Vasco outlines Unbabel’s translation pipeline and underlying technology, which includes MT engines, quality estimation capabilities, and AI-powered translation productivity tools. He describes the importance of the human-in-the-loop model and the ability humans have to impact the output of machine learning (ML) models.
The CEO talks about the challenges of ‘going remote’ overnight in response to Covid, and how Unbabel’s hubs policy will help them preserve the company culture in a hybrid-working future. Vasco says culture plays an important role in attracting top talent globally in the highly-competitive AI / ML space.
Vasco also shares his experience with investors, such as Point72’s Sri Chandrasekar, lead investor in Unbabel’s Series C, who provide actionable insights on how to further scale the business.
First up, Florian and Esther discuss the language industry news of the week — and the launch of the flagship Slator 2021 Language Industry Market Report. The duo share highlights from the 80-page, newly-released Market Report, which features a wealth of insights and data — on market size (by vertical, region, and intention), market dynamics, the supplier landscape, technology and investment trends, and more.
They also tackle over a week’s worth of M&A, as Florian unpacks Big Language Solutions’ acquisition of US-based interpreting provider Language Link, and discusses the backstory to Big CEO Jeff Brink’s “tenacity, honesty, and desire” to close the deal after Language Link CEO and owner Jeff Barger pressed paused in early April.
Esther talks about translation and interpreting provider Propio Language Services’s acquisition of Vocalink — also in the US — while in Germany, she highlights GEtraNet’s acquisition of Lingua-World. Florian closes by reviewing AI transcription agency Verbit’s acquisition of captioning provider VITAC.
Stream Slator webinars, workshops, and conferences on the Slator Video-on-Demand channel.
Florian: First up, tell us about Unbabel, what is the elevator pitch?
Vasco: There is a very simple insight, which is that everyone in the world speaks a language. Most people speak different languages and not everyone speaks the same language or we all speak all languages so this fundamental problem of different languages is going to continue to be a thing. The problem is that as a company, as you scale and you start selling your goods and services in other markets, you start facing issues that come from the fact that people speak different languages and those issues occur across the entire company. What we do at Unbabel is we are building a language operations platform that at the base combines artificial intelligence and human translation to enable companies to scale across languages throughout the different areas of the company.
Florian: Can you break the tech down a little bit for the non-tech people among our audience? What is at the core of Unbabel in terms of the technology?
Vasco: We started Unbabel in 2013 and a lot of times timing is a bit about luck and in our case, it was right at the edge when artificial intelligence was starting to be useful for translation. I think if you started a few years before and you put the output of a machine translation engine in front of a translator, the first thing they would do is erase everything. Right around that time is when machine translation started being useful enough. One of the core insights was until now solving language issues was primarily a human effort. Now that AI is starting to have an impact, how does the future look like from an AI-first perspective?
If you think about a translation pipeline, there were some very obvious first places to apply AI. One, machine translation extends throughout all of the different functions. What I mean is at the core of Unbabel, there is a transmission pipeline. This pipeline works in a simplified manner as follows. First, a machine translation engine is built by us, is customized, and customer adapted. As it is continuously evolving it is deep learning based. Then after that, there is a piece of technology that we have built over the years, quality estimation and that is a neural network that tries to make a real-time decision on whether the output of a machine correlation engine is good, or if we need human intervention.
Then if the decision is that we need human intervention then it gets sent through a smart task routing algorithm to our community of translators and in turn, they work on AI-powered CAT tools that we have built. That incorporates not just the typical grammar and spelling, but how to incorporate glossaries inside, how to decide which sentences are actually quite good and which are not so you can focus on them, and other aids that make the translator’s life easier. Then at the end, that is the output that gets sent to the customer and then the output of that also gets used to retrain all of the AI components. There are a few added metadata layers that happen after the fact, like annotations and other things to understand it and rate quality. Then the combination of that is what feeds the different AI components.
Esther: What would you say are the top three challenges you are helping customers to serve right now? What kind of functions and industries is Unbabel currently serving best?
Vasco: Part of the vision of Unbabel is based on this insight that one of the things that AI is going to do is enable the function of localization to increase in strategic value across the company. Localization is primarily focused on two of the sub-areas of a company, typically marketing and product. As a company scales, you have issues in different areas like how do I serve my customers in multiple languages? How do I sell my product in multiple languages? How do I enable internal communication? Those are all things that every global organization feels. Until now dealing with the solutions for all of the problems were ad hoc and a bit siloed so if you went to customer service, you would say, the only way to solve this is to hire natives. Or if you are talking about sales, you not only have to hire natives, but you would probably have to hire people that are in location.
For us, we felt that the right point to start was actually customer service and there were a few reasons for that. From a technology perspective, being able to be in a disruption market that typically localization could not handle and machine translation also could not handle and so there was a nice space there. What we do for the most part right now, as a first use case is enabling companies to provide multilingual customer support using translation as a layer. Typically we are focused on text-based, so email and chat. Let us say that you have a team of customer service agents in Germany serving the German market and sometimes not only is it very costly because the labor in Germany is expensive, but also a lot of times customer service is not seen as a very rewarding career, you have a lot of turnover.
There are all of these challenges and so what Unbabel enables you to do is serve this market from really anywhere, having a contact center in Romania or Poland, or maybe the Philippines. It does not matter. You can actually create centers of excellence and we add Unbabel on top and now every one of those agents can support customers in 30 different languages. The impact of this is quite interesting because not only do you have what you would expect in terms of cost reduction, but we see an increase in customer satisfaction. This comes from when you can use your workforce better so you have agents that can handle everything. It typically means you have a much faster time to first response.
For example, during Covid-19 Logitech grew quite a bit. Everyone is buying webcams for zoom and so they suddenly have an explosion on customer service requests and in their case, a lot of times they end up hiring engineers for their special tier two customer service to solve problems. Hiring a bunch of engineers to deal with 300% growth on customer service requests in the time of Covid-19 all around the globe was pretty much impossible and so with Unbabel they could scale linearly their customer service operations and reduce from 48 hours average first response to 12 hours because they could use all their agents, all the time. That ends up being a very typical use case for us.
Florian: Are the direct users of this usually your target segment, or is it big contact centers that are dealing with the enterprise?
Vasco: For the most part, the relationships end up being direct with customers, but BPOs which typically have the large contact centers are partners so they are definitely very much involved in the process because they end up representing about 80% of all customer service that goes through some agent that is in a contact center somewhere. They can also very much optimize a lot of the flow. It is very interesting, for example, for somebody in the Philippines, suddenly they increase in value. Until now maybe they can only do English customer support, but now they can do 30 languages right from there and so that means an actual impact on the economy.
Esther: Help us to understand the journey a bit more. What was the trajectory like to where the company is now?
Vasco: A lot of times in the case of startups it is easy to have hindsight and create this very smooth curve like everything was so neatly planned from the beginning. Paul Grant says something that I find very true, which is all startups are shit shows internally. It is a bit like that. You start out and you have a problem that you want to solve, and there is a lot you do not know so you go out and you start doing things and hopefully, you converge on a business model and product that gives you that beachhead. For us, that was definitely the beginning.
In the beginning, we just wanted to prove that this whole idea of a hybrid model of AI plus human works. It took us probably two or three years to really build out the core components, the translation pipeline, the machine translation. In the beginning, we used third-party tech for some of the AI components because we do not have the bandwidth or resources to implement everything. Then we did our Series A in 2016 and that was a change for us where we went from trying to figure out what is that initial use case to identifying that customer service seems to be that beachhead. Let us go out and build a sales team that enables this. Let us go out and continue to optimize and implement some of the core AI tech that we knew that we needed to develop.
Once you find that initial use case, it is much easier to then build for it. Then we did our Series B in 2018 and then Series C in 2019 and I think it is the phase that we are in. We have identified that initial use case, we are becoming more and more dominant in it but now it is time to start taking the platform and this whole concept of language operations to the next level. We have proven that we can bring benefits to customer service. Now, what happens is we expand to other use cases within an organization.
Florian: In terms of the leadership team, João is the CTO, but who else is there and what are the roles?
Vasco: The leadership team evolved quite a bit over the years. At the pace that we grow, sometimes the people right at the beginning are not the right people once you reach a scale. You are constantly improving and growing and you need to bring everybody along or find people that have the skillset to take you to the next level. We have a great leadership team with a lot of experience. We have Wolfgang Allisat as our CRO, he was the first employee Europe wise of Omniture, all his six previous startups were either sold or IPO’d, so has great experience with SaaS revenue. We have Sophie based in San Francisco, our CMO that joined last year. Her experience is a lot on customer service and taking platforms into deep customer service installations and how you market it.
We have Leah as our VP of people so culture is super important to us, especially where there are such magnitude shifts in tech company’s culture. How we deal with the current environment has been very crucial. We have James as our CFO based in London. James had been with MessageLabs and helped them with the acquisition, he has a lot of M&A experience and helps high growth companies. Then we have Sebastien as our VP Product. He came from onefinestay and Ocado, an amazing product guy from a systems perspective. Then me as a CEO and then Ana as Chief of Staff who vastly extends my ability to do things which has been amazing.
Florian: How was it going remote? What are the pros and cons of running a distributed team? If you can share what tools you are using to do that and your experience over the past year?
Vasco: I have to say we were not built as a remote culture, even from the beginning, we officially decided to start the company during a surf trip. That has been part of our DNA, even to this day on Thursday mornings in Lisbon at the beach, there is a lesson, instructors, boards suits, everybody at Unbabel is welcome. It is something we continue to do. We do a lot of events, retreats. It is a very heavy culture of spending time together. We had to, as most other tech companies switch to remote overnight. From a productivity perspective that was not an issue so things continue to function quite well. What we are seeing now is we did lose a bit of our culture and it is an important part and it is something that we need to get back.
I think there is an important part of our culture that has to do with us spending time together post-Covid-19. What will Unbabel look like from a culture perspective? Are we going to be a hybrid model where people are going to spend part of the time in the office or are we going to continue to be a remote company? My gut feeling is that we are going to be a hybrid model, at least in the beginning. Those interactions and serendipity that happens in the office are very important for us. I do not think that we have really been able to move the culture to a remote first experience. I think some companies are because they were built remote first, but that is not the case with us.
For us, we have been using the standard things, heavy zoom, communicating over Slack, over email, having things documented, trying to move to an asynchronous way of working. Some of that is very helpful and very valuable. We are not going to be able to go back to things as they were before, but what does the future look like? What we have done is we have embraced what we are calling a hubs policy so we have officially declared hubs, some of the offices that we had already, so Lisbon, San Francisco and New York, Pittsburgh, but we also opened up a hub in London and Berlin.
What this means is every hiring manager can hire for a position that they are hiring at any of the hubs. If you want to hire outside of the hubs, then we need to think about it carefully because if we now hire let us say 40 people randomly across the world, then effectively we are committing to a remote culture. We want to keep it within certain locations and then I think what will happen is eventually, we will all have offices and they will serve as focal points for people to have moments together as a team but it does not mean that those are the people they are going to be working with so your team might be spread around hubs and so the day to day experience might be more remote than they were before.
There are also other people that like the separation of office and private space, that like to be able to see people in an office and have a space they go to. Now people like that flexibility of not having to be there every day, but to be there maybe two, three times a week where they can take a break. I noticed that, at least in our organization, people with kids are more likely now to want to come to the office, because they need that separation. We have been doing polls and surveys at Unbabel and people started remote and now they are starting to migrate to two, three days a week to have a place to go to.
Esther: How have you found it trying to hire and retain these hyper-competitive roles, machine learning and computer science?
Vasco: It has been a big shift over the last few years. For example, we benefited from this concept that Portugal was a bit of an undiscovered jewel where you had a lot of great technical talent but it had not been explored much seven years ago. In the last seven years, that certainly changed. We are going after the top 1% of talent globally and in AI, that means competing with Facebook, Google, DeepMind and other companies. Especially now with Covid-19, physical barriers are much lower and so all our people have multiple offers from other companies. On one hand, this creates pressure to provide a basic set of conditions from a package perspective that you just need to be competitive. Then on top of that, you need to find the people that fit with our culture and that believe in the mission.
Some people prefer the smaller teams. Some people prefer the kind of work that we do and the challenge that we solve. Within AI, there are a myriad of challenges to be solved. Some people are very much vision focused or they are processes driven. If you want to work in machine translation, in AI applied to quality estimation, there are not that many places in the world that are doing proper research, that are going all the way from high-end research to very core applications, to having their work going from research to customers in three or four weeks. That environment appeals to a certain set of people.
Esther: Tell us a bit more about the startup environment. What is it like in Portugal? How would you say the differences are compared to the US and Europe generally?
Vasco: Overall I am very bullish when it comes to Lisbon. I have this whole thesis where there is macro Europe versus the US and the US has a sense of belief that is sometimes harder to find in Europe. It is changing slowly because you are starting to have unicorns in a strong startup ecosystem but it takes a while. In the West, it is almost like there is a stronger reality distortion field, of the ability to just believe that things are gonna work out and are going to happen and that creates an environment that becomes very conducive to risk and innovation.
If you look specifically at San Francisco versus other places in the world, for example, when it comes to Europe and the US, I would compare London and New York. What you see there are two cities that are amazing, that are thriving but they will never be startup cities because there are a lot of other competing industries. There is finance, there is fashion, there are a lot of other things and so the city does not focus all of those resources and attention to tech focus on a bunch of things. San Francisco and the Bay area has that. There is this insane critical mass of people that have decided that this is what we are going to care about. The benefit of that has been accumulated over the years and so it feels highly energized. I compare San Francisco, in my imagination, to Florence in the Renaissance. People move all over the world there to be part of the early conversations. A lot of the meaningful conversations on automation, AI, social experiments are happening there. Then those conversations percolate and maybe a few months later, you are having them again in other places.
When it comes to Lisbon, obviously it is a much smaller environment. The one benefit is we do not have competing industries that much and so the city embraced tech over the last few years and that created a very positive feedback loop. We have good technical talent, the weather is great, you can have a good life here. San Francisco is on a different level and it is easier to become a big fish in a small pond in Lisbon, and you need to not have that mentality from day zero. One of the great things about San Francisco is it puts you in a position of discomfort more often. It constantly reminds you how small of a fish you are and how big the pond really is and in a way that is more uncomfortable, but on the other hand, it propels you to bigger limits.
Florian: What is your take on a lot of people going to Austin and to Miami over the past six months?
Vasco: I do not think Miami is going to be a start-up city. Austin, maybe it is. On one end you have companies like Coinbase that are leaving San Francisco but others that are very committed to the Bay Area in San Francisco. Google’s continued to invest 14 billion in complexes around the Bay Area so there is going to be a shift. I have not seen yet an ecosystem that can compete with the Bay Area, but it could be that in 10 years people leave. I do not believe people are going to abandon cities now. There has been a big migration towards cities. Those reasons are still there. You have the ability to have serendipity in a city with a generation of ideas, collaboration, access to culture, access to entertainment, all of that.
Even from a carbon footprint perspective, cities are much more efficient. The idea of not having to use a car all the time. Those are things that continue to attract people towards cities. Now, I do think San Francisco in particular has a bit of a problem, it is a bit jarring when you are in San Francisco, the homeless situation, for example. It is one of the richest cities in the world, why is this happening? The lack of ability to really make advances might have more of an impact than Covid-19 because ultimately I think what people are going to want is a city, but in a place that they feel safe, that they enjoy, that they can have a vibrant, good life.
Florian: We hear a lot about the human-in-the-loop model and Unbabel is at the forefront there. How do linguists interact with MT in your system?
Vasco: As with language, generalization is hard and what we see is dependent on content types. If you go up the chain of how many people are going to read that content and the amount of time that someone spends creating that content the more you need human intervention to be able to do a proper translation. There is another dimension of language pair, so AI is much more developed in English to Spanish than it is English to Turkish, for example, and if you go Chinese to Turkish, then you need to go through English a lot of times, which adds a lot more complexity and errors so that is the other dimension. Language pair, content type, those tend to create very different experiences and the impact that AI has and by relation the ability for humans to have an impact in the output of AI.
In the case of chat, the linguists are behind the scenes, constantly improving the engine, but actually in production is the engine that is being used. Humans are not in the loop there in the sense of having to check with humans before delivering because in chat you need real-time and so there are constraints there, but AI has gone to the point where if you have a continuous adaptation process, then you can deal with a lot of the chat needs. When it comes to email, for example, what we are seeing is 70% of emails still need some sort of human interaction and it could be very little. It could be just correcting a couple of things and mostly it is because machine translation engines still work on a sentence by sentence space and so once you have multiple sentences, the probability of errors just goes up quite a bit. You might have inconsistencies across sentences, maybe one sentence comes out a bit more formal, or it is more of the fine-tuning.
If you go into marketing, you start getting into more of a traditional translating review process where you have two humans in the loop, one to do the translation and it has to do more with either expertise on particular topics or cultural adaptation and tone. If you translate with Google Translate, they do not have translation to European Portuguese. They only have Brazilian Portuguese so it is things like that where it is just data-related and so once you want to speak on behalf of a brand and you are translating someone that has all the brand knowledge, you need to have a lot of humans involved to do that too.
Esther: What are some of the challenges that you see when you are trying to scale this advanced technology into the enterprise systems, these real-life scenarios in companies that might have more bespoke or legacy setups?
Vasco: There are quite a few but also they have been changing a lot. Five years ago you would say anything about AI and machine translation, and there was a very strong reaction against it when it came to translation. Now that is certainly changing so companies are much more open. They have seen a lot of people interact with it, they are buying into it and seeing the results. It is now becoming a viable solution or at least we need to consider it as part of the solution because you have a lot of impact. I think that there are a few challenges.
One is there are still a lot of legacy systems, as you said, so integrations at the level that we want them to work, which is very seamless, translation as a layer should be almost invisible. You should be able to just speak normally and the other person should just understand you without having to actually pay attention to the fact that there was a translation. That should not exist and so in order to do that, you need almost native integrations into the content systems, whether it is, in the case of customer service, like CRM, Zendesk, Salesforce, Dynamics and so on.
Also if you go into marketing, Marketo’s and Pardo’s, and HubSpot’s, such that the end-user has a seamless experience and so there is a challenge there. When you add the technology that is disruptive and enables you to do things in a fundamentally different way, that means that you need to change the way you are doing things and there is change management involved. Covid-19 in a lot of ways accelerated the digital transformation of companies and so in a way that has been positive, but we are still talking about changing core processes of customer service or marketing or how things happen and so that is sometimes a challenge. The third is how the pace of rapid evolution of technology is increasing in itself and how do you continue to be at the forefront when there is pressure on all sides, but at the same time, you are trying to, in our case, build a vertically integrated end-to-end system that is able to deliver the whole package to a customer as a solution.
That is constantly a challenge as a startup because you are investing in a lot of different areas so we have a big surface area of tech that we need to continue to develop and so doing that in a way that is still competitive at all levels. At Unbabel we need to develop a machine translation engine that is at least as good as DeepL or Google Translate. We need to develop an AI-first TMS that is as least as good as Trados or anything out there. We need to develop an experience for the translator that is at least as good as Gengo or any other company that focuses on crowd, and so on. Every component needs to be best of breed, competing with a bunch of other companies that also are developing just those components and so that is definitely a challenge, but it is a fun one.
Florian: Talking about staying ahead of the curve, what is your take on all this NLP that is coming out like GPT-3? Is that something that you are really looking at and incorporating?
Vasco: We are in some areas, certainly the work we have done on COMET, the new framework for evaluation of machine translation that has been beating every benchmark out there. It benefited from large language models and so we are seeing the impact on the ability to generate fluidity. I still think that we are a bit in the uncanny valley territory, things are quite good, but I do not know if I am going to use them on a daily basis. I think in a couple of iterations, we will. We have not quite figured out how to incorporate them very easily and what I expect to see is that we are going to start seeing really innovative ways of using them, in ways that maybe we had not thought about before. I have not seen the killer app for GPT-3 but I have seen some apps that I am thinking are almost there.
Florian: You mentioned your CFO, James Palmer. In his LinkedIn profile, it says that he is responsible for finance planning and M&A, is there anything you can say around that? What is your planning around M&A?
Vasco: Until now we were very focused on one use case and now we are starting that journey of expansion throughout other use cases and the next phase for us is going to mean that we start having real options in build versus buy and how do we think about expanding to different areas? At this point, it is great to have someone with that experience on board but we do not have a specific acquisition that we are targeting. We are starting to consider our options as we continue to develop the company on what makes sense in each case. M&A is a very important strategy and so having the resources that can lead that is very important. Sometimes people ask me, when are you IPO-ing? If you want to build a big global company that solves a meaningful problem, you are most likely going to go through those stages. At some point, you are going to IPO, et cetera, but that is not the goal. Those are just things that happen along the way so it is good to have the resources that enable you to go through those stages but they will happen when it makes sense.
Esther: What is it like working with somebody like Sri who has a deep understanding of the industry dynamics?
Vasco: Everybody on the team was fascinated when we started interacting with Point72 and Sri in particular and I am saying this because Sri was the first time that I have talked to an investor that actually told me things about my market that I did not know, and about what I was trying to solve. That is very rare and it is not expected. If I think about Chris Tottman who led our Series A or Andy Vitus who led our Series B, they are all amazing investors in different ways. Chris is amazing with go-to market, sales and the early stage stuff. Andy is great at process and structure and how you keep focus on growth. They bring a lot of value to the table.
Sri on top of that also has a very interesting thesis on how the market is going to evolve and how we can unlock opportunities in ways that I had not thought about before. Most investors would tell you, we are here to help, but you need to build your business. It is rare when an investor on top of that gives you insight on how to build your business beyond pattern matching, so a lot of the benefit that you have from investors besides capital is visibility. They have seen a lot of things. They have this great pattern matching on seeing other companies go through similar issues and a lot of times what you see early on is different. Everyone is building a slightly different product and a slightly different market and so challenges are all different.
Then once you get to about 70 to 100 people, most companies have the same challenges. It is all about how do I scale engineering? How do I scale finance? How do I scale my go-to market? Pattern matching becomes much more useful because they apply to a wider array of companies. I think our investors continue to not only bring a lot of value but keep us focused on the big picture and that is an important thing because as you are developing a startup, it is very easy to get in the trenches and to be just focused on the challenge in front of you.
Florian: What is your outlook for the language industry as a whole? What are some of the most exciting things you are working on at the moment at Unbabel?
Vasco: The most exciting thing for me that we are working on is the language operations platform and for a lot of people maybe this is a fairly obvious thing but for me, this is an idea of a category that enables the growth of localization. I am seeing the same shift happening in companies and how they see language and language operations capturing that as the evolution of localization. Going from a siloed approach to something that has structured and strategic value for a company as they scale. I think it makes a lot of sense because as physical barriers come down more and more, especially with Covid-19, language becomes a bigger barrier.
It becomes more obvious even in a very simple example, if I were to ask about famous Chinese singers or famous Chinese writers, we probably do not know very much, at least I do not. In general, if you ask famous American singers or writers, people will know them, and it is this dissemination of content, dissemination of ideas and that applies at all levels, including companies, brands, content creators, and the big hurdle there is language. It is these weird language barriers that create an inability for us to access knowledge, content ideas, product services and we see that happening across the entire spectrum.
Customer service is a simple but powerful example. If you happen to be born in a country that does not speak English, your access to customer service is way worse than if you do speak English. If I want to call British airways in the US, I have a 24/7 line open. If I want to call them from Portugal, it is five days a week, nine to five and that is a very simple example. Now that we have solved a lot of logistic issues, like, how do we ship something? How do we build something? We can do it easily from anywhere in the world to anywhere in the world. Solving language as a barrier is an even bigger challenge and will have an even bigger impact. I think we can do it so I am very excited about that.