Episode Transcript
[00:00:03] Speaker A: Hello and welcome to Localization Today, the podcast that takes you inside the language services industry to highlight the innovations reshaping it. I'm Eddie Arrieta, CEO of Multilingual Media. And today we are uncovering a breakthrough in AI driven translation, segmentless localization.
This is powered by Burroworks Free Flow Editor.
Imagine translating entire pages or articles in one continuous context. Rich stream. No more disjointed sentence by sentence segments. We'll explore how this approach tackles the challenges of preserving meaning, boosting productivity and redefining human AI collaboration. Joining us is Gabrielle Furman, founder and CEO of Burroworks, whose platform is trusted by global brands to manage high scale language operations. Pour yourself a coffee, water, iced tea, whatever your preferred drink is, and get ready to rethink how we localize in the age of AI. Gabrielle, welcome once again to Localization Today. How are you doing today?
[00:01:16] Speaker B: Doing great, Eddie. Love that intro music. Very good to be here with you. You're a great interviewer. So always exciting to talk to you, always.
[00:01:23] Speaker A: All right, Gavriel, and thank you so much for doing this because as I was mentioning off the recording, we've had an ongoing conversation in Localization today with different leaders from let's say the corporates that would call them, those that buy the services and those that are producing the technologies and the services.
And we've seen this duality right now in terms of how it is done done. And on the one hand you have operationally speaking, so you'll have the orchestrated approach and then you'll have plug and plays and some companies seem to really like that.
And also you have the other free flow, let's say not free flow. I'm thinking now your concepts, the one stop solution where you just hire one company and that company, that's everything for you. And that sometimes might be just your internal corporate team. And what I've heard recently is that actually the big ones do both.
They orchestrate and they do, they just, they just make sure they have everything at their disposal and that they can, you know, you know, stress systems whenever needed for whatever their goals are and they just have a lots of safety nets to be make sure that they do it. Now I was mentioning to you, very happy to hear segmentless and semic less localization. So that's very interesting to, to bring that to the table. Thank you so much for being here once again and really lovely to get to see your, your ideas.
[00:02:53] Speaker B: Thanks Eddie, thanks so much for, for having me. And yeah, I'm looking forward to an exciting Conversation and hopefully getting people thinking about what a future.
And they say a on purpose. I don't think I believe in the future. I believe in multiple futures. But what a future could look like for translation content as AI continues to get better and better.
[00:03:16] Speaker A: Yeah, that is really good.
What are the different futures? And they are, as we speak, they are presenting themselves right present being that manifestation of that future that we thought of. And I do see wonderfully many different approaches to this stage in the adaptations and also the, how should we say, the interpretations of the uses of technology to solve language problems. So it's wonderful to see the variety and how we learn from, from each other's experiences. I think that's also very wonderful to see. Gabriel, what is segmentless localization? Can you describe how, you know, free flow treats entire documents or large blocks? What is it? How do you think about it?
[00:04:06] Speaker B: Yeah, I think the easiest way to describe it is to go in a journey back in time. Right. So when localization began early on, the first big challenge was how do you parse and standardize files? So, you know, how do you get, let's say, a DOC file or a InDesign file or a PowerPoint file, and how do you parse it so that regardless of the file, you're working within the same structure? Right. So parsing is this process of going from many different file types and possibilities into a stable, predictable structure. And that parsing also typically has segmentation, as I would say, a subcomponent of the parsing which is going to break up the text.
So parsing is kind of like understanding the text in some way. It's not in a syntactical level. So like breaking apart the pieces, extracting, let's say, the. The actual text from the pieces of code around it, because we don't think about it. But typically, for instance, a Docx file is an XML file deep behind it. So. Right. So the parsing process is extracting the content. Segmentation is breaking that apart. So whether you're going to segment by when there's a period or whether there's a line break or however you choose to segment text, that's kind of like part of the framework of a traditional translation management system or a CAT tool.
[00:05:35] Speaker A: Right.
[00:05:37] Speaker B: And the idea of segment list is going in the other direction. So the idea is, okay, you can still parse a file, but now maybe you don't have to break it up into small little pieces. Maybe now we can work with much bigger chunks of content.
And I want to frame that. I think that's not like a panacea. It's not a cure for everything.
It's good for very particular contexts, very particular situations. And it's not going to be that good for others. Right. So just as an example, maybe you're translating software ui and I would imagine that it doesn't make any sense to try to get rid of segments with software ui. It's part of the nature of the beast.
But a good example is if you're translating, let's say, a LinkedIn post.
A LinkedIn post is typically a short form content that has a beginning, a middle and an end. And it's typically all connected from a meaning perspective. So maybe the pun at the beginning is connected to something at the end and maybe a little, little reference in the middle is connected to something that was said in the beginning. And you have, you have humor and you have changes in the way you express things. So maybe, you know, maybe the original post in English had, I don't know, 20 lines. And maybe your post in Spanish for Colombia is going to have 30 lines because you're going to want to explain some things a little bit more. So you start to get into a lot deeper level of adaptation and change of that same content when you go from one language to another. So I think the idea of segmentless is enabling those deeper transformations to take place.
[00:07:22] Speaker A: So that's very interesting. So it's almost like there is an amplification on context, I would say. So segments have certain limitations.
How does this just to stay in segment localization, segment less localization.
How does this approach impact translators view of context and coherence? Because what I'm getting is you're mentioning there is a transformation and I'm curious to understand why is the transformation happening. But how does this impact the translators views on context and coherence?
[00:07:58] Speaker B: Well, a. I think it's still early for us to respond with data. As far as there's too little data for us to say assertively. How does this impact the translator experience?
In my opinion, it's very liberating because when you're translating like let's say a LinkedIn post and you're doing it in a translation editor just like ours for instance, or any other editor, you're kind of stuck as a translator and you have a sentence in English and now you have a sentence in Spanish and maybe, sure, you can join segments and some editors and maybe you can add a segment here and there, but you're fairly constricted.
And then when, and when you do that as well, you start to mess up any kind of translation memory as well. So, you know, whenever you join or you, or you add segments, all of a sudden you start to get disjointed translation memory. So that's a, that's another issue. But you're basically confining people to a more transactional role where they're getting a sentence and they're packaging that sentence in another language.
You don't have that much degree of authorship over that.
When you have, let's say a whole LinkedIn post in front of you and you can delete a few sentences, you can change the orders of the paragraphs, you can add a bunch of text, remove a bunch of text, you can re. That gives you immense authorship capability. And that's why in our opinion, the, the, this whole segmentless thing that we call free flow is about really pushing the boundary of what is translation versus content creation. Because in my opinion, translation is content creation. It's just content creation that has a very specific briefing, a very specific idea of what needs to get done. But nonetheless, in my opinion, translation is a form of content creation.
[00:09:54] Speaker A: So you believe there is like a transcreational process when you like whatever the original text is, it's the original text and then anything that goes from there is transcreation.
Just very, very, very.
The wording you use is perfect. How did you say it was a very restricted instructions, correct?
[00:10:18] Speaker B: Oh yeah. I mean, I think think of the original text as a briefing, right? And like, you know, you get a briefing of, you know, maybe you want to create an ad or maybe you want to write a blog and you know, the briefing says, hey Eddie, I need a blog that is about 500 words long and it does this and it does that and it uses these SEO terms, right? And that's typically like content creation, you have some kind of briefing and someone goes ahead and writes it out.
Now typically with translation, in my opinion, it's the same thing, but the source text is a very, very strict briefing.
This is a sentence. Now say the same thing in another language. This is the sentence. Now say the same thing in another language.
By breaking from this, you know, one to one correspondence of sentences and thinking about the text as a whole, we can allow ourselves to think of the source more of as a, as a briefing and less as something that needs to be followed bit by bit. So for instance, again with the example of the LinkedIn post, if you're translating it line by line, you're going to get a result that's going to be line by line. Like I said, maybe it's going to have a few lines more, a few lines less, but you're still kind of stuck now with the free flow. What we see is maybe you have a LinkedIn post, let's say in English, that has two paragraphs and is super funny and provocative, but maybe in Japan you're going to have four paragraphs and it's going to become serious and it's going to become conversational, it's the tone is going to change, maybe it's going to add content that's not there. Why? Because now you can treat that original sentence, the original text in English not as a definitive, solid, inflexible briefing, but just as a direction. Right. This is the direction that we want to take things.
Now the translator becomes enabled as a content creator. Right. So that's, that's the idea behind it.
[00:12:21] Speaker A: And this is the part of that conversation that we've had before about the sophistication of the craft in many different places. And how would you see the future of work, you know, unfold and the different interpretations of the, of the translators, craft interpretations and then those that work in culturalization and you know, global content creation.
That's a very important take for those that are listening to some of these concepts for the first time and they are more familiar with segment based tools and you know, the line by line that you're mentioning and each line to translate each line. Can you compare the free flow with the traditional segment based tools? Like what are the limitations of the one? And then how does the segmentless free flow help with this?
[00:13:14] Speaker B: So segment based tools, they're working with what people call relational databases. So relational databases, they have a one to one correspondence, right? I don't want to get too technical, but the idea is, let's say you translate a sentence, my dog is beautiful, and you translate that into Spanish as mi pero.
That's going to store that, think about it as an Excel, it's going to store that in an Excel file. Obviously it's just a database, it's just an example. And then that, that entry can be retrieved later in the form of a translation memory. And now you have, let's say now my dog is really beautiful. And you now it's gonna, it's gonna find the similarity and you're gonna get a match, probably like a 77% match, something like that. And the translator is going to fill in the blank to mi pey lindo. They're going to add the muy and, and the edit the cat tool or the editor is going to facilitate that retrieval of knowledge in a very specific and syntactical way. Right, That's a translation memory. Now in the free flow, because you're not working with segments anymore, the search for things isn't relational anymore.
It's what we call non relational or vector driven search. So in the vector driven search, what we can find are what we call semantic correlations. So for instance, in the syntactical correlation, the system has no idea that there's a similarity, let's say between Perro and Gato, or like dog and cat, because after all they're both pets, as an example, mammals.
But in the free flow, because you're working with semantic correlations like vector driven things, you can actually, and this is not our work, this is just what embeddings do and what large language models models do. You can understand that there is a similarity between a dog and a cat and that there is a similarity between, let's say, beauty and aesthetics. So these concepts, they're stored in a, in a less tight and a broader way, which makes it on one hand a little bit less predictable. On another hand it makes it a lot more leverageable. So let's say you know, you write, you translate this text that's all about dogs, and then you translate another text that's very different, but it's all about cats.
We can still find correlations between dogs and cats.
So it's a different way of retrieving past knowledge. You're not just retrieving knowledge in the form of translation memories, you're retrieving knowledge in the form of context. So it's a. And you could retrieve, ultimately it's open to retrieving any kind of knowledge. So you know, if you have videos as reference sources or websites, you don't need, you don't need a structured translation memory to serve as a guide. Really anything ultimately becomes an embedding. So it's multimodal by design. So you could have a video, you could have a song, you could have a website, you could have a transition memory, you can have whatever you want and it's still in some degree leverageable. Right? So it's, it's very different ways of managing knowledge.
[00:16:31] Speaker A: That's a really good way to put it. Gabriel, thanks. I think finally starting to get it. And I think one important element of this or one important question for me to ask would be is there still a place for segment based tools or would you be ready to call the death of the segment based tools? And this is the Future, Right, because you will need this type of contextualization to do better, right?
[00:16:57] Speaker B: Yes and no. That's why I don't like the idea of the future. I like the idea of a future because, like I said, I think there are many, many, many, many, many different futures. And like I was saying, you know, for software localization, I don't imagine software localization, at least for now, going down that approach. I don't know if it makes sense. I think it's still important for I think there's a lot of value in working with specific strings that have separate key values that are not in some way jumbled together.
But like I said, that same segment based approach doesn't work really well for a blog or for sometimes even a legal disclaimer that's going to change very significantly from one country to another. So I think this is really, really powerful for when there is a deeper need of adaptation in that content type.
So, and again, with software ui, just as an example, maybe you can adapt the labels, but you can't really adapt the structure because that would be a new ui. And so you have to translate within the constraints of the ui. So when you have to do that, segments make a lot of sense. But with a LinkedIn post, what's stopping you from having three paragraphs in one language and two paragraphs in another and five paragraphs in another? From being really funny in one language and really serious in another, from having an image in one language and not having an image in another. So there's nothing stopping you in that kind of content.
And it's in that kind of content that I feel that this is particularly strong.
I also think that there's another very important use case which is for many different content types. What ends up happening is that at some point in the content production life cycle, the content gets translated, okay, gets localized, like let's say a blog, right? It gets localized, but then that blog gets placed with it gets published somewhere either as a draft sometimes, and people in that market will typically tweak it and change it and adapt it and do things to it. And those things are typically then not reflected back into translation memory because often they can't. So it's like, you know, let's say you have a product, but the product works a little bit different in country A versus country B.
And you start getting into these real variants, right? These real variants. It's not just like you're translating product A to country B, you're actually changing the product. Maybe, maybe a feature was added or features were removed and all of a sudden you need a framework to deal with that. And there's a lot of last mile changes that are done by markets. I just don't get captured by these tools because they, they're not really compatible with them.
And the other thing as well is when you get a content writer, they don't. I've never seen anybody in my life, for instance write content in a CAT tool, right? It's not designed for that. You write content in Google Docs, you write content in all kinds of like content production systems. But, but not in a CAT tool. It's not designed for that.
And we're trying to create a space where you're still translating but you're also writing.
And, and, and then in order to do that you can't really work with the constraints of a segment. But like I said, this isn't for localization in general. This is for when you really need a certain degree of adaptation.
[00:20:20] Speaker A: Really good way to put it.
And that it's great to have there so that there is clarity for those that I know that were going to ask me about this. And it's very interesting what you're also painting in terms of a place where you can write and also get this type of feedback or these translations that you could do, right? This localization that you could do from there. I use Grammarly for certain things. I usually just copy paste things there. But I end up writing just to get kind of like my grammar correct that and it is a great experience rather than having, you know, second men to correct my grammar like that would make no sense. Right. So I agree with you Gabriel that this is probably really good for. I can already see how we do for some of our, you know, content clients where we'll do like two languages. Like it becomes really complicated for us to do two languages like oh my God, we have to folders. And it's because we are not going to do it in a line by line app. You know, it will be one of our translators trying to use like ChatGPT or something. So for us marketeers, I think what you are proposing sounds very appealing in terms of the, in terms of the quality this approach is. Is there a benefit? Have you seen anything from your experience how does specifically preserving paragraph and page level context improve that translation quality?
[00:21:47] Speaker B: I don't think it improves.
It's a great question. So again the, the key thing is what is translation quality? Right?
So when people think about translation quality, the first thing that comes to mind is is it saying the same thing as the source. And if it doesn't, it's bad quality translation. If the sentence in English says I love my dog, and the sentence in Spanish says, quieron po qua mi perro, that's bad translation quality.
You're not communicating the same thing. So it's bad, typically. Right? And everybody has their different definitions and error typologies and all of that, but this is different. Quality doesn't become a matter of relaying meaning. It becomes a matter of effectiveness in the text. Right? So again, you're right in placing this as a marketing tool. Right? And what do you, what do you do when you write marketing text? You're trying to get an outcome. You're trying to get a reader to click on the get to know more button or the subscribe button or the buy button, or, you know, you're trying to get people curious, you're trying to get people enraged. Whatever it is you're trying to do with content, you're, you're, you're looking for an outcome, right? That's any, any content writing is looking for that outcome in form, etc.
So what if you change completely the translation to get to that same outcome? So again, you know, maybe that your translation in English is going to be really provocative, but maybe you're. Whatever you're saying in Spanish is going to be for some reason, maybe your audience in Spanish is going to be of a more senior demographic, more conservative demographic, and you want to tone it down and you want to make it more conversational. You also want to explain more things. All of a sudden, whatever the source says is going to be really different than whatever the target says. And in my opinion, that's actually really high quality, not really low quality. But again, not quality within the traditional localization framework. It's quality within. Does it do what it's setting out to do?
[00:23:55] Speaker A: And this is a very, very important point that you're making, Gabriel, because I think one of the things that I'm always curious about is how, like, um, I don't know if the listeners are familiar, but in the human body, we have different types of, like, nervous endings. So some are smarter than others. So when you touch yourself, you know you're touching yourself. Some of the nerves in your mouth, they. Everything they perceive is pain because they are not supposed to be exposed. So even heat or coldness just feels like pain.
Touch feels like pain. And in some places in literature, they are referenced as, like, dumb nerves.
Something like, they just feel like pain is like, bad, terrible. But they are not understanding whether it was like it was pointy or if it was soft. So this probably speaks a lot about that evolution on the concept of, you put it really well, effectiveness. And the effectiveness then it's a more human, let's say a measurement than 1.2 than 9 over 10. Effectiveness is something that it's to a certain level, very subjective and that probably could be trained. So I think that's a really important, really important way to do it or a way to think about.
Seems to me very complex still in terms of how technically to do this. And probably you are bringing it down to our level. But you know, there are some other questions that we have here on where do sentence by sentence workflows break down? Is there a limit to this context that you're getting? And you were talking about vectors as well. You're talking about like the places, the repositories of context. It comes down in a lot of the conversations here, they always tell me, you know, it always comes down to the training data. Does it then get to then yet another kind of like limit on contextualization for where you are proposing. And I'm still trying to think about it in terms of workflows, I still don't understand how does it actually work. So if you could give us a few more words on how it actually works, that would be great.
[00:26:17] Speaker B: Let's say again, you're Translating like a LinkedIn post, right? Or a blog. Let's talk about a blog now.
And let's say you have SEO terms or a glossary, something that you want to follow in the free flow editor, you're still going to be able to follow that glossary. If you've translated things that are similar in the past, those things are still going to be suggested to you. The difference is that what's behind the scenes is a different mechanism than let's say a translation memory. So that's the key thing. From the user experience perspective, it's pretty similar.
From a behind the scenes perspective, it's very different.
And the reason why it's different is this previous way of working.
And this is something that I think is super important. So this previous way of working, it has a lot of value because like parsing files is something very important. Segmenting files is something very important.
If you, you know, run a little experiment, try uploading a document that's a few pages long into GPT, asking it to translate it for you, it'll say that the document is done and you won't be able to download it because it's, it can't parse and restructure the file. It'll lie to you, it'll say a bunch of different things. Maybe it'll generate a file if you push it too much. But it's not going to be translated. There's going to be gibberish in the middle. So parsing is important, right? And parsing is a very.
Yeah, it's just a very important thing. And it's, and it's useful. I'm not here to talk poorly about parsing. The thing about the free flow is it's about being able to work with bigger chunks of content and dynamic chunks of content. So when you're parsing, for instance, let's say you translate sentence A, it's going to look for something similar to sentence A in the translation memory and it's going to use syntactical deviations. So if you have like a little comma now, all of a sudden you don't have 100% match with something. Now you have 98 or 97 or 99% match, depending on the, the length of the sentence, etc. So that's typical. That's typically how that's going to work. Now, let's say in, in, in, in this context, maybe it's not looking for a specific sentence, it's looking for things that they can leverage. So, you know, maybe you have the concept of dog somewhere translated and it happens in these sentences and you have this concept of beauty and it happens in other sentences in other contexts and through embeddings, you can put this all together.
So one is working through various, you know, conventional search processes that are looking for textual similarities. The other one is looking for embeddings, which are mathematical representations of text, which can be way more multidimensional than just, you know, letters and characters.
[00:29:07] Speaker A: And you said, you know, perhaps translation quality might not be, you know, not necessarily effective, but effectiveness would be, would be comparable in many cases. And, but in, in terms of mismatches and rework and, you know, efficiency, how much time and rework do segment mismatches cost at scale? When we're thinking, you know, large amounts.
[00:29:35] Speaker B: Of information, again, I think it's, I think, I don't think that's actually the problem in my opinion, Eddie, because the, the problem really is this is the problem.
AI, in my opinion, is only going to get better. Humans are still the owners of culture. So when you get a File that's like 99% there, let's say, and all you have to do is tweak it. If you don't have the freedom to really tweak it. And all you're doing is adding a comma here or there in every segment. You're not adding a lot of value. But if you become an author and you get the translation, the translation is pretty much there. But now you can think about it as an author and you can change the order entirely of a paragraph, you can strike out a huge part of the text, you can change the tone, you can explain things more. Now you have freedom. And because you have freedom, you have authorship. And because you have authorship, you're adding value. Right? So our vision is there's still so much for humans to do once we figure out the translation bit. Because if you think about translation as one little step in a much bigger content creation process, then the translation bit is just the beginning, right? There's everything else that happens after that, and that's what the free flow is all about. It's about everything else that happens after that. It's about having one version of the text that's perfect for you versus one version of the text that's perfect for me versus having one version of the text that's perfect for someone that's riding on the bus.
So that's what the free flow is about. It's about going able, being able to go from source to multiple different variants of that same text really easily, predictably, in a way that leaves you with audit trails, in a way that leaves you with an understanding of what's going on.
And you simply can't do that. Right. So it's not, I wouldn't. I think the framing of the question suggests that it's actually kind of there to replace segment based translations. And it's not, it's just, it's a different tool for a different world and a different purpose.
[00:31:42] Speaker A: Gavriel. So it's great because then we can get into the fun part for me here. What are the use cases that you see? What would this type of approach be useful for you? You've talked about some social media. I definitely agree. My marketing team and the type of work that we do as we are thinking of different languages, that's the type of scenario that probably we want to see. What other use cases do you see?
[00:32:09] Speaker B: Well, anything that involves some level of adaptation.
So email marketing, campaigns, blogs, copywriting in general, letters, videos, any content that's really going to require like a deeper degree of intervention so that it's effective in a given purpose is going to be perfect for it. So the, the, and the use cases like I Said are they're really broad, right? Sometimes you're translating a company memoir, something that could be super dry, super simple. But maybe again, the way you communicate with your employees in the US is going to be different than the way you communicate with your employees in Brazil. Maybe their, their level, their academic level is going to be different. Maybe you want to make it simpler, maybe you want to make it harder. Maybe you want to make it different for Germany versus Japan.
And like I said, these decisions, they change things, right? So let's say you're making the text just more accessible.
That's probably going to be. Make you unpack. One sentence is maybe going to become now two or three sentences.
And that's not just the vocabulary, that's going to change. Maybe you're going to want to explain things more. That's the kind of thing where the free flow comes in. Because now you can, you're not worried about, oh, I only have a segment and now I can add segments, but now the segments. You know, think about it. When you add segments in the translation editor, they don't have a source segment to correspond to, so they become lost in the ether.
In this case, you can add an entire set, paragraphs and paragraphs to something and it's fine because we don't need a one to one correspondence with the source.
You know, like I said, I think the best idea to understand this is the source. It's just an abstraction. It's just a briefing. You know, just like someone would tell you, Eddie, write a post about the future of localization.
Make everybody really excited about it. You know, we want to paint a really beautiful picture and you get all this stuff, you know, and you're like, okay, I'm going to write the post, right? It's, it's broad strokes. And then the writer does the fine strokes. This is kind of the same idea. The source becomes these broad strokes. The translator becomes, does define strokes within the broad strokes. But you know, that could mean going bigger, going smaller, that could mean going in a different direction. It's just a briefing. And that's why I think it's so, so important because I really think in. And this is kind of our, our vision is that it's not a matter of, well, whether or not translators are going to be valuable in the future.
The question is how I'm sure they're going to be valuable. The question is how are they going to be valuable?
And the reason I think the, the reason I think or the way I think they're going to be valuable is through the richness in cultural contextualization that they have, they have a lot. Translators have so much knowledge about where they are, about where they're from, about the people, about the culture, about the little subgroups of people, people about the zeitgeist of that place, about the history, about the tendencies, that knowledge is priceless. But they have to have a way to channel that and that knowledge. In my opinion, when in a segment world, it becomes very pasteurized, very watered down. In a segment less world, that knowledge can be used however they want to, but it will require almost learning, like a new trade. Because, you know, in my opinion, it's fair to say that a commoditized translation, so something that just says the same thing as the source is for most parts of very little value. Because models and AI and neural machine translation can do a decent job at ensuring that some of the preservation of meaning, if not all of it, is there. But like I said, I don't think that's where language stops. I think that's where language begins, even in English, right? You can write one thing and you can write hundreds of versions of those things until you're happy with whatever you have, and maybe you're going to be happy for a given use case, but not for another, and etc. And I think that's the same thing for translation. So in our minds, what we're trying to build is the framework where it still makes tremendous sense to have people owning text.
[00:36:44] Speaker A: And that approach paired with uninterrupted flow, that brings new possibilities of all those features that you were mentioning to imagining conceptually, what, what do you think are the possibilities that arise when you can translate, when you can review and publish everything in one uninterrupted flow?
[00:37:04] Speaker B: Well, I think it's really ownership. That's what arises. You know, it's like if you're restricted to these little boxes, you don't own the text. You didn't create the boxes, you didn't decide how they're going to be laid out, you didn't decide how long it was going to be, you didn't decide how short it was going to be. If you think the text is too wordy, you can't really make it more concise. If you think the text is too complicated, you can't really simplify it. If you think it's too simple, you can't really add to it, you're stuck. The briefing is very, very. You know, it's kind of like receiving a briefing for a post where it's so long that it's pretty much the post itself and you have zero creativity. It's not going to be much fun creating that post, right? If someone gives you just a few broad lines and says, eddie, can you think about something in this direction, in that direction, blah, blah, blah. And they're, they're, then they're instigating you to think and provoking you. You're probably going to have a lot of fun writing that post with the free flow. It's kind of the same thing, except the text doesn't have to be provoking you necessarily. It's like it's the idea that you can take it in whatever directions you'd like and you can ultimately even deliver dozens of different versions of that text as well. So you can deliver, you know, a very witty and funny post for some audiences and you can deliver a very sober and serious post for other audiences. And those posts maybe look very differently too. You know, one may be longer, one would be shorter one, one may be brighter, one may be darker.
But you have now the tool to do that and to integrate that knowledge, right? Because if you could do that in Google Docs, you could do that in Word. But what about all of the little decisions that you made? You know, what about the glossary? What about all of these, these editorial decisions? They'll get lost. And when you get a new briefing, you're gonna have to do it again. So in the free flow, you're preserving that in a different way than the translation memory, but you are preserving that knowledge, Gabriel.
[00:39:05] Speaker A: And of course, in this context and in these different scenarios where free flow can work, how does it currently integrate with, you know, empty machines and post editing workflows?
[00:39:19] Speaker B: Yeah, it integrates 100%. So when you start translating in the free flow, you're going to get a translation suggestion and it's going to benefit from our context sensitive engine that is a combination of neural machine translation plus a ton of other data sources, both pre edited by large language model, etc. So you still get all that benefit. But that's just the starting point, right? That's just the starting point. You get the translation. You're like, okay, now where do I take it from here and for what purpose and in what directions and how many different versions do I want of this?
So you still get all of those benefits, but those benefits are just the beginning, that's not the end.
[00:40:00] Speaker A: I think one of the biggest things I'm getting out of the conversation, Gabriel, is the interactions that you see humans having with artificial intelligence and with this New technology, right? This new tool segment, less, as you're thinking about it, the editor and this context rich scenario or ecosystem in which we'll see now translators and creators, how is it going to affect the relationship between human linguists, human translators, human interpreters and artificial intelligence?
[00:40:31] Speaker B: So in my opinion, it's going to make give translators, not interpreters, because interpreters wouldn't be using this. I mean, they could if they wanted to use their knowledge to translate, but this is meant for written text.
And it's just going to make them, if they can adapt to this, they can, they become, their work becomes infinitely more valuable because now it's not only packaged as translation work, it's packaged as content creation work. And content creation is a more valuable and more relevant kind of work.
So I think it just gives a much longer Runway for meaning and for relevance for translators. Because again, now the argument isn't like, well, I can just put that into a large language model and it'll be translated for me. Now the argument is sure, but I am the one who's going to be able to make sure that it is adapted and perfect for each and every one of your use cases and audiences, and that we are building together a language culture around your brand, around your product, around your community.
Right. So it makes it so that language can actually have owners. And without owners there's no, no value. Because I mean, the, the large language model can translate, just can't own the translation. That's the difference. Right. Humans can own things. You can own something like, let's say, you know, you're in a bad mood and you're mean to someone. You can own that. You can say, I decided to be mean to that person, or, you're really nice, I decided to. It's a decision.
Large language model doesn't decide things. Sure. You know, people can argue. Yeah. Agents can make decisions, they can't own the decisions. Even from a legal perspective, human can own a decision. The large language model cannot own a decision.
So. But in order to own textual decisions, you need the right tools to do that. And if you're an owner, you have a lot of value because now you're not just translating, you're owning the culture of communication of a certain brand, a certain company, a certain product community.
That's the big shift.
[00:42:36] Speaker A: And I really like the way you put it, Gabriel, because in terms of the current power dynamics, that there is this feeling that translators, interpreters, linguists, creators have no leverage.
And you know that the leverage is on, let's say, the technology Owner's side, not even the technology side, but the technology owner's side. And then that the ownership is completely rebalanced, or rather that it's imbalanced. And do you think something like this can reposition and help reposition and at the end of the day, you know, we are in a capitalistic world. Do you think this becomes a competitive advantage? I mean, looking at content in this way where you give ownership instead of restricting that ownership, do you think at the end of the day, this translation. Of course you do, otherwise you wouldn't be doing it, you wouldn't be supporting it. But I do understand that. Can you tell me why you. Why you think that gives that competitive edge?
[00:43:42] Speaker B: I don't think it gives a competitive edge. I think it's, it's. It just enables relevance in the future. I don't. I think without it, you just. You're irrelevant without something. And again, it's just a tool. The harder part is the change management from an intellectual perspective in people, right? Because a lot of us have been beaten to death with repetition in our jobs, right? So if you're translating for the past 20 years and everything you're doing is going segment by segment and making sure that whatever is in English is in the other language, and that's all you did. And let's say you got hammered by someone doing LQA for an agency, say, no, Eddie, this was wrong. The translation said big and you said very big. And your brain got beaten up for 20 years thinking like that. It's going to be really, really hard for you to release yourself from those shackles and say, you know what? To hell with whatever is lqa. I am the owner of this. And I decided that I want to use big and not very big for a reason, and I own that reason and I stand by it. And there's a rationale for that. So I think if you're able to do that, you have value, right? And that. But at the same time, that's very hard because it's the opposite of what so many of us were taught all of our lives.
So that's, that's kind of like, I think, the conundrum that we're all in. You know, it's like, can we change? Can we. The tool itself, honestly, it's just, it's just a tool. It doesn't do anything. It's like.
It's like a kitchen knife, right? What good is a kitchen knife if you don't have a good chef? I mean, if you're going to use a kitchen knife, like the way I cut onions and tomatoes. I could have the best knife in the world. I'm still a very, very crappy chef, but, you know, you give a very crappy chef a very crap, A very great chef, a very crappy knife, and they're, they're going to be 100 times better than I am. They're going to have a ton of value.
And it's the same thing, right. It's just a tool. The hardest part is getting people to shift mindsets.
[00:45:45] Speaker A: And that's probably going to be our evolutionary challenge, how we go to use these tools over generations and how many years has it taken us in our relationship with all the tools that you're referring to, right. And the knowledge that surrounds the culinary skills to be where we are at. It might take many centuries before we refine and those futures become all available and, and it's great to see you again, Gabriel. It's always great to, to talk to you. We are unfortunately coming to an end of our conversation. It's gone by really fast.
Before we go, is there anything else you, you, you might think we, we should let those that are listening know from our conversation?
[00:46:31] Speaker B: That's it.
[00:46:34] Speaker A: Great, great, great. And I think it's be gonna be very inspiring. I think there are many, many instances of this conversation that we're going to be resharing over social media. This is the end of our conversation.
So that was for everyone who is listening today. That is our deep dive into segmentless localization with Gabrielle Ferman of Burroworks. If you're eager to see free flow in action, check out the demo links in the show notes.
Thanks for joining me, Eddie Arrieta, CEO here. Multilingual media and we are in localization today. Be free to subscribe, rate us wherever you listened and follow us at multilingual media on LinkedIn and Twitter for updates. Until the next time, keep questioning, keep innovating, and keep localizing. Gabriel, thank you so much. Goodbye.
[00:47:23] Speaker B: Thank you, Eddie. Great being with everybody.