You know, this, this gentleman asked me, hey, my daughter's in college. Here's what I think she should be focusing on. Said, what do you think are the, the most important skills in AI? You know, is it, they said, is it prompt engineering? Is it this and that? And I said, you need to have strong language skills, right? Because if you have strong language skills, strong writing skills, it can take you very far in this world. So, all right, let's, I think we have a couple questions. Maybe we can go rapid fire. We can see if we can get some quick question and answer here. Olga. So let's, let's give it a try. So woozy. Asking any thoughts about some of these new stories you Hear about them translating lost languages or some of the groups that are working with understanding language and animals. What are your thoughts on that one, Olga?

I know a lot of my friends started teaching their dogs how to speak. I'm not quite there and I don't nurse high hopes about my specific dog, but let's. Okay, animals aside, long tail languages spot on. Actually, Geni helps us take language in general, take translation and democratize translation and make it AI. Translation go hand in hand with translation goes hand in hand with risk accessibility. So with advancements in large language models, first of all, you don't need to pivot through a particular language. You, you can go between two languages. Even if you don't have sufficient amount of training data for a particular language, you can still compensate. There is a huge field of generating synthetic data. Like for instance, there is a known language and well resourced language in a language group. You can actually draw parallels between an existing uncovered language and language that's less covered and you can actually develop a core corpus, a synthetic corpus for long tail languages.

It's really fascinating.

It's fascinating. Now, again, training animals and understanding animal language, I might not quite be the expert in the field, although there are a lot of. I know that there is a lot of work being done in that direction. I'm not quite there. I still focus on human translation.

All right, all right, we'll go, we'll go two more here, quick. So Frank, asking how much of this job replacement talking about versus acquiring new skills in retooling the work that will be done. Also, will I help teach others how to do translation?

Absolutely. So I actually have a lot of conversations with professors from the Middlebury Institute, if I'm pronouncing it right, what used to be Monterey Institute. So the reskilling, it's not as much job replacement. I would probably go with reskilling. The reskilling can happen, learn at work, right. Or there are a lot of classes, for instance, offered in translation schools, localization schools that actually teach you that I don't know, Python basics, prompt engineering, syntax. So there is quite a bit that's already being done for reskilling. Again, I'll be very careful about replacement. I'll really talk about reskilling. And it's very much true. If you look, for instance, what's happening in language learning application, there is way more AI plugged into how you learn the language. It's much more personalized. Like again, I spend a lot of time in Mexico, so I spend a lot of time in duolingo and I'm watching it evolve more and more and plugging more like, okay, Olga, you failed here, let's go back, let's revisit that. So AI powered education applications are fantastic. My own personal passion is how can I help kids with learning disabilities actually acquire professional skills? Translation or not. Right. It can suggest options, it can act as a copilot. So that's one thing that I'm personally very passionate about. But absolutely, it's you.

Yeah, love that, love that. All right, here we'll do our last audience question. So splat, Lana, asking as a consumer, how do you validate the output assuming you don't know the language to which most the model translates to? That's a fantastic question. How can you do that as a consumer?

I know how you can do it as a professional, but let me think how you would do it as a consumer. So let's take professional first. As a professional, like as a language. Professional AI is equally used for producing language, but also for vetting language, vetting the output for accuracy, for like how good or bad your translation or your generated content is. And I would imagine that, and I know that some of the GPTs in the GPT market do exactly this and I'm pretty sure there are widgets that can do exactly this. Large language models can beautifully produce content, but they also can self heal and self judge. For instance, you take GPT 3.5, you validate it with 4. So I would say I know that it's happening in professional translation and I'm pretty confident there is something, something out there that can help you, help you vet it. Let AI judge itself and make decisions on how good a bad it is.

That's great, that's great advice. So, all right, Olga, we've covered a lot here. We've talked about different ways that you know, this is used in our daily lives, how important, you know, translation is to businesses, you know, trying to expand into new markets and also some of the pros and the cons. But you know, as, as we wrap up here, what is maybe your, your one most important piece of advice specifically when it comes to AI impacting, you know, translation and just the jobs and careers of people in there. What's your, your biggest piece or your best takeaway advice for, for everyone?

Somehow manage to wake up in the morning and listen to podcasts and glance through AI news and know what's the latest, making sure that if you are in school, you're learning what's relevant. So that would be the first advice? Absolutely. Stay appraised of what's happening in the world of AI and another important skill that's extremely demanded now. Use your best judgment to deploy AI where it's fit for purpose and don't just deploy it across the board. And actually, AI analysts and advisors that are appraised of the capabilities and shortcomings are in huge demand now. If you can pass that judgment, you will always have the most fantastic job in the world of multilingual AI.

I love that. So, so important, Such great advice. Olga, thank you so much for joining the Everyday AI show and sharing your insights with us. We appreciate, appreciate it.

Thank you so much for having me.

And hey, as a reminder, y', all, there was a lot there, so much good information. You know what I'm gonna do as a human once we get off this call? I'm gonna go listen to this again and write more, more information. So, so all of this great knowledge that Olga just shared with us, you can read about it, see how you can apply it to grow your career, to grow your company. So if you haven't already, go to your Everyday AI dot com, sign up for that free daily newsletter written by humans. And thank you for joining us. And we'll see you back tomorrow and every day for more Everyday AI. Thanks, y'.

All. Thank you.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.