categories

HOT TOPICS

Thought Leaders in Artificial Intelligence: Spence Green, CEO of Lilt (Part 2)

Posted on Friday, May 14th 2021

Sramana Mitra: Give me some examples of things that Google Translate cannot do, but you can.

Spence Green: Google Translate can do the same thing a person can do. You can give a string input and it can give you a string output. People can do that too. What you don’t have is any certificate of correctness. You don’t know if it is right or not.

What you need for that is a production process that goes through quality checking and delivers to you this translation with a certificate of correctness. That is where a human in the loop process comes in. The key point is to make the human intervention or the human component as efficient as possible because that is the limiting factor. It is also the slowest and most expensive step in the production process. 

Sramana Mitra: What is the difference between an agency that has a lot of human translators who take documents through Google Translate and you?

Spence Green: This was the subject of my dissertation. What you just described is a cascade of machine output that people correct. This is called post-editing. People have been doing this since the early 60s. There is nothing wrong with it. It works. There are two issues that we have tried to address in that workflow.

One, this renders the work for people to be tedious and boring. Instead of active cognitive tasks like translation, it turns into fixing machine errors and output. That is a less interesting and less fun job. Professional translators are college-educated and, in many cases, with a graduate degree. There are other things that they can do besides sit around and correct machine output. This tends to drive away the good translators, which I think is bad for humanity.

The second thing is that it shortchanges machines too. You can think of what translators do as creating training data. When you have a linear pipeline like that, the machine never gets the opportunity to learn from the corrections that human beings make.

Part of our research was creating new interfaces that are more interactive and engaging for people to work with. The second part was online self-training so that when people make changes to the machine output, the machine can train on it immediately and update its model so that it doesn’t make the same mistakes in the future. 

Sramana Mitra: What about the human intervention piece? Is this something that you have on your staff?

Spence Green: Yes, we have a community of translators in a variety of different languages and areas of expertise that we pair with our customers. For example, Intel is one of our customers. For that customer, many of our translators have an engineering background. We also have a company like Asics, the shoe and apparel company. We have a translation team there that has more of a marketing and retail background. 

Sramana Mitra: You said that this is a community of translators? Are they freelancers? When you take on a client, the client is paying and you put together the manual workforce that completes the workforce?

Spence Green: That is correct. It’s a human-in-the-loop solution. We build the machine component for it and we also provide the people component of it. That is this community that we have built. When we work with a new company, we provide the people and the technology to them to solve the enterprise language problem.

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

Hacker News
() Comments

Featured Videos