A year after naming Narrative Science one of our Top 10 Tech Companies to Watch, we checked in with CTO Kristian Hammond to find out how the young company is doing.
The company's software, called Quill, uses artificial intelligence to analyze numbers and turn them into, well, narrative. An investor's quarterly portfolio performance numbers can be turned into a story about how the individual stocks, bonds and funds are performing. A bank branch's performance numbers can be automatically turned into a natural language report about that facility's strengths and weaknesses.
Narrative Science's early work was in market alerts. Detecting, for instance, that a company's stock experienced a 52-week high or low would trigger the automatic writing of a story.
More recently, the software has been used in banking for fraud detection and compliance reporting. It can parse through transaction streams, looking for violations of a bank's business rules, which are often based on compliance requirements. If the bank has to report all transactions over $5,000, the software will automatically find all transactions just under $5,000 and higher and issue alerts and compliance reports.
Artificial intelligence in the software is tuned to provide a broader context for each transaction - this could take the form of a paragraph or a full-page description.
This reduces the workload for overburdened compliance officers. "We take the commodity work, the grunt work," Hammond says. "If you take someone aside and say, what is it you love to do, they're probably not going to say they like to do the same repetitive task. Most people want to stretch the limit of their creativity."
One bank is using Quill to handle suspicious activity reports, others are in talks.
"Financial services are primed for this moment where there's all this data, all this work and analysis," Hammond says. "I need to be able to say it," meaning, tell the story and provide context. "Saying it at scale is hard unless you bring a machine to bear."
The basic engine behind Quill has three main components. The first analyzes data and finds facts, such as Company A beat expectations or Company B's long-term debt is larger than any other company in its sector. "Those facts are not enough and they're too much," Hammond says. "You don't want to know everything about everything."
The second piece is a module that evaluates the importance and interestingness of the facts. Importance is defined by the domain of interest if you're looking at the stock market, a company going bankrupt is important. "Interestingness" depends more on the user. "If I'm generating your portfolio analysis and you don't own stock in that company, it's uninteresting," Hammond notes.
The third component is what Hammond calls a micro language generator. It has tiny blueprints and micro-changes: for any given idea you have hundreds of ways to say things," Hammond says. "What we're doing to always improve the system is we're making those modules better - creating a clearer and cleaner set of data analytics, a tighter version of what's important and interesting, and improvements in natural language understanding, so more of the context of what the system has said drives what happens."
The reports are configured by humans - a team of engineers and eight journalists who are now called "meta journalists," "configurers" and "content architects" in Narrative Science's world. They determine the patterns the software should look for, such as a company performing better than expectations.
"Once it's configured, the system says, I know how to talk about sales performance review, I know how to do the analysis, you don't have to teach me anymore," Hammond says.
When the software has learned how to handle a report well, the meta journalists will move on to new engagements and nuances of existing ones. "For intelligence, there's always a next horizon, there's always a smarter version," Hammond says. "That will go on forever."
The software can produce a gamut of outputs from simple headlines or alerts to "an eight-page exegesis that's a full and complete reckoning behind a single number that includes text and visualizations and is rendered as a PDF," Hammond says. "It's a genuine research report."
The company uses a software-as-a-service delivery model. Data from banks and other clients comes through Narrative Science's API, through an FTP drop or a direct feed.
It's developing a private cloud model for one client that is expected to go live in the first quarter. "We'll give them the software and they'll do what they want with it," Hammond says. "We want to provide a model for organizations that are concerned about security issues where they can manage their own private cloud we never see the data."
Time and volume have improved Quill, Hammond says. "The more we do, the better the technology gets we can see more examples and can adjust things," he says. "As opposed to when you outsource, the more you do, the worse it gets, because you've used up your high talent pool and you've got to go deeper into lower skill levels."
The software can be applied many places - banking, insurance, journalism and beyond, Hammond says. "I believe within my lifetime I'll see a day when people will look at spreadsheets the same way we look at computer punch cards - as a mechanism for communicating with a machine and letting it communicate with us. That mechanism and that time are gone. By humanizing the machine, giving it voice, we can rid the world of an awkward and painful mechanism."
He's never met anyone who relishes using a spreadsheet, Hammond says. "They relish what they get from it. What if you could get that without struggling?
"Data is there to serve us," he says. "It's not there to beguile us or annoy us but to serve us."