Building an Effective Localization Process with Microsoft's Soeren Eberhardt
In this episode, Redokun co-founder Stefano Bernardi chats with Soeren Eberhardt from Microsoft about his extensive career in the language industry. Soeren shares insights on tackling challenges in creating content for diverse markets, tools and strategies for managing large projects, lessons learned from past mistakes, and the role of AI in localization.
[00:00:00] Interview with Soeren Eberhardt from Microsoft
Stefano: First of all, thanks so much for being here and for allowing us to have a conversation with you.
Soeren: Oh, my pleasure.
Stefano: And my question, first question to you if you can, describe yourself as a person besides the expert in the language industry.
Soeren: What I'm always saying actually working for Microsoft is my second job. My first job is being a dad being a family man. I have two boys, 16 and 13, so teenagers need a little less attention than toddlers. They need different attention. And my, my proud career as a soccer dad just ended because my younger one decided he doesn't want to be a goalkeeper anymore.
Soeren: He wants to play American football instead of the sport world wide known as football. But I'm digressing. So that's definitely important part of my life. And yeah, I also have two cats and two dogs. And sometimes I feel like the cats need more attention than the dogs. Although dogs need to be walked.
Soeren: Yeah. And [00:01:00] I love reading. I love traveling. And I do travel with my boys, with each of my boy every other year. That's one of the things that I established when my older one was six years old. So over 10 years ago, we went on a road trip together. And then we realized it's tons of fun.
Soeren: So the last. Big ship that I took with my younger one was actually to, to Japan where I'd never been and was really cool. And it's fun traveling with kids because you see things. differently again. Sometimes, especially I've traveled before for both for pleasure and for my job. At some point, especially when you're traveling for your, for work sometimes don't see this things as, as excitedly anymore.
Soeren: But yeah, traveling with a, at that time he was 12, traveling with a 12 year old through, through Japan was just absolutely awesome. Now I would have probably not looked as much for Manga and anime as I do with a [00:02:00] kid at home. So yeah, just a few words about me.
Stefano: If you come to Asia, hit me up. I have a lot of recommendations.
Stefano: I also noticed that you wrote a book the Little Book of Languages, or maybe several, that's what I found. It's super interesting. I already have it in my cart on Amazon to order. So I noticed that you share a lot of stuff on LinkedIn. You, of course, have created a lot of content related to languages.
Stefano: Where is the best place that people can follow you?
Soeren: LinkedIn is, that's where I'm posting things. Yeah. And I love writing about language. I realized there's actually a lot in the news and I think for people in the language industry there's so many different topics. It's such a, such an interesting field where there's technology and.
Soeren: We can talk about changes in technology later. There's so much, so many things that are happening just in the tools area. And then [00:03:00] there's language itself. I'm always saying we actually working on one of the best problems to have we have different languages and different cultures, and I don't really consider that a problem.
Soeren: I think that's really super beautiful. and just learning about those differences is awesome. But then we also have to work on bridging those. That's our challenge, right? That's the problem itself. And yeah, you cannot reach a gap that you don't understand. And so understanding just the different cultures and languages themselves, I think is super helpful.
Stefano: Yeah, I can see just with English, when you travel around and you go to Singapore, you immediately see that people there speak a different English compared to the UK, you can feel if you live there a little bit. What's the fundamental in the society that are making English that different compared to what we were speaking in Europe, let's say, so that's really interesting.
Stefano: So I've noticed that you've mentioned in some other [00:04:00] interviews that you guys. Your team is focusing on producing content for 39 languages. Of course, the markets are many more than 39. I've heard that you are more focused on marketing, if you could share a little bit more about your work in Microsoft and what you like about it, maybe also.
Soeren: Yeah in the meantime, I've moved to a role where I'm working with a partner. So I'm not working on that localization project anymore, but I can outline what. Yes we worked on marketing pages for selling Microsoft 365 and doing that as you pointed out, 39 languages across many different markets.
Soeren: And actually the fun of that was that you have to be in when you sell products the nice thing is you can really see your impact. I could see how much we sold in each and every market. We could also web pages. It's so beautiful. You can see the engagement. You can see where people click, what people [00:05:00] like.
Soeren: So you have to be hyper focused on, on optimizing things per market. You should, if you want to do better. And I led the experimentation program for those, for all our markets, meaning that I was running a lot of tests people that aren't necessarily familiar with that concept of an AB test.
Soeren: So yeah. You basically don't change anything and you have a variant or maybe even more than one variant where you make a change that you get that track against the hypothesis that the behavior of your users will be changed. And so we would try things where we just change the messaging of the headline on a page and say, okay, for emerging markets like Mexico and Brazil, maybe people are more interested in this particular
Soeren: feature of the whole suite that we're offering and we don't use the English headline translated but we create a new one for those markets and see Will that actually change how many people buy? And it did. Which [00:06:00] that was a successful test. And while I should not say that the other tests were not successful you're always successful in learning something.
Soeren: But we also, of course, had tests where things did not play out as we expected. And we were like, okay, we tried to change behavior. No didn't do it. But yeah, that was super, super interesting. Just seeing how you can make specific changes. Like for some European markets, we, for example, we're pointing out all the products help with GDPR, obviously for the US we never had to think about that the small and medium businesses that we sold to did not think need to think about well, they need to think about data privacy, but not in terms of legal regulations in Europe.
Soeren: We would try out all these different modification on the pages. So we had different content for different markets. And we actually had people that were located in the market. So I always had an inroad to our subsidiaries to get additional input. It was really super, super helpful. So only looking at getting [00:07:00] things translated getting things localized, but going that extra mile and adding content, maybe shortening content in other areas. Yeah, adding different presentation of the content to make sure that it's really fit for the individual market. And I know that it's a luxury that not all companies can afford, right?
Soeren: It's you really need to put effort in things like that. But I think that there's ample opportunity even on a smaller scale than what we did.
Stefano: In fact, I wanted to step back a little bit to, to make sure that we can provide a little bit. Maybe just to those people that maybe are not working with this advanced technologies, let's say.
Stefano: So my first question was probably what are the key strategies for managing, a large amount of translations, right? Can this amount of people that are working on this project that is, of course, very big.
Soeren: I think, yeah, there's so many different components to it. How do you make sure that first, the tools obviously are important, and [00:08:00] I do think that the language industry by now, there's really good and solid tools.
Soeren: You should not have to think too much about your translation pipeline. I was I know that I've been spoiled at Microsoft because a lot of that was really basically fully automated. I didn't have to think about handing files to somebody, for example, things got pushed out automatically. That's not something that, that everybody has.
Soeren: But. The more you can make sure that the tool space works well, that you don't have to think too much about that. Also that your translators, I don't, are not getting too distracted by tools that they don't know how to handle. And that's the second piece is obviously the whole education of your ecosystem, right?
Soeren: So Microsoft has excuse me, for example we have a site with all our style guides. We have a big terminology database. So really giving translators the right guidance on how to [00:09:00] translate. And it's not only about the regular language mechanics, right? You want to tell people this is our brand voice.
Soeren: This is how we want to come across and really making sure that you get, even before you start a project that people really have the necessary information, also the information about whatever you. Localize whatever topic it is, whatever subject it is. When you look at for example a certain product, is there additional material that you can give to people?
Soeren: And I know that for translators that is overhead but ultimately it makes their lives easier, right? They need to invest. upfront work of consuming content. So you also need to make it concise. And I did work on language, on the language quality strategy a while back when I was on a different team at Microsoft and we would have kickoff calls, for example, where we invited key people from the translation teams and told them about the product that we were going to localize with them.
Soeren: At the time I worked on Microsoft T Teams the very [00:10:00] initial. Localization of that, we started with 18 languages and. I wanted to have a chance to explain to people what is this thing? What can it do? And these are some of the key features. Of course, everybody needs to sign a nondisclosure agreement for things like that.
Soeren: Tell them about a new product, right? But that's standard in the industry. Giving people some information to, to work with that they have in the back of their heads so that they really understand, okay, this is beyond just writing grammatically correct and without any typos.
Soeren: I also need to think a little bit about this is how things should come across. And I also know, okay, yeah, this, these are the features I understand what this whole thing is about. There's less chance of mistranslations.
Stefano: Yeah, I can imagine how challenging it was to, translate Microsoft Teams and everything around it because the growth that you guys have with that tool was impressive.
Stefano: So I feel like the speed that the team had to work on [00:11:00] everything around it must have been really challenging. So could you describe what was. Your typical workflow, typical process in this type of very complex projects, if there is a typical workflow. 'cause
Soeren: Yeah. It's it's hard because it's very different when you localize like a contained project or when you have ongoing localization, right?
Soeren: For that initial release of Teams. It was a lot more traditional with you basically take files that you sent it sent to people and you can give them all this initial information in a kickoff call, like we're getting started now. And you have a more thorough review phase.
Soeren: Yeah, you were asking about the project flow, so you should have. A quality control step somewhere in between where even if you cannot review everything, you want to sample, sample things and understand how's the [00:12:00] quality. We also try to give it to native speakers to review things.
Soeren: So basically after translation, you also have a base. And then we also obviously we did a lot of localization testing. So you need to have these very distinct project phases. Giving files to the translators, getting them back from the translators, having the localized versions those being reviewed, but then also being tested.
Soeren: Going to all the bug fixing and catching up with any updates that might have happened on the source side. For ongoing localization, things look very different, right? If you keep working on a project on a product becomes less of a project. You have these mini phases, but where do you fit in quality assurance now?
Soeren: So you can do that on an ongoing basis, but you know that. Whatever you find, it might already be in the product. You might need to involve your users a little bit more and see whether you can get feedback from them. But you also need to be very agile in your [00:13:00] bug fixing. So everything happens more in parallel.
Soeren: So as a project manager, that can be really difficult to wrap your head around that. Oh, you don't have these distinct phases anymore. Now you put them on top of each other. And when do you do what, right? So you might have a week where you focus on the language quality work and you do it more in parallel and you still have to get files or strings out to your translators, but you say okay they're working on that, but like we have the review phase in one week and in the next week we fix all the issues that we found you on that review.
Soeren: Although we know that they are already in the product and we just keep going. With that in, in parallel so you need to, yeah, you just basically need to change your thinking for that ongoing localization a little.
Stefano: But you touched on proofreading, so we talked at the beginning about what trail difference is we're translating something.
Stefano: I saw today that you posted something on LinkedIn, or it was yesterday, [00:14:00] today for me, it was yesterday, probably for you, but I saw that you. You posted or you shared an article about false friends, let's say. So words that in English mean something, but in other cultures, maybe they're not very nice words, let's say.
Stefano: So how you, who does that control for you? Who do you think is the best fit to proofread your content and to make sure that these cultural issues. don't pop up, in your
Stefano: translations.
Soeren: You can have lots of different layers of how you control, right? You can have a peer review process where you have one translator translate and another one reviews or proofreads right away.
Soeren: Actually it's. how I started my career in localization long time ago. And my role was called copy editor. I would just review the translations [00:15:00] of other people. At some point it got so boring that I decided well, not so boring. I was more frustrated. And I was like, I'll translate myself.
Soeren: I do, I can do this better than those people. Which was not true because then of course those proofreaders found something. You can also have or on top of this, you can also have a neutral reviewer. And At Microsoft, we've had lots of different models we would have that at the vendor internally, they would just have a different team, right?
Soeren: And we would say, hey we leave that to you, but you need to have people at some point who review and as I said if you have a lot of content, you might just need to take a sample, right? And then you'll and there are all these methodologies of error categories, there's this whole quality framework, the TOWS S1, the TOWS DQ framework.
Soeren: We say, these are the different categories of mistakes that people make. And then you also have severity. How bad is it? A [00:16:00] long time ago, somebody actually translated Windows consistently into the Portuguese word for windows for Brazilian Portuguese. And so that obviously is a lot worse, right?
Soeren: When you just throw out the brand. And yeah, these different categories and I always felt it was helpful if you also have built in an arbitration process so that the translators learn what they did wrong, because you don't want just to fix the issues, you also want to give that feedback.
Soeren: And then you also obviously have a way to measure if you normalize it against a thousand words and you can calculate a score and that's something that we did very regularly. So those are two layers. And then obviously there's your users as well. And so you can we shouldn't go to the users right away because sometimes you have like internal reviewers that are not necessarily involved with the product and in Microsoft.
Soeren: We have native speaker communities, for example. Sometimes I get asked if, my German is still good enough. I always saying, I don't want to translate anymore. I'm losing touch with the market, but I can still check [00:17:00] a product and see whether things are mostly correct. So I can look at translations and I can also use the software, a new product in my language version.
Soeren: So if you have that. I know that sometimes companies actually use that as their very first review process that they just ask internal speakers, you need to be a little bit careful with that because at some point they tire out and they're like, Hey, I have another full time job. I cannot constantly volunteer for your localization team.
Soeren: But it's definitely something that you can leverage. And of course, the bigger the company, the easier it is also to spread out that work. And we did that. In the marketing team that I was working in we would have contacts in the subsidiaries and for pages that we stood up freshly of when we said, Hey, we want to advertise this new product.
Soeren: We have a completely brand new page we would go to the subsidiaries and say Hey, can you have some volunteers just looking at this, and they would review it and give us feedback [00:18:00] so that is a good process as well. If you. If you have people that you can leverage for that. And then the next step is your users are willing to give feedback.
Soeren: Often, it's just the lack of Ways to give feedback, right? If you find a, an error on a Microsoft web page, who do you go to? Oh, it's not that you just enter like localization [email protected] and send them an email. But fortunately we have feedback mechanisms. We actually have surveys.
Soeren: But then more and more web pages have a thumbs up, thumbs down or some sort of mechanism where you can tell people hey, there's maybe there's something wrong, but that is one of the challenges often you get feedback where people will complain about product features, something that's unrelated to the translation, you get a lot of noise and we've seen that with our office products that in product.
Soeren: We would pop up a survey and a [00:19:00] lot of the complaints are not about localization at all, but other things. And it's interesting as well, but having to filter that out and obviously you get a lot of feedback in other languages than English. Yeah, you might need to run things through machine translation to get at least a feeling for what people are complaining about and then process that, and that in itself is can be quite a bit of work.
Soeren: You need to see what can you do. It's something that's in between, like your internal reviewers and every user. If you can afford it to have beta testers for a product and I hadn't talked about like specific localization testing. If you have native speakers and you give them the full experience sometimes that can be paid.
Soeren: But you can, obviously there's products, especially games, for example, if you give people an early release, they will be happy and they are willing to give you back in terms of translation feedback, for example. So that's always a means as well. So you have all these different [00:20:00] basically circles of possible reviewers, testers that look at your localized versions.
Stefano: So I can really tell your passion for your job, sorry, and I got actually understood what I need to do if I find a type of Microsoft website, they send you a PM on LinkedIn and for sure it's going to get fixed.
Soeren: Yes, people have done that on LinkedIn. Yes,
Stefano: You shouldn't have confirmed my suggestion.
Stefano: No, you're going to be flooded. Moving along. So what would be your advice for a company or a team that is starting to dip their feet into this world of translating their content. How should they do it from scratch? What's the first step?
Soeren: So for a team that just starts translating content I would actually say start at the source.
Soeren: How much do the people that create the content for you, how much do they understand international markets? Who are they writing for? And since we were [00:21:00] talking about Microsoft Teams and that very first release that I worked on, I spent a lot of time with the developers, spent a lot of time with designers, just telling them, Hey, you have some beautiful ideas for stickers and you put an emoji, but let's talk about things that could be offensive across markets that could not work.
Soeren: And sometimes you feel like the fun killer, where you're like, Hey, you need to think about this, you need to think about that when you go into international markets. And the same is true for just for text, right? Whatever you translate how suitable is it for translation? And I know that is often not the thing that actually the people who give you all that content expect from you.
Soeren: They're like, okay, now go away and get this back to you, to me in all your different language versions. Then we can distribute it and instead you are saying like, wait a minute, can we first take a look at this together? But writing teams don't necessarily know how to write for the audience and even for English.
Soeren: Because when we're frank, there's [00:22:00] often if you don't cover lots of languages, there's always those people who use English as a second language, maybe even as a third language. Those writers might not have written for, they might have written really specifically for their U. S. audience or for their U. K. audience. So having that conversation, I feel that it's something that translation teams, localization teams should always do as basically just giving a few simple guidelines. Just think about humor, for example. And as I said you don't want to be the fun killer if they have a sort of humor, but then let them explain to you what are they going for with those like specifically funny phrases, idioms.
Soeren: Because then you need to go back and think about, okay, how will we tackle that in other languages? Do we want to keep that or do we tone it down? And then if we want to keep it, and that's the next step for translation team by thinking about what is that additional guidance that we can give to [00:23:00] our translators.
Soeren: Do we have terminology and you can always start very small with terminology can have simple glossary you don't need to have this massive database solution, but the good thing is there's so many tools out there in the language industry, actually when I started, Trados had just come out with their term base.
Soeren: So it was still fairly new. Oh, that's how old I am. And by, by, by now things are established, right? And you can quickly stand up a term base. So making sure that and that's for the writing team as well, that they, you might need to have their help and for them it's beneficial as well if they are aware of terminology and it's not always the case across teams.
Soeren: It really depends on how much experience time they might have. And then the other guidance that I mentioned is definitely style guides. If you can provide people with additional information, as I mentioned, like if there's something humorous, then how should people tackle that in French or in Japanese [00:24:00] and some languages would really say okay maybe just tone it down, make it sound more professional, although the original British version is really funny.
Stefano: You touched on, you've touched on an amazing point because. I have an Italian speaker as a first language, we like to build paragraphs that are super long with a lot of sentences that are all connected and there is a thousand concepts in one single sentence because it's a full sentence just with commas, no full stop, no anything like that.
Stefano: And I cannot even imagine when an English speaker, for example, has to translate the content. It's going to be crazy. Like a lot of the concepts that are. Inside that sentence are for sure going to get lost because they or the translation is going to be very long. So the, I really appreciate your suggestions to start first for the, from the content that we've [00:25:00] written.
Stefano: And try to maybe tone it down, maybe review it thinking, okay, guys, we need to go to market with this in different languages. Let's align what we're trying to say here to make sure that we can actually communicate what we want to say. So that's, that I think it's the most, one of the most powerful suggestions that I've heard from people in the industry, because it's very simple.
Stefano: You can do it immediately, but everybody forgets about it. So I have a few more questions. Maybe I'm gonna go to them. The one that I find more interesting. Have you made any mistakes or do you know someone that made some of your teams maybe that made without making names that maybe made a big mistake and this mistake told you guys something and you learn something from that and you'll make sure you're never going to make that same mistake
Soeren: again.
Soeren: Oh yeah, I mean there's so many mistakes. And I shouldn't tarnish my brand's reputation. [00:26:00] We fixed them all but I think if you're working in the language industry you will inevitably make mistakes just trying to think about the mistakes that I can blame on myself because I know that Microsoft, for example in, many years ago, we made mistakes where we would show the wrong borders between countries.
Soeren: And then from then on people were aware of the fact that you have to be very careful when you show maps. And now there's established processes and they're all around the industry where Google would show, if you go to Google Maps and you look at contested areas it looks different.
Soeren: Kashmir, for example, contested area between India and Pakistan will look differently when you are in Pakistan or when you're in India or when you're somewhere in a neutral, so to speak, neutral country. And so Microsoft has actually addressed that organizationally just seeing that certain [00:27:00] things you really need to have a team that, that store some knowledge and as a way of things.
Soeren: On a personal level yeah, I worked on a program, for example, you mentioned my Little Book of Languages that actually evolved from there. I released a lot of languages at Microsoft that never done. So we started with Catalan at the time for there was in the early 2000s, Windows was available in a lot I don't know how many languages, maybe 20 languages.
Soeren: And then Microsoft started to have this, it was basically like a skin on top of Windows, like a language version that only had 20% localized, but the most important things. So it was really like 80 20 rule where you cover a lot of things that people use, like the settings, the important settings, but not something that mostly administrators on a computer would use.
Soeren: And we worked with all these languages that we had never done before. And that's where it was ample of opportunity for me to make languages just like the order in which you read these languages in a country [00:28:00] because some languages might have a different status than others. Yeah, and that's where I eventually, I just wanted to share the information of all these languages.
Soeren: There were lots of languages that I had no idea about and was like, oh, maybe that's some interesting information to share with people. And I always send out a little bit of information on those languages. But eventually triggered my idea. Hey, I should just write something about languages at large.
Soeren: But, yeah, those were the I also learned what mistakes people can make when they don't know how to use tools. And I've made, yeah, I remember one of my, that was many years ago, even before I started working Microsoft. I remember that my very first HTML file that I, I localized at that time, I was using a tool.
Soeren: That did not separate the tags from the content. And I just removed all the tags. I was like, I don't know what that stuff is. And the engineer next to me was like, what are you doing there? And I was like, there are all these things in this. And [00:29:00] yeah. He said, no, those are really needed. I feel like, I think Yeah, it's probably fun to to think about some of those mistakes.
Soeren: And as you said it's really important that you learn something from them. In that case, I just learned keep things in there. And also hope for better tools that eventually help you with that. Because that many years later, actually, I had something similar where in that small, in that language program that I was working on, we did Nepali and the Nepalese translators, they translated.
Soeren: We did have exposed some information on how big the font was supposed to be, which font type, and they translated everything. They translate the even the numbers into their local numerals. So that, of course. The page couldn't be rendered correctly because the browser didn't understand what to do with it.
Soeren: I was like, Oh, yeah, they basically did the same thing that I had done at that time. So then how do you address mistakes like that, right? [00:30:00] There's many ways. Sometimes you just take these learnings and you keep them with you and apply them in the right moment. But.
Soeren: Sometimes you have the chance to address it in a tool. Today it's really hard to actually remove all the tags if you're a localizer. Because your CAT tool will not allow you to do that. And who wants to translate in a notepad? Sometimes it's the process where you basically say Oh, people might make this mistake.
Soeren: So how do we make sure that in the process maybe we need that extra review step for this extra tricky thing. Yeah there's so many areas where you basically can think about small optimizations. I had this one, it was not my mistake, my, maybe it was my mistake now thinking about it, we had the word PIN just by itself and half of the people I think we translated at the time, we translated into a little over 20 languages, half of the translators translated it as personal identification number, because it was all capitalized the pin [00:31:00] that you enter somewhere, but the other half translated it as to pin something on something as that English verb.
Soeren: I couldn't blame them. It was really not clear from the context and we realized like, Oh, maybe we need more developer comments for words that are tricky. And at the time we actually developed an automated process that would comment for the developers. If they hadn't put a comment in, it would say please explain ambiguous words.
Soeren: And we had this word list. Whatever, or if we just, if somebody just had a single word in a string, which just automatically say, hey, can you please put a comment in on how that word is specifically used so that translators understand it. So yeah, sometimes you can, while it's. Between tool and process, right?
Soeren: You can instill something. We say, okay next time, what could we do to preempt a similar error? How can we make sure that this doesn't happen again? But yeah, there's so many mistakes that that I've learned from. As you say, like I'm experienced and while the experience comes [00:32:00] as much from the things that went well as from the things that didn't go well oh, can we do this differently?
Soeren: Next time.
Stefano: Yeah, there's always a next time, right? Hopefully. So last topic that I wanted to touch on is, you work at Microsoft, the big word right now, it's AI. So how do you feel about AI? Are you leveraging it? What's the impact that
Soeren: you see? Yeah, I want to be exact because what we're seeing right now or what we've seen for the last little bit over a year, basically with ChatGPT 3.
Soeren: 5 being released was the large language models because the language industry has worked with AI for a long time, right? Machine translation, the different iterations become more and more important. So the generative AI yes I'm using it a lot. That's the nice thing. I I do have a subscription for ChatGPT and I'm not advertising it, but I found it very useful.
Soeren: And then Microsoft has been working on integrating That's the large language models into [00:33:00] products, which I find very useful. So there are things where I feel what it's super helpful with is that fear of the blank page. In December I sat down, I was like, okay, I need to do planning for the next half year.
Soeren: And I really, I just prompt, I had a prompt and told the prompt can you write me a document planning the next half year with, and with a few things in there. And it's sped out this really long document, structured with a lot of blah, blah. Of course I couldn't use it like that. But then going in and filling that framework was, it was so easy for me to just, okay, I have a structure and even the structure was not a perfect one, but something to get you started with.
Soeren: And that's where it can be super, super helpful. Basically just for creating text that I need to create. I'm not really writing a lot of emails that way. That's more, sometimes I do that [00:34:00] privately. Like I, when I have to write an email to one of my kids, teachers, just like one of these standard blah, blah emails or like looking forward to
Soeren: meeting you at the parent teacher conference just write me an email because there's a lot of boilerplate text that we use in everyday life, right? If it's really something like that, I'm like, I might as well. And of course, I go over those things. You always need the human in the loop. So there's definitely some some areas where I'm using it quite a bit.
Soeren: And as, as Microsoft is releasing more and more products with it. Teams, for example, talking about Teams again if you have your meeting transcribed, it can also create a meeting summary. It can create the meeting notes, which I think is really awesome because if I'm driving a meeting I don't want to take notes at the same time, right?
Soeren: I want to be involved in the discussion while I actually want to steer the discussion. So do I sign one of my peers all the people that are there for discussing things? Should I really [00:35:00] relegate one person to taking notes? And of course the notes need review. It's not it's not perfect, but there are so many applications where it can come in really handy and does work that otherwise might be really cumbersome for the language industry at large.
Soeren: I feel like there's so many things. And I know that there's tools developers that are on it. I'm actually also teaching, teaching localization. So in one of my classes, I, for one of my classes, I had this little terminology mining exercise and was like, Oh, I should just try that with a large language model, shouldn't that be able to do that?
Soeren: And I tried it out. It was like, Oh yeah, that's probably one of the applications where you just say from this long text, please extract some important terms. And as always you need to take that with a grain of salt, right? You need to refine that. You probably need to look at what would be the best prompt for this.
Soeren: So there's a lot of things that you need to do to make it work on a regular basis. I'm also thinking about you can [00:36:00] use it for translation as well. You can also use it for creating text in parallel. Not everything needs to be translated, right? You could also say. If you have a really clear help topic, why not just say write that help topic again in French that obviously you need to understand how good the models are in those different languages.
Soeren: And there's languages for which there's really no good models yet. And you will need to check it for accuracy. And that's something that I've also played with where I ended up with ChatGPT 3.5. I just had it. Write a biography of a not that well known author that I knew and just invented all this stuff.
Soeren: Then the next ChatGPT 4 was a lot better but that taught me that yeah, you have to be, you have to be really careful when it comes to, to accuracy it might be good. Ultimately it's the most powerful autocomplete on the planet. That's what a large language model is, right? It's basically.
Soeren: Coming up with a [00:37:00] thing should be next in a text. And for certain things that can do it really well, I've actually done recipes with large language models. And then, because recipes always follow the same format, right? And it would never say take rice, meat and gummy bears, because it has, there's probably not a whole lot of recipes out there that was trained on, they would have that combination.
Soeren: And the recipes all so far have worked out well, my family has not complained. That cooked something based on ChatGPT.
Stefano: Yeah, your family is not Italian. We will complain about a recipe written by AI.
Soeren: Try it out. See what it comes up with. I found a recipe, Edward. And as I said, like you need the human in the loop, right?
Soeren: Somebody needs to take a look and say oh, wait a minute. That sounds like far too much salt and far too much pepper. And put that in there. Yeah but I think there's there will be lots of there's already applications in the industry, but I think that everybody working in the language industry should look [00:38:00] at large language models, just play around with what is
Soeren: what is available and try to think about, hey, how could we use that in, in, even in a process in translation, right? I've, I even tried it for rewriting texts or the like advanced copy editing given it very mean text and said, Hey check this on whether this is inclusive, whether this text is inclusive and completely rewrote the text because it was not inclusive at all.
Soeren: But I just want to know, Hey, does this work, right? So there's so many applications. You just need to know, what do you want to get out of it? And then really understand how to write a proper prompt. And all that. Lots of people are learning that, how to prompt the large language models best.
Soeren: So start there and experiment a little bit with it. And yeah there's a lot of things going on at large. So try to see as well, just get inspired with what other people are doing. So
Stefano: basically either evolve or be [00:39:00] replaced,
Soeren: Yeah. There's a lot of conversations about how translators will be replaced or writers might get replaced.
Soeren: I think People were already afraid of translators being replaced when CAT tools came out and were like, oh, then everything will be recycled. No. And in certain areas recycling became less because people got bored by text that sounded too boilerplate, right? When I was talking about humor a lot of the text is a lot funnier now, and you need more expertise for that.
Soeren: But then humor is also challenging for machine translation, and It's also challenging for large language models to come up with something really funny that appeals. And so I think people will not be replaced that, that quickly. They might need to shift some of the work, right?
Soeren: But I, I think that whenever we get a new tool, it's normally the more cumbersome tasks that get replaced. And yeah, as I said, like I'm. I wouldn't write an [00:40:00] email where I have something really personal to express with a large language model. I, as I said, like I do that for the boilerplate things where I don't even want to write the text in the first place.
Soeren: I'm like, yeah, this is basically blah, blah or I need to have, otherwise I would have probably looked it up on the internet. Like, how do you write a text like that that I find helpful, and I'm not afraid. That AI will replace people throughout.
Stefano: Yeah, I like the legal dilemma, especially in languages, because, for example, if you're driving a car, you're not driving a car, it's the AI driving a car, and you have a car crash.
Stefano: Is it your fault? So that is, opens an opportunity to say, in language translation, let's say, if the AI makes a mistake and Microsoft, let's say, publishes, That mistake, who's, which entity, who's fault is it? Is it Microsoft or is it AI? I feel like [00:41:00] humans are not going to be replaced because we want a human to say, yes, I sign off on this.
Stefano: So maybe we're going too deep into the realm of the nerds here. Soeren I think you might need to go to work and start your day. So I don't want to keep you here. We don't want to keep you here for too long, but I was. That's for us a super interesting conversation and I hope you have some fun. You had some fun.
Stefano: Oh yeah. And maybe we can do this again in the future. If you, if your cats love you,
Soeren: I actually had to quickly feed the cats before. Before the call started a little earlier than they normally get fed, but otherwise they would be complaining. You would have heard them. And despite all that, which is also AI, right?
Soeren: People don't notice that, but like the noise filtering that you get in zoom or in teams. The fact that you don't hear all the background noises is great use of AI. Did we need it? Absolutely. I think it didn't [00:42:00] replace anybody. There was nobody who would shush your any background noises for you.
Soeren: There's so many applications where I'm like, yeah, it just makes life easier. So I'm optimistic about AI. I'm excited about all the opportunities. That there will be because I, as I said I feel like our main challenge is bridging that, that gap. And as you were saying humans also like to take to talk to humans ultimately, you need to have, you need to have the humans in there.
Soeren: Yeah, I don't think that people in the language industry can really be replaced. You need the experts that bridge the gaps between cultures, languages, markets that understand two of them and or even more and can make sure that things work across those gaps. I think we have lots of exciting work in front of us.
Soeren: Very optimistic. Yeah.
Stefano: I love your optimism, optimistic approach. So I guess we can close here the conversation. Thanks so much. And have a good day. Sorry. Yeah. Very nice meeting you.
Soeren: Take care. [00:43:00] Thank you. Bye
Stefano: bye. Speak soon. Bye bye.
Streamline your translation workflow with zero downtime. Discover how.
