Bill Kerr: The internet has a language problem. The narrow range of languages in which useful content is consistently available limits its utility. And while machine translation has advanced significantly in the past decade, there’s a long way to go. Spurred by the coronavirus pandemic, businesses and governments are adopting new AI-powered neural machine translation technology to expand their reach and boost productivity and efficiency. Yet, even as the technology improves and business models change, human translators aren’t going anywhere anytime soon. The US Bureau of Labor Statistics forecasts a robust 20 percent increase in translator and interpreter jobs between 2019 and 2029. This reflects increased globalization and increased linguistic diversity in the US. Translation is shaping up as a growing category of AI-augmented jobs. And while it’s suited for some uses, completely automated translation falls short of the mark for many more. This is especially true in creativity, originality, and when cultural nuance comes into play.

Welcome to the Managing the Future of Work podcast from Harvard Business School. I’m your host, Bill Kerr. I’m joined today by computer scientist Spence Green, CEO and co-founder of Lilt, an enterprise language translation company whose services combine machine translation and a network of freelance translators. We’ll discuss Lilt’s core technology, the future of translation work, and how companies can integrate localization services into their business models. We’ll also talk about the prospects for reliable and widely available United Nation’s–style simultaneous translation feeds and their implications for the workplace, as well as the ways Covid-19 has highlighted machine translation’s importance in e-commerce, enterprise systems, and governments. Welcome to the podcast, Spence.

Spence Green: Hey, Bill. Thanks for having me.

Kerr: Spence, why don’t we start with a little bit of background about yourself and how you came to co-found Lilt.

Green: I’ve been working with computers and programming since I was very young, and after college, I worked in defense. And for part of that time, I ended up going to the Middle East—I was working on an airbase—and my personal goal at that time was to learn Arabic. This was not long after 9/11, and I thought this was a part of the world that would be important for me to understand during my lifetime. As part of that process, two things happened that were really important. The first was, I learned that, in that part of the world, people who don’t speak English make less money and don’t have the access to opportunity that you and I have. And the second thing that happened was, Google Translate came out in 2006. So I got excited about this technology that I thought was interesting from a computational point of view, and then had this potential to impact people who didn’t have the same access to opportunity. So I left that job and went to grad school at Stanford in 2008, started working on machine translation research. And then a few years later, I was at Google working as a research intern, where I met my now co-founder. And for the last 10 years, we’ve been building these systems to do precisely that: to make more of the world’s information available to everyone.

Kerr: That’s quite the journey, beginning with the recognized need in the Middle East and then coming over to the technology side. Could you tell us a bit about your core technology? And how does artificial intelligence and machine translation improve upon what humans do? And what technical advances are at the frontier right now?

Green: In the case of machine translation, this research started in the late ’40s, right after digital computers were invented during World War II for two purposes: bomb making and cryptography. And the original idea was, “Well, Russian is just encrypted English, and if we could just decrypt it, then we could do translation.” That was literally the original concept. The modern computational approach to machine translation, which is using machine learning, that dates to the late ’80s, actually. So that sort of approach was current up until 2016 or actually 2015, when people started applying neural network approaches to machine translation. And that currently is the state-of-the-art. So our system, it’s a neural network system. And the two things that it does that are unique, which we started working on 10 years ago, are 1) a system that learns as it’s used, so it can correct its own mistakes, given feedback, and 2) a system that is interactive, so it supports human interaction in the form of predictive typing. So while I’m translating something, it’s going to be reacting and predicting words and phrases while I work. Translation you can think of as bilingual writing, and almost every organization of a certain size does it in some capacity. We have two relatively different businesses that share a common technology. And the technology is common because, as I said, translation is just bilingual writing. And you do it in a lot of different places. And what differs is what you’re translating. And that’s the key bit. So in the private sector, our business is a managed service, which enables companies to do global customer experience—meaning they can offer all products and services with the same experience that you and I have with English. And, actually, in some sense, the internet is broken in this way, in that you go to almost any company’s website, and if you change it into Chinese, let’s say, and you click three links, and you get kicked back into English. That’s true for almost all companies right now. And so we try to fix that problem. In the public sector, the problem is the United States government or any government, really, is monitoring parts of the world that it cares about. And it just doesn’t have enough people to process all the information that’s being created in the world. So if you can use machines to scale the human resources that you do have and use technology to process the rest, that’s really valuable to government.

Kerr: Going back to your wonderful analogy of Russian being encrypted English, for that Chinese website, or for translation into Spanish or many languages around the world, Arabic, how complicated is it? You can imagine changing the banner, three or four words, but then you’ve got probably many levels of detail, in terms of how the webpage was meant to be displayed and the structure behind it. How far can it take you down the journey of transforming what was originally meant for English audiences into something that is ready for Arabic?

Green: There’s really two parts of that problem. The first is organizational recognition, that language is important. So consider, Bill, if you bought a Toyota, and the instrument panel and the LCD screen and everything associated with the car was in Japanese, and you had no way to change that, you’re not feeling like the most valued customer at that point. But very few businesses recognize that, because primarily, they operate in English, and they don’t have a more global point of view. But as soon as you recognize that, that’s the hardest part. Then the second part is operationalizing what your quality standard is. Because if you don’t have one, then you should just use Google Translate. Just throw everything through Google Translate; problem solved. But that doesn’t work for any business, because all businesses have some kind of identity that they want to project. And operationalizing that so that it’s globally consistent across all languages and is relevant in all cultures is actually quite difficult. Once you operationalize that, then you can put it through a production process like ours that’s going to make that possible to realize in all languages.

Kerr: So tell us a little bit about, among the types of content, what’s the hardest to translate? And then, how do you work with customers to think about what should be done with the AI backbone versus have human translators involved?

Green: It’s a little bit deceptive, what’s hardest. But if you think about it a little bit, it becomes obvious. So many people think that the hardest would be highly technical content—there’s jargon and language of a bespoke nature. But it turns out that technical content is usually written according to a convention. So think of an extreme case, like a patent. It’s highly routinized, the way that these patents are written. So it’s not very hard for a machine to learn that and start to mimic it. What’s much more difficult is creative content—things that require common-sense reasoning or human ingenuity. So that would be like translating a novel or translating a movie, let’s say, that has a bunch of colloquial culture references and making that relevant in another language. It’s very difficult for a machine to do. So I would say that’s the really hard one—anywhere that a significant amount of world and cultural knowledge is required—because the machine doesn’t have access to that information. So I think that’s probably the hard spot. Then, in terms of where the technology comes in, we think of it as anywhere that content is being authored, you would want to make it possible to publish simultaneously, in all the languages that you care about. So that’s a combination of technology first to connect to and integrate, wherever words are, they’re being produced—whether it’s in a word processor on a website or in a forum or in a code repository for creating an app. And then, having the operational and production capabilities to route that, produce it, run a quality check through it, and put it back where it’s supposed to belong, in an efficient, scalable, and affordable manner.

Kerr: Who owns the data? Who owns the knowledge that gets developed in this learning process that’s happening with Lilt?

Green: As you can imagine, our customers are pretty sensitive to how their linguistic data is used. This process is the one place in most companies that acts as a convergence point. So everything that’s being written by product marketing, legal, sales, wherever, it’s all going through this kind of funnel. And so there’s a lot of very sensitive information. So, at least in our business, we make strong representations that we don’t use that data across customers in any way. And in the government case, which is sort of at the more extreme end, our system deploys in an air-gapped environment, where they control the entire installation of the system and how it’s used.

Kerr: Let’s turn now to the human side. You work with a lot of translators. Tell us a little bit about their primary demographics. Are they full time, part time? And then, how do you go about recruiting them into your business?

Green: Professional translation is, in some sense, it’s like one of the original gig economies, I guess. It used to be the case in the ’80s, that businesses would employ translators directly. And then in the ‘90s, as the internet became more available, this function was increasingly outsourced. So today, over 80 percent of the professional translation workforce is outsourced. And it’s distributed globally, because the convention is that you translate into your native language. So I’m an Arabic speaker, but it would be very, very difficult for me to translate from English—even something that I know really well, like a computer science text or something, would be very challenging for me to translate from a language that’s not my own. So we tend to source and hire them in the geographies where those linguistic communities exist, so they’re all over the world. It’s a skilled labor force. So the vast majority of them, I think over 80 percent of our translators, have college degrees, and a high proportion of them have graduate degrees, as well. And it also tends to be a majority female profession.

Kerr: And how has technology been changing their jobs, starting to dial it back a little bit to, say, 2000—before your technologies and Google Translate and so forth got off the ground? And then, if you look ahead for another 5 or 10 years, what do you anticipate being the changes in tasks that translators will find?

Green: It’s interesting. The technology development, really since the ’70s, has all been driven by machine translation. And so, in the ’70s—well, actually in the ’60s, when the first MT systems started to be used—it was found that they produced so many errors that correcting the output took longer than just translating from scratch. So as a response to that, some translators in the EU in the late ’70s suggested, “Well, why don’t you just create a database of all of our past translation, and when we translate something new, you could just look up in the database if we’d done it before and just give us that so we don’t have to do everything from scratch.” So it seemed like a pretty good idea. So that was called “translation memory.” And that was a response to the state of machine translation. And that was the dominant technology through—well actually, even today—usually the core of the enterprise localization program is just this database. And it’s really only been in the last couple of years that machine translation has gotten good enough that what I said before is not entirely true—that it takes longer to correct it than to do it from scratch. What that mode does—which is, give me a machine output and then I’ll fix it—has two actually really negative effects. The first is, it turns what is a pretty interesting and cognitively active task into a tedious clean-up crew task that these skilled people don’t actually like to do very much. So it renders the job pretty unpleasant. And the second thing that it does, it shortchanges the machine, too, because the machine has no capacity to learn from what the people did, if they’re just correcting Google Translate output. So I would say that, recently, the effect of machine translation on this job has been pretty negative. And what our work for the last 10 years has been about was, “Well, let’s just use our brains a little bit and think if we can go beyond this simple cascade of correcting machine output and coming up with a more engaging and efficient mode of interaction with the machine.”

Kerr: If I’m not mistaken, the early technology behind Google Translate was documents that, for government purposes, were written in English and French or in multiple languages and pulling the text from that.

Green: Yes, that’s true. The machine learning–based machine translation systems, which as I mentioned date to the late ’80s, they’re all trained on what we call “parallel text,” which is a sentence in English and then it’s translation in French. So think like the Rosetta Stone. So the technology is completely language independent—it learns a model with that as input. So the natural place to find this text is in government proceedings. The Government of Canada publishes everything in English and French; the United Nations publishes everything in its six member languages. And indeed, those were the original data sources for researchers building these types of systems.

Kerr: Localization—that’s a word that is very important in your business. So maybe you can start by defining more precisely for us, what does localization mean for you? And then tell us a little bit more about how that has been built into the operations of Lilt and your staffing models and so forth.

Green: Well, localization is a kind of ill-defined term that is used interchangeably with translation, usually. But I think that the actual definition of it is that it takes in more local context than just linguistic issues. So it could be design items. For example, we have a customer that they create these designed templates that get used for invitations and flyers and posters and things like that. So if you have an invitation that’s like, “Let’s get drinks at 5:00 PM,” and you’re going to translate this into being used in Saudi Arabia, well, that culturally doesn’t really work that well, so you need to come up with the alternative that’s more culturally acceptable. So that’s the broadly encompassing term for localization. But technology, specifically, applies to the act of converting a sentence in one language into a sentence in another. And all of that cultural context is where the human comes in, because it’s not presently known how one would codify that drinking is not customary in a Muslim country, and the system should incorporate this into its model when it’s making a prediction.

Kerr: Seems like this is a tricky design question that maybe you have to work with clients on, as to whether you integrate this localization with the primary functions of the organization. Or do you treat it as some kind of add-on downstream, right before it goes live, then you’re trying to make the adjustments or the corrections?

Green: Typically, in businesses, it’s at the very end; it almost is an afterthought, it’s at the very end of the value chain. And this is not great because changes could be made more upstream that would make it much easier to produce content in other languages. So for example, if the source language isn’t written very well, or it isn’t quality checked in the original production step, this can have pretty negative effects downstream, which is why we’ve tried to build our technology directly into where the original content offering is happening. And if you can intervene there, then you can make the whole process much more efficient downstream.

Kerr: Spence, I was recently on a video call conference with China, and we were attempting to use some software that was live translation of Chinese-into-English text, and my voice-into-Chinese text, back and forth. And it was clear that we weren’t quite at that United Nations level of simultaneous translation.

Green: Right.

Kerr: So is that in the future? Is it something that we can accomplish? Could it ever be completely automated to do that? Or will there always need to be human intervention and touch when you’re trying to do something as sophisticated?

Green: Yeah. I mean, I think it’s obvious and inevitable that some of the issues associated with the current implementation of simultaneous interpretation will be fixed. One is the latency issue—just as computers get faster and machines get faster, it will be practically instantaneous. And I think it will also help when in the future there’s current work on building more-sophisticated speech translation systems that train directly on the audio data, rather than going through transcription to text and then going through a machine translation system like ours, and then going through synthesis. So that whole pipeline, which is the way these systems work right now, it’s slow. There’s a lot of computation there, and then there can be cascading errors through that pipeline. I think that will all obviously get fixed. The key question really goes back to the quality guarantee. If you’re having a conversation with your friend in Chinese, and you’re using technology like that, wonderful. If you’re negotiating a nuclear agreement, then probably you want some sort of quality guarantee. And then, again, how do we get that in a machine-oriented workflow?

Kerr: So you’re clearly a technologist. I think all of our listeners will have picked up on that. So as you think about you and your co-founder, John DeNero, balancing both research and running the business, how as an entrepreneur, have you struck that balance?

Green: Yeah, I think we’ve had kind of the same working partnership since the beginning, which is, John is chief scientist, and he’s actually on the CS [computer science] faculty at Berkeley. And he manages our research team and our research group. This is somewhat uncommon for an early-stage startup, but we have a full research team, that they spend about 50 percent of their time publishing in the academic community and then also working on our production system. So John works on that. And then I run the business. And that’s been our arrangement since the beginning, which has meant I’ve been getting a PhD in a different topic for the last six years than the one I originally got.

Kerr: Well, then, this is the right question to tee up next. There’s a lot of interest in this area. Microsoft recently acquired Nuance. How secure is the business model of the future? And throw in there Covid-19. How has the pandemic and so forth changed where Lilt is arcing toward?

Green: I think the one way to think about it is that, for a specific enterprise problem like this, you need not just the technology, but you need a service that goes around it. Because the key problem is actually the change management and catalyzing and supporting enterprise change from the traditional way of doing this, which is just having people type to an automated workflow. And the tech companies usually don’t do business like that, where you need account managers and customer success people and product managers and enterprise salespeople, and all of that that goes around the technology. So I think there are lots of opportunities to solve problems that are not just sort of horizontal, cloud-based technology. I think it’s very challenging to build a business in that regard. So I think that still remains the case in our field because so much human labor, we’ve automated a large fraction of it, but there’s still a lot of meaningful work that’s done by people. And that doesn’t tend to be the type of business that the tech companies go into. And your second question, about how the work of the translator changes, I mean, I think that’s really just the history of technology, that we build tools to routinize tedious work, to free ourselves to go do more value-added work. And just generally, we, as people, have a very difficult time imagining the jobs of the future. If you had gone back 100 years or something, and you would talk to somebody and be like, “In the future, there’s going to be this job, that’s occupational therapist,” nobody would have a clue what that job is. And so I think that’s generally, the history of technology and innovation.

Kerr: So Spence, as you think about this evolving job with the translator, one thing we spend a lot of time on this podcast working through is, what type of institutions are necessary to prepare the workforce for new technologies and to keep them in front of this augmentation curve?

Green: Almost every job now is in some sense a software job. And facility with software and knowledge of programming is essential and vital. And I think, historically, the education that translators get is very much focused on the act of translation, independent of the technology that’s used to do that work. And that’s not actually how it works in the enterprise. The actual job, in practice, has very much to do with using technology to achieve an outcome for a business. So I think that’s really important. The second point that I will make is that, I think that there are a bunch of jobs around translation that are really interesting, that are being made possible by the technology. So, for example, one of our customers is a financial information company. They have had a team of internal translators for a long time for security reasons. And when they introduced our technology, the efficiency that they got meant that some of the people on the internal team had time for different things. And some of them just started learning Python and becoming corpus linguists, basically, where they’re now writing code to clean up and collect data to train the system to make the whole workflow more efficient. Now, that would never have been possible if they were spending all their time translating English sentences to German. And so now you have people who are still doing language work, but they’re doing software work that’s really high value, and it’s right at the edge of new technology that’s being created in the enterprise in developing those skills. And I think that’s just really exciting. And that really should be taught in schools, as well.

Kerr: In one of our other recent podcast recordings, we had a discussion about the National Security Commission’s new report on artificial intelligence, which is getting a lot of private-sector, public-sector interest and so forth. What’s your take on it?

Green: Yeah. I think this is a matter of utmost importance. And we submitted a position paper to the commission, and it’s obviously an important part of our business. An important thing to note is that artificial intelligence research, machine-learning research, it happens out in the open. Most of the manuscripts are published on archives. The ideas are discussed in international conferences, and this is a really, really good thing because it means that the pace of progress is very, very fast. It also means that everybody has access to this information. And the pace at which different countries are able to use these ideas and implement them to protect, to ensure the primacy of national security, is of the utmost importance. And I think if we’re bogged down—politically and bureaucratically—because it’s very difficult for new innovative companies to do business with the government and to provide their technologies to civil servants and to active personnel in the military, then we can’t possibly keep up with other countries that are making it more of a priority to get these new technologies, international security applications. And that was the motivation for the commission and for the report. And I think it’s just really important that we, as a country, decide that we need to expand the national defense base and make it possible for more companies who have technologies to offer that are not being built by the large historical defense contractors and make it possible for us to do business with the government, too.

Kerr: Spence, I have one final question, which relates to you being an entrepreneur. We’ve talked a lot about the technology side and the future of translation. If you had one parting piece of advice for new entrepreneurs trying to get off the ground, run a technology business, all the things you just talked about, is there something you reflect upon on your journey that was really important, really vital to share?

Green: I think there are two things that I’ll say. The first is, I think it’s really important, if you are going to start a business, and your objective is to build an enduring business, then it’s just so important to understand who the customer is, and more crucially, how they buy. So I think when you come at this from technology, you think, “If we develop this technology that’s obviously useful, and it has a big market, and it’s something that businesses need, then that’s all great.” But there’s this really important detail, which is, customers are people, and they operate within systems that adopt and procure things in certain ways. It’s just vitally important to understand that. The government is the most extreme case, where it’s actually quite difficult for people within government to procure things, even if they want them. So understanding that is really important. And the second point that I would make is that ideas are pretty cheap and in generous supply. And I think it really is the case that a lot of business is execution and just a lot of persistence. So I think that, if you’re going into business, you have a new idea or a new technology that works, that’s really great, and a lot of people should do that. But then you also have to stick with the operational grind part, which actually is the real challenge in building a business and then having impact on the world.

Kerr: Spence Green is the CEO and co-founder of Lilt. Spence, thanks so much for joining us today.

Green: Thank you, Bill.

Kerr: We hope you enjoy the Managing the Future of Work podcast. If you haven’t already, please subscribe and rate the show wherever you get your podcasts. You can find out more about the Managing the Future of Work Project at our website hbs.edu/managingthefutureofwork. While you’re there, sign up for our newsletter.