Thought Leaders in Artificial Intelligence: Spence Green, CEO of Lilt (Part 4)

Spence Green: One of the first applications of digital computers for cryptography and bomb-making were developed with machine translation. People started working on this in the late 1940s and early 1950s. Machine translation research surged and flowed over the following decade and it took off after 9/11 when the United States government realized that it didn’t have enough Arabic language speakers.

It started investing money and research in the university system to build these machine translation systems. Out of that came Google Translate and Microsoft Translator Hub. The latter was what put my co-founder and me through grad school. There were a bunch of us in 2000 working with machine translation.

The goal within governments, with scarce linguist resources that they have is to give them tools and software to be able to amplify each person. The goal is to enable them to process more information but then to also enable people within the government who don’t speak another language to have a capability to translate documents rapidly using machine translation. They can at least get an understanding of what the information is about.

The key difference is one of domain. In the public sector, the languages that we usually care about are Chinese, Japanese, French, and German. You train these systems in public domain texts like news wire and government proceedings. Internally in the public sector, you often care about the low resource languages like Farsi, Urdu, Hindi, and dialects of other languages.

The things that you are trying to translate are very much not government proceedings and news wire. You have to do a lot of domain adaptation to make these systems effective in government.

Sramana Mitra: Talk to me about competition. We did a large story about Unbabel recently. They are specifically going after customer service. They have a good partnership with Zendesk and a lot of Zendesk customers are using Unbabel for multilingual customer support. In the general area, this is one company that we have found interesting. How do see the competitive landscape?

Spence Green: The real competitor is the market that we are in. This is a large market estimated at $30 billion. It is a large multi-billion market for sure and it is expanding. The challenging part of our market is that it’s historically a BPO market. It’s highly fragmented. There are a lot of BPOs and their services are undifferentiated because they all do the same thing.

What is happening in the automation of that market? There is a lot of structural resistance to automation and it is also somewhat difficult to turn a market that’s historically services-based into one that is now driven by technology and machine learning. I would say that our greatest competition is the market itself and the effort required to consolidate and change a large, mature, and slow-moving market.

Sramana Mitra: It has very high potential from the technology point of view. Also, don’t you think that this is going to be a huge job loss phenomenon? 

Spence Green: I don’t if you think about the amount of information being created versus the number of people who are capable of doing multi-lingual work. One is going up at an exponential rate and the other is flat and the only way to bridge that is through technology. We need to amplify the scarce resources that we have and make technology that is good enough to process the rest of it in a fully automatic mode. This has to be done if we want to bridge this language gap across the world.

Sramana Mitra: Very good answer. Switching questions, what are open problems in this general domain that you think some new entrepreneurs should go after? 

Spence Green: There is a lot of interesting speech translation. For example, to build a cascade in the state of the art for building these speech systems, you have an ASR system and then it goes through the system. There is some research right now on building direct speech systems. That is very expensive because of the training data requirement.

I think that for speech translation and speech systems, you have a significant drop in capability once you have moved away from English. Multilingual speech translation is an area of some growth and interest. Within the enterprise, companies are moving to different modalities for their assets. Right now, they are not great solutions for making those multi-lingual as well. I think that is one important area. 

Sramana Mitra: Very interesting interview and work. I am very happy to meet you. Thank you for your time.

This segment is part 4 in the series : Thought Leaders in Artificial Intelligence: Spence Green, CEO of Lilt
1 2 3 4