S4E3 AI Translation - Dos and Don'ts with Markus Seebauer

Zohra Mubeena-Mutabanna: 0:08

aloneness notes. Welcome to Inside tech COMM With your host Zohra Mutabanna. In season four, I hope to bring to you different perspectives and interests that intersect with our field. Let's get started. Hello, folks, welcome to another episode of Inside Tech Comm With Zohra, Mutabanna. Today, I have an amazing guest, who I have chatted with a few times, and have gotten to know over the last few years, his name is Marcus C. Bauer. Hey, Marcus, it's awesome to have you on my show today.

Marcus Seebauer: 0:47

Thanks for having me, Zohra It's a pleasure.

Zohra Mubeena-Mutabanna: 0:49

Yes, it's my pleasure. So without further ado, please, can you share a little about yourself?

Marcus Seebauer: 0:55

Sure. Yes. I'm from Germany. And originally, I studied computer science in Munich. So I have quite a tech background. And then a bit randomly, I got into the translation and localization industry you over a decade ago.

Zohra Mubeena-Mutabanna: 1:15

And you have been doing based on your LinkedIn profile, you are focused on elearning localization,

Marcus Seebauer: 1:24

We are focused on elearning, documentation, and technical documentation to help companies grow internationally.

Zohra Mubeena-Mutabanna: 1:33

So tell me a little bit about your company.

Marcus Seebauer: 1:36

So I started Gateway Translations around 11 years ago. And over time, we grew a lot. We work in 45 languages. We have clients in the US and Canada, in many European countries. We worked with GitHub, General Electric, some start ups, byte range of clients, on the technical documentation elearning, and website translations.

Zohra Mubeena-Mutabanna: 2:07

45 languages. That's amazing. Now, I have never encountered or rather her in my career, I have not had to translate work with translation companies, unfortunately. So tell me convert translate into 45 languages? Is the process similar for each language? Or is it different? Like what does that process look like at your company, specifically,

Marcus Seebauer: 2:30

one of the most important things is to start with what you want to achieve the many ways how you can do translation, human translation, you can have the four-eyes principle, you can have AI translation, AI assisted translation, you need to properly prepare files, or you will multiply the mess you already had. Imagine, you translate your documentation, sometimes we translate into 20 languages for a client and a file was not properly formatted. And it's fine for English. But when you translate it in, let's say, in Spanish or German, the tax expense a lot, then it might look very different. And then when you have to fix that later, in 20 languages, it's quite a lot of work and not an enjoyable work. It's a cost, it can delay things. So it's important to properly prepare files, and choose the right translation company or even can be internal translators, or freelance translators. There are different setups, how to approach translation.

Zohra Mubeena-Mutabanna: 3:47

So one of the things that you touched upon was cost. And I want to come back to that. But before that, I'm just very curious to know. Are there any exotic languages that you have translated to?

Marcus Seebauer: 3:57

Yes, for example, maybe for you is not exotic, like Marathi. But many people might not even know that language in the US.

Zohra Mubeena-Mutabanna: 4:10

Oh, wow. Yeah. I never thought I generally I would think of Hindi. But for my listeners, Marathi is a language spoken in the Western state of Maharashtra. Right? And the script may be very similar to Hindi. But yes, it's its own unique language. And very rich and I'm so thrilled to hear that is what characterize it as an exotic language. Yes, it's not exotic to me. Yeah. Is the process very similar for other like, you know, for an exotic language in this context, would it be very different from other European languages?

Marcus Seebauer: 4:46

It can be challenging when it comes to encoding, let's say software localization. If a software is not prepared for some scripts that could cause Some issues or some funds, they don't do well with certain languages. Now we have declined in Arabic, which is a right to left languages. And that changes many things. The bold print was not visible. And one of our project managers caught that and then discussed with the client, how can we solve this? Because it changes the user experience if the blueprint basically disappears?

Zohra Mubeena-Mutabanna: 5:29

Oh, yeah, I never thought about the user experience aspect of it. That's a very important aspect that one needs to think about, right?

Marcus Seebauer: 5:37

Yes. And there are also other right to left languages, which are less well known, but especially with Arabic, with Hebrew. So for the Arabic documentation, when we have a figure, let's say, with two x's, then we also need to invert the figure. Yeah, you can imagine if you had like a figure that was the other way around, right? In English, that would be super weird and confusing, and not professional and hard to understand. So for Arabic people, it might be similar.

Zohra Mubeena-Mutabanna: 6:20

Yeah, a lot, a lot of things to think about here, challenges. I guess curveballs that one should keep in mind. So it looks like there's a lot of good pointers that you're giving our listeners if they are ever considering having their content translated into because I thought it is it's a pretty straight line. But it's definitely not it doesn't seem like a straight line to me. From point A to point B, there could be a lot of zigzags.

Marcus Seebauer: 6:43

Yes, it's the thing you don't know what you don't know. So clients make assumptions based on some things they don't know. And that can cause issues later. And then many translation companies, they just see themselves as a company that okay, please send me the translations, and then I will send you back the translations properly translated, but that's not enough.

Zohra Mubeena-Mutabanna: 7:13

So there is a lot of hand holding involved here. Yes, that's adds a lot of value, it adds. Okay, so now I think we've talked about cost and value. So please tell me, what is the value versus cost for translation? You know, for companies that may be holding back from translating, or thinking of shortcuts? What are some things that companies should keep in mind? I think you have touched upon some good nuggets. But what else can you think of and share with us?

Marcus Seebauer: 7:42

Yes. So for example, as a German user, I have used different company services, their documentation, and then it's interesting how I approach it, and how they didn't really consider that. So let's say, when I look for specific information, I would often search on Google, but some other people, they might go on the company website. So they search in German on the company website, I search in German on Google. And then you wouldn't find anything if the documentation doesn't exist in German.

Zohra Mubeena-Mutabanna: 8:22

Okay. So can you elaborate a little bit on that? I think I'm kind of trying to understand in terms of value versus cost, can you add some more meat to it.

Marcus Seebauer: 8:31

So especially for products used by developers, software products, they like to research features to look into details before they buy the product? Or maybe it's a it's a freemium product, maybe that's less common in with enterprise software. But people do a lot of research about very specific technical aspects of a product in when someone searches in German, and then nothing is coming up might be even a competitor's product coming up. At least, that's the case for publicly facing documentation.

Zohra Mubeena-Mutabanna: 9:14

So is this where Search Engine Optimization becomes critical?

Marcus Seebauer: 9:18

Yes, because in different languages, it's maybe not an exact translation. But he see what you mean, people use use different keywords,

Zohra Mubeena-Mutabanna: 9:29

Keywords, I see. So what would you recommend to your client, whose content may not appear in German but maybe appearing in English in Google? What should they keep in mind?

Marcus Seebauer: 9:41

It's good to see how is the English documentation doing which content performs well? And also, what do you want people to do? Do you assume that if people spend a lot of time on a page, then that's good because they have more understanding what does it mean? That it's a little bit hard to use that feature, the product? You need to holistic perspective.

Zohra Mubeena-Mutabanna: 10:11

Makes sense? Makes sense. Yeah. Thank you for shedding more light on it. Another thing that I came across in your profile was you talk about machine translation versus AI assisted artificial intelligence assisted translation versus human translation. Now human translation is evident, it is a human translating it for you without any, I suppose. Right? Would that be? Right? But I thought that machine translation was AI assisted translation. But seems like there is a distinction. Can you please tell us what that distinction would be?

Marcus Seebauer: 10:43

Sure. So what you see these days, some companies, they assume, let's just use AI to translate our documentation for a smaller market and see how much traffic we get if it's worth it to then kind of correct this a bit or even use human translation, just as a first step. That's something in the field of technical documentation, or, let's say, on social media, on Twitter, people click on a button, and then the tweet is translated to another language. There's no human involved there. And sometimes you can see that from the quality.

Zohra Mubeena-Mutabanna: 11:31

Right? So is that machine translation? Is that AI?,

Marcus Seebauer: 11:35

I mean, AI translation is more the term that people outside the industry use in, in the translation industry, people usually speak of machine translation, and then AI assisted translation where you have a machine translation, AI translation, and then a professional translator, will correct that. And it's a kind of particular skill set.

Zohra Mubeena-Mutabanna: 12:03

Oh, is it? So somebody who's been doing human translation would not be a good fit for content that is machine translated?

Marcus Seebauer: 12:11

Yes, like, could be similar to a technical writer and a technical editor. They are different skills. So it's okay, creating something from scratch, especially translators who have been doing this for some decades. They might not like to work with AI translation.

Zohra Mubeena-Mutabanna: 12:34

I see. Okay, lots, lots to think about there. Lots to think about. Now, ChatGPT is here. And do I didn't intend this to be my team for Season 4, I think it's very relevant that we talk about its impact from your perspective. I mean, AI has been used in machine translation for quite some time now. Right? I think in my season three, I spoke with Lisa Zarfl. And she talks about machine translation and how it has evolved, but it hasn't eliminated, you know, human editors, you still have that human component very strongly present. What is your take on that question? Before I go on to my ChatGPT question? And specifically,

Marcus Seebauer: 13:19

yes, there are some market segments where people just use something like Google Translate and publish it. But when it comes to professional documentation for a software company for an engineering company, or other fields, then there are always humans involved.

Zohra Mubeena-Mutabanna: 13:40

Okay, now with ChatGPT entering the arena, what do you think the future will be like?

Marcus Seebauer: 13:46

In the translation industry, there are lots of manual tasks, and they take a lot of time for the people on the client side, people on the translation agency. And I think there will be more and more use cases to reduce the effort for these tasks, like let's say, terminology extraction, to find the key terminology in an English text, extract that so that you can build a term base glossary.

Zohra Mubeena-Mutabanna: 14:19

I see. So that's, that's an example where you think there's that manual element involved, which ChatGPT it would not be able to replace

Marcus Seebauer: 14:26

Potentially there are many use cases that we cannot even imagine now. I think the biggest aspect of ChatGPT is that AI translation really went into the mainstream.

Zohra Mubeena-Mutabanna: 14:41

Oh, that's all you think. I see. So you don't feel like ChatGPT is a threat to human translators, like translation editors, it's not going to replace them or eliminate those professions.

Marcus Seebauer: 14:55

In some ways. It is, for example, a fortune In 500, Klein, they decided now, because they sold less in the last quarter in that they will not have translators work on their documentation anymore, and they will just use AI. And then they will have internal people correct the mistakes. So basically, by getting into the mainstream, like the visibility for ChatGPT, that reduces work for translators, because clients feel that they can quickly edit this internally. At the same time, there's a lot more work because the content has been increasing. There's a lot more content in the internet, generally speaking, we translate into more languages because the world is more connected. So over all the translation volume has been increasing, especially when it comes to the AI assisted translation.

Zohra Mubeena-Mutabanna: 16:04

Okay, so from what I understand, it seems like companies are already leveraging ChatGPT. And eliminating or sort of reducing that manual involvement. Is that right? Okay. I see. So there is a worrisome trend. In that sense.

Marcus Seebauer: 16:23

Yes, in some ways, the quality of documentation will go down a lot.

Zohra Mubeena-Mutabanna: 16:30

I'm so glad you're talking about the quality of the content going down. So you you see that that as a real sort of taking away from the value of the content and that sense,

Marcus Seebauer: 16:40

Yes, and there are a lot of hidden costs, because I don't want to mention this company publicly, of course. But when we look at the engineers, who have a very different role than editing translations, and they validated translations before, but now, they should really clean up the whole translation without any terminology management, and they don't have this knowledge how to set up a process. That's quite a recipe for disaster in, I wonder how many companies are in a similar boat, there are great applications and ways to set up AI and translate a lot more content faster at lower cost, but you need knowledge or you need a tech partner for that. And you cannot just do that from today to tomorrow. It's it's quite an effort.

Zohra Mubeena-Mutabanna: 17:42

Okay. So that is where the value of what value versus cost. That's another aspect where companies are compromising on value to save some cost.

Marcus Seebauer: 17:51

Yes. And now there was an issue with Samsung, where employees used a ChatGPT. And you should never enter confidential data also into Google translate, it's quite risk. And some companies, they even block such websites for the whole company.

Zohra Mubeena-Mutabanna: 18:11

Yep. Yep. I read about where user history user chat history was revealed accidentally. And, you know, companies that are experimenting with ChatGPT ended up exposing their intellectual property. Right. source code, but so those are those are valid concerns. Absolutely. Another thing that comes to mind is we talk about bias and data. And with translation, there is a higher chance for any bias or inappropriate content to kind of get through, right, without the right filters. And would you agree?

Marcus Seebauer: 18:49

Yes, I saw that on some big social media sites where some very normal innocent posts, they would be very insulting in some profanity, in sometimes racism because of errors in the translation.

Zohra Mubeena-Mutabanna: 19:11

Marcus, as we've talked about, you know, all these challenges, what advice would you have for your clients or companies out there on when to use AI assisted translation versus human translation?

Marcus Seebauer: 19:23

It's important to really look at the accuracy you want to achieve with the translations. If it's something very critical, then it makes sense to use human translation. And we should also consider in the past, we often had a translator, and then we had a second person, proofreader editor, who would fully check the translation against the let's say, English source, and edit it. And now if we move from this four eye principle to a is Since the translation, we just have one person there. And then the role of the project manager is really important. So for some applications like medical devices, life sciences, where it's very critical to have extremely high accuracy, AI translation might not be a good starting point.

Zohra Mubeena-Mutabanna: 20:22

Oh, okay. That sounds like one of the limitations of AI assisted translation. Are there any other limitations that you can talk about?

Marcus Seebauer: 20:32

Yes. So when you have very specific terminology, then it's challenging because you have to change a lot. And there's a certain type of blindness, when you as a, we call it post editor. So basically, linguist, the translator, who edits the AI translations, when you have a huge project, and you don't create it from scratch, you just try to find the arrows. Another limit is really the existing translation. So if you have a huge translation, memory of existing translations, translation memories, bilingual format, let's say, of the English source in the Chinese translation, and then you can also train AI. And that is really powerful.

Zohra Mubeena-Mutabanna: 21:31

So this is the positive, the pros for using AI, it's not a limitation.

Marcus Seebauer: 21:36

Yeah, if if you have that existing translation memory, then it's great. But if you have the very specific terminology, use the translation from scratch, or you get into many new languages, then it's really a limitation. And we call it low resource language, like a language where there's not so much training data online, let's say something like Marathi or maybe like Vietnamese, there can be many languages where the quality of the AI will be much lower. And also another limitation is that you cannot use the same machine translation engine and expect good quality for every language. For the set Chinese documentation translation, you will get much better results with some Chinese engines, but then you would also need to consider data protection.

Zohra Mubeena-Mutabanna: 22:37

Yep. Excellent, excellent pointers. In everything that you've shared with us. So far, I believe I've covered a good set of questions that come to my mind, is there anything that you would like to touch upon other than what we've talked about?

Marcus Seebauer: 22:51

Another interesting point, with AI translations, machine translations, let's call them self learning, AI, basically, it learns just like training data based on user input. So when the translator translates one sentence, for the next sentence, it will consider the previous translation. And that is something very unique and powerful. And we have some technology for that. So when there's specific terminology, or it can also consider a specific style, that goes a long way compared to very generic engines, where it doesn't matter. If you translate that one page, or 1000 pages, the AI will not learn.

Zohra Mubeena-Mutabanna: 23:41

I'm a little confused here. So I think what you just said about AI being able to self learn. So is machine translation, we touched upon that machine translation versus AI assisted translation. With machine translation, not have the ability to self learn versus AI assisted, like using ChatGPT would be an advantage. Is that where you're trying to draw the distinction?

Marcus Seebauer: 24:06

Yes there's another technology we use. There's a lot of technologies available for different use cases where you translate one sentence in one way, and one technical term in one way. And then when this technical term occurs, again, this will be considered. But if you use something like Google Translate, that wouldn't be the case.

Zohra Mubeena-Mutabanna: 24:33

I see. So I My apologies because I when I think of machine translation, AI assistant translation, I did not know that there could be nuances in the technology that you're using or the engine that you're using.

Marcus Seebauer: 24:45

Yes, it's huge. What's out there.

Zohra Mubeena-Mutabanna: 24:49

Yeah, apparently, of course, my short sightedness that I missed on that. This has been a very revealing discussion to me because I like I said, I know a bit about machine trans. I mean, the Tron Thracian industry, but this has given me a very different perspective and insight. Again, I believe we've covered a good set of questions and we've been, you know, something very new way different. Anything else that you would like to add?

Marcus Seebauer: 25:14

I think that's all from my side.

Zohra Mubeena-Mutabanna: 25:17

Yeah, this has been a great conversation. Marcus, thank you so much for coming on my show for your time. And for enlightening our listeners with, you know, with all the hidden gems.

Marcus Seebauer: 25:26

Thank you, Zohra. has been a pleasure.

Zohra Mubeena-Mutabanna: 25:28

It has been my pleasure. Thank you so much. subscribe to the podcast on your favorite app, such as Google, Apple or Spotify. For the latest on my show, follow me on LinkedIn or visit me at www.insidetechcomm.show. Catch you on another episode.