Manga is one of Japan’s best known exports, but it’s surprisingly hard to make money here.
Today we dig into exactly why this is. We sit down with Sho Ishiwatari, founder of Mantra, who explains how is company is trying to expand the global market by streamlining the translation and global marketing processes.
We also talk about why manga is so much harder than books for AI to understand and a few ways Japanese universities are trying to develop and inspire the next generation of Japanese founders.
It’s a great conversation, and I think you’ll enjoy it.
- The surprisingly complex manga translation process
- The real problem with fan-translated manga
- How to think about getting a 10x( or 100x!) improvement
- How the University of Tokyo supports startups and what other schools can learn from them
- Why translating manga is so different from translating novels
- The downside using contextual hinting wit AI/ML
- How to expand the global manga market
- What every Japanese university should be doing to encourage startups
Links from the Founder
- Everything you ever wanted to know about Mantra
- Connect with Sho on LinkedIn
- Friend him on Facebook
- Follow him on Twitter @mantra_ja (Japanese)
- Sho’s published academic research on machine translation of manga
Welcome to Disrupting Japan, straight talk from Japan’s most successful entrepreneurs.
I’m Tim Romero and thanks for listening.
Manga and Anime have been two of Japan’s most visible and influential exports. Japanese manga has earned its own section in US bookstores. And in the movie industry today, many of the world’s most successful directors and cinematographers cite Japanese manga and anime artists as some of their biggest inspirations and influences. But surprisingly, despite manga’s global popularity and influence, the global market is pretty small. There’s not a lot of money in manga. A
And today, we’re going to dig into that. We’re going to sit down and talk with Sho Ishiwatari, CEO of Mantra. And we’re going to find out if a startup can disrupt or even survive in the manga industry. Mantra has created an AI that can translate manga. But, as is the case with so many startup stories, the journey is far more interesting than the destination. You see, before the AI could translate manga, Sho when the team had to teach it to understand manga. Not just read the words but understand the context and the layers of implied meaning.
Sho and I talk about the nature of human understanding, how Japanese universities can better inspire the next generation of startup founders, and AI’s role in helping people understand each other. But you know, Sho tells that story much better than I can. So let’s get right to the interview.
Tim: So we’re sitting here with Sho Ishiwatari of Mantra, who’s bringing Japanese manga to the world. So thanks for sitting down with us, Sho.
Sho: Yeah, thank you for having me, Tim.
Tim: In the introduction, I give a really high-level description of what you guys do. But can you explain what Mantra is?
Sho: Yeah, sure. So what we are doing is to deliver comics, deliver manga, across language barriers. So we are building two products. The first one is a manga translation tool that is based on our machine translation technology. And another product we’re making is Langagku, which is a language learning tool based on manga.
Tim: That’s two really different lines of business.
Tim: Let’s dive into both separately later on. But the core tool, do you just translate Japanese into English or do you support other languages as well?
Sho: We currently support for language peers, Japanese to English, and English to Japanese, and Japanese to Chinese, and Chinese to Japanese.
Tim: Well, that’s interesting. So far, most of your business has been Japanese to other languages. But do you also have companies bringing in English or Chinese language manga to Japan?
Sho: So previously, almost all the Japanese comics read in Japan are made from Japan. But recently, there are more and more comic platform also in Japan, for example, from Korea or from China. They are having more and more content that have super high quality, so there are several contents coming in from outside Japan.
Tim: So tell me about your customers. You mentioned that you’re working with publishers and creators, but walk me through your use case. How would a publisher use this tool?
Sho: Basically, Mantra Engine, it works on browser. Before Mantra Engine, people has to use Photoshop and Excel and Adobe Acrobat to create the translated manga. And there are several people are involved in the project, like translator, checker, and then designer, who put the characters on the images. So the process will be very complicated and you need to connect to others by using email and send the files to everyone who are involved in the project, which takes time. And the translation and design, all the process done manually. But in our product, they can do that on the browser, the same tool for everyone. And also, it is partially automatically done by AI. You put the image into the tool, it automatically translate that. So it helps people do each process as well as the communication between those different people.
Tim: So it is AI translation but it’s not completely AI, there’s a post translation editing step too, right?
Sho: Sure. Sure. You’re right.
Tim: So from a publishers point of view, what do they gain in terms of like how much faster is this process in the real world, or how much less expensive is the process compared to traditional way? What are the publishers gaining by using Mantra?
Sho: One of our clients who is not a publisher but they’re translation company of not only the comics, but also other media. But what they say is that they can accelerate the translation speed, whole process by 200%.
Tim: So cutting the translation time in half.
Tim: And does that result in cutting the costs in half as well?
Sho: I think so. Because they’re hiring people, so I don’t think that that’s easy, but basically it should be.
Tim: All right. Well, in the ballpark anyway, like customers often won’t share those details.
Sho: Yeah. And it works on browser. So previously, the translation had to be done by the professional translators. Because if you want to publish in different languages at the same time, you have to share your unpublished data to translators before you publish, like, one week before or two weeks before that. But the thing is there are so many fan translators who want to translate but they are not professional translators. So what Mantra Engine enabled was that you can share that to not only the professional translators but also to the fan translators.
Tim: Have publishers been taking that up? Because I know there’s far more fan translated manga out there than there are officially translated. So, have publishers been using that feature in Mantra?
Sho: Yeah, one of the publishers, yeah, we are working with fan translators. And we let the fan translators use Mantra Engine.
Tim: But if you’re looking at an overall 50% reduction in price and time, that’s really impressive. That’s always been the promise of automation of any kind, right?
Sho: Yeah, you’re right. But we are not satisfied with this result, you know. It should be like 10 times faster or 100 times faster.
Tim: How do you get to that number? So, I mean, in startups, we talk about the 10x a lot. It’s a really good goal. And actually, you know, 50% is really impressive, honestly. But let’s say we’re going for that 90% reduction, right, what’s the bottleneck? What do you have to change to make that happen?
Sho: That’s an interesting question. Yeah, we are thinking about that. And it’s not an easy question. But as I had mentioned, there are three people who are involved in a single project — translator, checker, and then designer. So even if each process can be accelerated, but if you are working with other person, you need to wait for something. You need to wait until others finish, after which you can begin checking for designing. That is one of the bottleneck.
Tim: So changing the workflow to allow these processes to happen in parallel to some degree.
Sho: Yeah. And another idea is that if people can do multiple process, so let’s say if a translator can also do the design part.
Tim: Yeah. I mean, it’s a different skill set, right?
Sho: Yeah. But if we can assist the process, we can let the translators to do the design very easily.
Tim: That’s true. The more advanced AI becomes, the more likely it is you’ll be able to do that. Well, I want to dig really deeply into the technology and the markets in a minute. But right now, I want to talk a little about you. And let’s back up a bit. So, you founded Mantra when you were doing your post doc at the University of Tokyo, right?
Tim: Was this part of ongoing research you were doing? Was it a brand new idea that you and your co-founder had? How did you get rolling with this?
Sho: The thing began when I was doing the last year of our PhD. I and my co-founder, we were talking about the idea of Mantra. We were trying to make prototype and we were finding customers. We’ve already decided to make a company then but we haven’t made the company. We just did that as a side project.
Tim: Why this particular area? Was it just you both were very interested in manga or was it directly connected with your research? Why this particular niche?
Sho: Both, both. One strong motivation for me is to involve in entertainment area. I wanted to use artificial intelligence technologies to accelerate the cultural exchange. Very important culture of Japan should be manga or anime or video games. And also, I was doing research on machine translation. And my co-founder, he was very good at doing image processing and comics, consists of text and image, so we can make it.
Tim: And what kinds of support did you get from the University of Tokyo when you were starting this?
Sho: A lot of support. There’s so many programs that support the side projects or the startups. For example, we had summer founders program during your summer vacation. They don’t teach you how to make product because all the members, they have ability to create product but they don’t have ability to–
Tim: Right. It’s a bunch of programmers and engineers and researchers. They know how to build, they just don’t know how to sell it, right?
Sho: Yeah, you’re right. So that’s a program to tell you how to sell the product, how to understand customers, because that’s super important, too. And we also have FoundX, that is a kind of acceleration program for you UTokyo alumni. So we were the first batch of the FoundX program. We spent for nine months. They let you use a room for free, and also, they will teach you how to not only understand the customers but also find the investors.
Tim: So it’s not an investment program. It’s co-working space and coaching and advising.
Sho: Yeah. And I cannot list up all the support given by UTokyo, because there are a lot, not only FoundX but also Todai IPC program.
Tim: Does that support continue after you’ve graduated?
Sho: Yeah. It’s more about the network thing. Because when we were in FoundX, we have batch mates who are in the same batch, which is nice because everyone, we didn’t know anything about how we should make company, how we should make a key. So we begin with the same point. So there is a alumni network.
Tim: I think that network is far more important than the investment funds or the mentoring in the long run. I think that network is what builds a real startup ecosystem.
Sho: Yeah, I do think so. Yeah.
Tim: Let’s get back to the Mantra product itself. I want to dive into the technology and the markets here. So the Mantra Engine, is that based on a standard machine learning model like GPL or TensorFlow, or is it based on something else?
Sho: Yeah, it’s based on the one made by Facebook, and we modify several thing. It’s based on the normal transformer where we use algorithm but we modify that.
Tim: Now see, that’s the thing that I find interesting, because so much — translating a manga is really different than translating a novel or an email.
Sho: Yeah, yeah.
Tim: Because so much of the context is not in the text.
Sho: Right, right.
Tim: How accurate is it? How much editing usually has to be done after the AI goes through?
Sho: That is my most favorite topic. It’s so difficult to translate comics, right. Because if you translate novels, all the information should be included in the text.
Tim: Yeah. But in manga, I mean, the images are part of the story. They’re the main part of the story, right?
Sho: Right. So we should consider context more. And the context is not only the textual context but also the visual context, right. For example, in manga, a single sentence can be divided into separated speech balloons. When you read it, it’s easy for you, but the AI doesn’t know the order of reading.
Tim: Since so much of the context of the manga are in the images, how does the AI pick that up?
Sho: First of all, you need to extract the order of the phrase. In a single page, you will be having like five to six frames, which are not very clear, right?
Tim: Yeah, they can go in different patterns, right?
Sho: Right. So the first step is to detect the frames and recognize the order of the frames. After which we decide the order of reading, then we can extract the context.
Tim: Well, tell me a bit more about this context extraction, because anything could be going on in the background. You could have three or four different characters who are talking and none of that’s in the text at all. So what clues do you look for to extract that content?
Sho: The one thing is the position of the text indicates the reading order. Another thing is more visual information, like a girl is speaking or boy speaking, and how many people are speaking in the frame.
Tim: So, men and women speaking, boys and girls speaking, is that something that the AI just detects? Or when you’re setting up a manga for translation, do you say, okay, this character is a 15-year-old girl and this character is a 9-year-old boy? How much hinting do you have to give it?
Sho: No hint, they will be automatically detected by the image recognizer.
Tim: Okay. So the image recognizes this is a teenage girl or this is a elementary school boy.
Sho: Not like elementary school boy but just like boy or the girl with long hair.
Tim: Would be AI track those characters through the entire story in the manga? So the girl with long hair is speaking the same way and the middle-aged salary man is speaking the same way through the whole story?
Sho: Yeah. That’s a super nice idea. But currently, our system only track the context inside a single frame.
Tim: All right. Getting back to accuracy, AI is fascinating. It’s very difficult to intuitively know what will be easy for an AI and what will be difficult for AI, right? I would imagine the AI is probably pretty good at guessing the ordering and extracting the speech bubbles.
Sho: Right. It’s very accurate to extract text and the context. Regarding the translation accuracy, 20% of the pages are almost perfectly accurate. And if speech bubble level, around 30% to 50% of the speech bubbles can be translated accurately.
Tim: So what is the main metric you use? How do you measure success? What’s your metric to tell your customers this is X percent accurate? How do you measure that?
Sho: The easiest way is how many of the speech bubbles can be used as they are.
Tim: And you’re seeing right now you’re at about 30%?
Sho: Yeah, 30% to 50%. It really depends on the type of the comics you are translating.
Tim: Actually, do you give the AI hints about that at all, saying, hey, this is an action manga, this is a romance manga? Or does the AI just have to figure it out for itself?
Sho: Oh, that’s another interesting question. If you’re translating the action comic, the style should be changed. But the thing is, if you create the domain-specific model, you need to divide the training data into several parts. So there is a tradeoff between the accuracy and the domain-specific.
Tim: Interesting. Let’s talk a bit about the business model, though. So how big is the market for this?
Sho: If we talk about the comic market size, that’s huge, and that’s growing. But if you’re talking about the translation of market size, it’s not as huge as the comic reading. The market size of comic translation is around 10 oku-en.
Tim: Okay, so that’s around 8-9 million dollars.
Sho: There is no huge market right now. But what we are thinking is that if more webtoon is coming to Japan, or if the Japanese publishers are willing to go outside Japan, then the market size will be growing.
Tim: Well, I can appreciate that. I mean, theoretically, it could expand the business. It could open new markets and allow for Japanese manga to be sold in greater quantities overseas. But a $9 million market is pretty small to go after.
Sho: Which is one of the reasons we are doing different business at the same time.
Tim: Okay. Your current business model is you charge between $1000 and $5000 a month for a publisher to use the platform.
Sho: Mm-hmm, mm-hmm.
Tim: So, would your maximum market in Japan be like 20 customers or 30 customers, and then you’re done with the publishers? How would that scale up?
Sho: Yeah. So one thing is to go outside Japan, and another is change the way of selling this product. So, we provide the Mantra Engine to some of our customers if they are a comic platform, they sell comics. So we don’t charge the monthly fee of Mantra Engine, and we ask them some portion of their sales, the translated comics.
Tim: Well, looking into the future, before, we were talking about the goal, of course, is 100% accuracy, 100% of the speech bubbles are translated correctly, doesn’t require any human intervention. Do you think we’ll ever get there? How close can we get to that?
Sho: Okay. That’s my another favorite topic. I think about it every day. But my current conclusion is that it’s almost impossible to create enjoyable content without human translator, at least for 10 years or something like that. You can translate something that is understandable, but it’s not easy to create one which is enjoyable, which is fun. Because there are so many efforts are spent on, for example, you can change the name of the characters, a new name for other language. That’s a very creative thing.
Tim: Yeah. And before, we were talking about how one of the challenges is that so much of the meaning is not in the text, it’s in the illustrations and the graphics. But I think, actually, there’s a level deeper than that. So much of the meaning is not in the drawings or the text. It’s in this shared human experience that we humans have because we live in the world and AI’ don’t. Do you know what I mean?
Sho: Yeah, yeah, you’re right, you’re right. For example, one character, they may say “dattebayo”. So that kind of weird way of speaking, those things are very, you know, it’s not just a text. That is a character made by the author, and it’s so difficult to translate those things.
Tim: It sounds like you had a really good experience at the University of Tokyo, and University of Tokyo is really special in the way they support startups. Well, what was some of the most valuable things you got from Todai, that that you’re like, wow, every university should do this for its students?
Sho: So, there are several classes that teach how to create a company. And those classes are not for the business guys but for the people who can build things.
Tim: So the early stage stuff, this is how you can get on the right track and this is how to set yourself up for success.
Sho: Yeah. Because once you’ve decided what to do, you can find ways to figure it out. But before that, if there is no opportunity to know about the startup or know about what you can do, instead of being researchers or being engineers, for most of the students, there’s no chance to even get to know you have that kind of opportunity.
Tim: I think that’s so important, because there is so much innovation in Japan. But even after all these years, it still surprises me — so I was down at Stanford just a couple of days ago, and every computer science student and most of the engineers are thinking startups. They may start one, they may not, but they’re thinking about it. And in Japan, we’re still not at that level yet.
Sho: Yeah. But it becomes better, because more and more students began startup already and they share knowledge. They share their experience.
Tim: Yeah. Role models are so important. And I think the role models much more so than the Masayoshi Sons or the Elon Musks, just to have someone who’s like three years ahead of you.
Sho: Yeah, exactly. Exactly.
Tim: Who’s like, “This is what I did. You can do what I did.”
Sho: Yeah, yeah. I remember that on the day I decided to try a project with my co-founder, I took the class. And at the class, the guy who was running a startup, four years older than me, he talked about making startups. And he was graduated from the same department that I was so it’s very easy to imagine that, oh, maybe I can do that because we learn the same thing at the same department by the same teachers.
Tim: The gap is just much smaller, right?
Sho: You’re right. You’re right. If Masayoshi Son talk the same thing, I don’t think I can imagine that, right?
Tim: Yeah, yeah. Well, listen, Sho, before I let you go, I want to ask you my magic wand question and that is, if I gave you a magic wand and I told you that you could change one thing about Japan, anything at all, the education system, the way people think about risk, the way people view startup founders, anything at all, to make it better for startups in Japan, what would you change?
Sho: I love that question, by the way. And what I thought was maybe I would change the language education. Because if you can learn and use other languages, you can get more information from the place outside Japan. There are so many opportunities outside Japan and there are several things that we need to learn.
Tim: Let me dig down on that. Because, I mean, I think foreign language education especially English language is super useful. But do you think that people need better access to, for example, information on how to start startups and information on how to program and the latest research? Or do you think it’s more of what’s happening in other markets and like market intelligence, or to mean maybe like a way of thinking? What is it the English language ability would give them?
Sho: I don’t think that’s only about getting the information, but it’s more about understanding other cultures. If you have different cultures, you have different way of thinking, then there might be some other opportunities. And more importantly, it’s fun to understand different cultures. I love that.
Tim: I think so. I mean, that’s what Disrupting Japan is all about, right.
Sho: Yeah. I think that’s very important, especially for the people who are living in a small island, to get to know other cultures and think about the world outside the island and seek the opportunities. I don’t say that Japan is a bad market. I’m saying that it’s nice to have several options.
Tim: I think that’s really important. And it is something that more and more founders and investors and even business leaders are saying that Japan is a great market, and that’s good and bad. I think there needs to be more of a focus on Japanese startups going global and not just looking at the Japanese market.
Tim: I think the English skills really would play into that well like you said.
Sho: Yeah, I think so.
Tim: But I think it’s also getting better year to year. It seems like every year, we do have more new Japanese startups that are talking about and thinking about global markets.
Sho: Yeah. But I’m not sure what changed. Do you have any idea?
Tim: I’m not sure. I think it might just be awareness. I think it might be role models as well. We’ve had a small number of Japanese startups who have been successful overseas. And I think that gives founders the confidence to say, “Hey, I can do that, too”.
Tim: That’s important, to have a role model that’s not that far from you.
Tim: Yeah, exactly, where the gap is not so big, where you can imagine yourself doing what they did. Well, listen, Sho, thank you so much for sitting down with me today. I really, really appreciate it.
Sho: Thank you so much, Tim. It was so fun.
And we’re back.
The manga industry is really interesting. In fact, the whole entertainment industry is really interesting.
But as a startup founder, that’s kind of a problem.
You see, the more fun or glamorous a business seems to be, the harder it is to make money in it. Fun jobs attract lots of competition willing to work for free or for exposure, and that makes it harder to turn a profit. However, there are two ways to succeed in a glamorous industry.
First, is to do something important that no one in the industry wants to do — improving the supply chain, streamlining, accounting, or perhaps automating translation. You focus on a non-glamorous niche inside a glamorous industry.
The second way is to disrupt it. You introduce something that will upend the entire industry or define a new category. You can make the case that anime itself followed this path in regards to the movie industry. The disruptive path, however, is not glamorous, at least not at first. No one will take you seriously. People in the so-called “real industry” will look down on you. And you will probably find your initial customers outside the mainstream industry. That is, until you eventually change the entire industry.
The disruptive path is high-risk, and it really only works for true disruptors like cloud computing or digital cameras or hydraulic excavators or perhaps manga and anime themselves.
Mantra is taking the first path. And at this point, their success in their current business model depends not so much on the quality of their product but whether the industry they serve can grow large enough to support them.
And at the moment, that’s kind of questionable. But the product is good. The team is smart, passionate, and already experimenting with alternative revenue models. I predict that when we check back in with Sho in the team in a few years, that either Mantra or the global manga industry is going to look very different.
If you want to talk more about manga or disruption or the disruption of manga, Sho and I would love to hear from you. So come by disrupting japan.com/show193 and let’s talk about it. If you leave a comment, I guarantee, Sho or I or maybe both will respond. And hey, if you enjoy the show, give us a review on iTunes or your podcast platform of choice, or just tell a friend or two about it.
But most of all, thanks for listening and thank you for letting people interested in Japanese startups know about the show.
I’m Tim Romero, and thanks for listening to Disrupting Japan.