Building a Cool Technology Company from Chicago: Narrative Science CEO Stuart Frankel (Part 5)

Sramana Mitra: How does it work? Can you explain the science behind it?

Stuart Frankel: We would have to fast forward a little bit as well, because at that time this was a baseball story writer. That’s the prototype that was built in the lab. It worked really well but all it did was write stories about baseball games. Ultimately, we have turned this into a horizontal technology, which I’ll tell you about in a minute. In short, the idea is that you’ve got lots of data and within that data, there’s a story to be told. In this case, it was a baseball story writer. It would look at relevant data associated with the baseball game—game data and also contextual data—and the system would write the story.

Sramana Mitra: Are you talking about all structured data? Is this software working on all structured data and then turning that into a natural language story?

Stuart Frankel: Yes. Maybe I should pause for a second on the chronological piece and tell you a little bit about the technology.

Sramana Mitra: That would be great because, obviously, it sounds like science fiction. We have a very sophisticated audience of technology entrepreneurs. I think they would enjoy listening about the science behind it.

Stuart Frankel: We started with this baseball story writer. We, ultimately, licensed that. We started a company and we set about building a commercial piece of software, which we wanted to be horizontal in nature. Instead of just writing stories about baseball games, we wanted to take any kind of data, as long as it was structured—basically, numbers, symbolic unambiguous text, regardless of the subject matter.

Sramana Mitra: For each domain, you need a certain data structure that we specific to that domain, correct?

Stuart Frankel: Yes. The technology would have to be configured for each domain and each use case, but the core technology is that it can take this data, and essentially generate language that sounds like it was produced by a human. It could be a tweet or that could be a 15-page investment research report that sounds like a Wall Street research analyst prepared it.

Sramana Mitra: I would love to see more of the algorithm but let’s get back to the chronological story.

Stuart Frankel: They had this prototype at the university. We started talking about creating a company and started talking about licensing the intellectual property in the university and building this technology that would, at first, tackle long-tail journalism. It was between 2009 and 2010. Traditional media had just gone through yet another gut-wrenching period—the financial crisis. Editorial organizations were just laying off reporters left and right. Stories were not getting as much coverage, especially long-tail areas. High school sports, for example, used to be a really big part of a local newspaper. Every Saturday, all of the Friday night football games would be written up. It was a big part of that local news experience. That stuff was really starting to go away. It was starting to away very quickly.

We felt that there could be a market opportunity to have a machine create those types of stories instead of a person. We formed the company in May 2010. We hired some early engineers from this research lab. We got started with five or six people. We started to take this prototype and built this horizontal technology. The first idea was to take the prototype and build on top of it. We’d essentially take this baseball writing functionality and copy that in order to expand the types of stories that the system could create. One of the best moves that we had made is we scrapped that idea after a few weeks. We decided to throw away the prototype and start from scratch. We spent the better part of 2010 building the first version of Quill, which is what we call the platform now. By the time we got to the end of 2010, we had a really interesting first pass at the technology, which was a truly horizontal technology that could take just about any kind of data and create stories from it.

We were really fortunate. We got some early press around the company and we got some early customers. We started working with organizations like Fox and Forbes. We expanded it outside of baseball and did basketball, football, and hockey. We started working with financial news organizations like Forbes generating earnings stories. You look at a typical news publication. If it’s financial news, there’s a lot of coverage around earning seasons of the top 100 widely-held stocks. When Apple announces their earnings, there’s thousands of stories about Apple. But on smaller regional companies, you can find almost no coverage. Again, we applied the technology to areas where we felt were underserved.

Sramana Mitra: Let’s say its earning seasons. Forbes wants to cover 100 earnings reports using your automated technology. What does this cost Forbes to avail of your services to do that?

Stuart Frankel: First of all, it’s typically thousands of stories. The beauty of the technology is that it can create thousands, if not millions. The example that I often use is we still work with a company called Game Changer to generate Little League baseball stories. Game Changer is a company that has developed an iPhone app that allows coaches and parents to capture Major League baseball quality data for these Little League games. At the end of every game, we get the data and within a second or so, have published through this app a story about that baseball game. We’ll do probably three to four million of those stories in 2015. It tends to be a lot or it certainly could be a lot. In terms of what we would charge organizations like Forbes, when we first started, we charged by the piece. It would be anything from a few dollars to $10 or $15 per story. Over time, we moved into more of a traditional software licensing model.

This segment is part 5 in the series : Building a Cool Technology Company from Chicago: Narrative Science CEO Stuart Frankel
1 2 3 4 5 6 7