That wasn’t the core part of Thursday’s panel discussion on “computational storytelling” organized by Nick Diakopoulos of the Tow Center for Digital Journalism at Columbia Journalism School, and moderated by yours truly, but it threw up a number of smart, interesting ideas that get us beyond doing what we’ve always done and into truly taking advantage of technology to help fulfil journalism’s mission.
Among the ideas: Different versions of the same story for different people; hypothetical stories to explore different theories, and “perturbation analysis” to understand how solid a story’s conclusions might be.
But first, some background on the panel, which featured three top-notch computer scientists: Jichen Zhu of Drexel, Mark Riedl of Georgia Tech and Larry Birnbaum of Northwestern U and Narrative Science. We spent nearly two hours exploring a wide range of questions about machine-generated narratives, from biases inherent in algorithms, data and interaction design to the importance of the provenance of data in a world where stories are created purely from data to our inherent (if unjustified) trust of narrative and anthropomorphic avatars.
You can watch it – telling insights, corny jokes and all – here. It’s pretty interesting, even if I say so myself.
For example, we touched on the subject of how there’s some research to suggest that people respond better to avatars than to impersonal interfaces. It may make sense, if that’s the case, to try and create more human-like user interactions for news sites. But does it also mean that Amazon would sell more stuff if recommendations came to us in the form of conversations with a vaguely human system powered by machine-generated text?
And if humans are inherently biased towards creating and believing narratives, how can we help audiences better understand what assumptions are baked into the stories they’re reading? If those stories are machine-generated, then there are likely fairly explicit ideas about how they were created; and if so, asked Larry Birnbaum, why not let people “tune” the assumptions to see how the story might change? His analogy was that of a self-driving car, which can be – and presumably is – programmed to trade off speed for safety, or vice versa. Why shouldn’t machine-generated stories offer the same kinds of controls to readers?
Similarly, if there are questions about how accurate the data is that’s used to generate stories, how about allowing readers to explore different scenarios by tweaking the data – what Mark Riedl described as “counterfactuals” and Larry framed in terms of “perturbation analysis.” Regardless of what we call it, it’s a fascinating use case to help bring data to life for people and to help them explore the impact of it on their lives. It’s also a way, as Larry noted, of understanding how “stable” a result is – something that any sports fan can understand when looking at a game result and trying to figure out it hinged on a few turning points or was a complete blowout.
Speaking of sports, Larry also described an early iteration of sports story that he had tried to build – one that emphasized suspense (“it was late in the eighth inning, and John Smith stepped up to the plate with the fate of the championship resting on his veteran shoulders…”) over results (“John Smith’s three-run homer in the eighth inning sealed the championship…”). It turns out, of course, that most sports fan prefer a story that tells them what happens up front; they don’t need to be seduced into reading a story about their team.
On the other hand, non-sports fans may need such suspense to keep them interested in a topic they don’t care about. With human sportswriters, of course, we can only write one version of that story, and generally we choose the one with the biggest audience. But if machines can write such stories – and they can – why not write a different type for each type of reader? Why not do that on stories about education, or health policy, or economic data?
If the purpose of journalism is to better inform society, not solely through the provision of dry facts but with engaging narrative and tales, can machines help us democratize such storytelling by providing the kind of story that each person needs? Does letting readers explore hypotheticals help them better understand why things happened the way they did?
Maybe technology isn’t able to give us these capabilities yet. But shouldn’t we be thinking about how best we can use them now, so that we’re really using them in the most useful way possible?