This algorithm reads you like an open book

Artificial intelligence can beat human players in complex tasks, such as games of chess or Go. Ironically, though, when it comes to mundane human activities, computers fare much worse. To address this computational challenge, a group of Stanford researchers designed an intelligent system that used fiction to learn about, and accurately predict human behaviour.

Ethan Fast and his team trained the algorithm, which they called Augur, to analyze over 1.8 billion words of fiction sourced from a publishing platform Wattpad.  

"There's not that many data sources that really contain this really rich understanding of the.. many things that people can do." Ethan tells us. "Maybe if you look at their journals or track them with sensors, you can figure out exactly where people are and what they do. But short of that, fiction really provides a great signal."

Ethan explains that fiction can help AI better understand the complexities of human behaviour by establishing common object-verb associations. "You have a lot of scene context that describes the relevant objects that are around you when things are happening." he says.

For example, Augur learned that when we encounter a bench, we're likely to spot it, sit, take a seat, slump or even plop on it. Similarly, Augur finds relationships between behaviours to predict one based on the other. If someone orders coffee, this action will likely be followed by one of these: eat, take sip, take bite or pay.

Augur's algorithm already shows some promising results in preliminary applications. Ethan explains that knowledge bases like Augur can help computers anticipate the user's needs in the ever-evolving world of ubiquitous computing.