Probabilistic Narrative

Rob, the lead engineer on The Demo Graphic Replicator, has gladly stepped forward to explain some more about the reasoning behind the design. It's a pleasure working with this kind of thinking. Over to Rob...


Stories have structure, and that structure can be represented and manipulated by computers. In the mid-19th Century Ada Lovelace only imagined generating musical compositions using programs, not stories. But as early as the 1960s programs were being written to use Vladimir Propp's system for representing existing Russian fairy stories from his book "Morphology Of The Folktale" to generate descriptions of new stories.

The Artificial Intelligence programming of the 20th Century relied for the most part, like Propp's notation for folktales, on discovering and encoding rules. Narrative generation systems followed this lead, and tended to produce series of events that lack story arcs, character development or other high-level features of stories. The state of the art by the end of the 1980s is described in the book Possible Worlds, Artificial Intelligence and Narrative Theory by Marie-Laure Ryan. At best the systems described produce Aesopian fables without the moral; narrative without narrative.

In the 21st century, the rise of Collective Intelligence means that statistical methods based on information gathered from the activity large numbers of people have replaced rule-based approaches derived from expert opinion for many tasks. This isn't a new development for artificial intelligence or for generative narrative, Markov chains and other statistical methods have long been used to generate the text of short stories. What is new is the availability of vast amounts of structured text on the Internet for those statistical methods to be used on.

DGRs are story characters represented as lists of words that represent their personality, interests and activities. Each time the DGR is run it chooses some of those words randomly to represent its current interior state. It searches Twitter for tweets containing any of those words. Then it retweets the first found tweet to express its own current situation in the story.

The tweets may contradict each other over time, but at a demographic level they will tend to match the character. A significant amount of the retweeted tweets will match the DGR's personality and situation, and incongruous retweets can help to shake up the story. The DGR is a probabilistic character in a probabilistic narrative.

The breakthough with DGRs is to recognise that although some retweets won't fit the character the average of the retweets will be "good enough". The reader can fill in the gaps and file off the rough edges, rendering the character imaginatively in their mind as they would with a scripted character in a story written with a pre-planned plot by a single human being. They give the storyteller a powerful new tool to work with, and they give the reader engaging new characters to imagine and follow.