TECHNES Encyclopedia Database

The picture above shows an excerpt from a fable generated by Tale-Spin. Here it is apparent that narrative was not at the heart of the project, as this narrative served only to set out explicitly the characters’ internal reasoning, the actions they undertake, and their consequences. Automatic planning remains one of the main approaches in research into generating narrative.^[7] One of its advantages is the ability to take changes in the context into account: if a player takes action in the world, the “plans” of all agents can be updated. And yet this depends on a strict formalization of what constitutes a “good” story, and the results may quickly appear rigid. Incidentally, many systems are based on quite classical theoretical structures, such as Propp’s Morphology of the Folktale and Joseph Campbell’s The Hero with a Thousand Faces.

Simulation and Emergent Narratives

In contrast with the planning approach is the so-called “emergent” narratives approach. Rather than spending time defining, formalizing and generating narrative frameworks, an emergent narrative models and simulates a complex fictional world in the hope that interesting stories will arise there “naturally,” like in the real world. In a system such as this, the “narrative” also adapts to the players’ actions not as narrative, precisely, but simply as a dynamic world. This approach makes it possible not to define beforehand what constitutes a “good” story and gives players the freedom to determine and design on their own what they deem to be interesting narratives. Some existing video games are strongly associated with this approach, such as the popular The Sims (EA) franchise, but also Crusader Kings (Paradox) and Dwarf Fortress (Bay 12). The fundamental problem with emergent narratives is that it that they may not emerge, may emerge rarely, or may have no interest.^[8] Typically, one must play a large number of hours on the games mentioned before witnessing an anecdote worthy of being told.

Deep Learning

Research into automatic scriptwriting has not escaped the deep learning tsunami of the artificial intelligence sector. This method has been explored both for generating stories and for the direct production of narrative texts. The former is more problematic, as deep learning relies on the massive availability of standardized data. As no such bodies of data exist for entities of abstract narratives, these must be laboriously constituted through the manual labelling of narratives.^[9] Direct text generation, for its part, can easily draw on millions of texts available on the Internet, and has made impressive advances. Consider, for example OpenAI’s GPT-2 and GPT-3 systems, which can generate perfectly credible texts. Here we find the same problem, however, as that described with respect to non-contextual grammars, meaning that by generating the story and its narration through the same process, the system does not naturally preserve an internal representation of its constituent elements, and problems of coherence quickly arise. Nevertheless, approaches are being developed for “keeping in memory” certain elements of a text in order to influence subsequent generations.^[10]

Another issue is the question of “creativity.” Deep learning excels in finding common motifs in its data and tends to generate prototypes. As such, it goes against what is frequently sought in fiction: surprise and novelty (although genre fiction and genre cinema regularly contradict this supposition). One way around this is to mix the kinds of data in order to force hybrid and potentially surprising learning. Consider for example the work of Janelle Shane, who loaded cooking recipes and Lovecraft short stories into a neuronal network and obtained extremely disconcerting results.^[11]