Computational Models of Narrative: Review of the Workshop
Finlayson, Mark A., Richards, Whitman, Winston, Patrick H., AI Magazine
Narratives are ubiquitous in human experience. We use them to entertain, communicate, convince, and explain. One workshop participant noted that "as far as I know, every society in the world has stories, which suggests they have a psychological basis, that stories do something for you." To truly understand and explain human intelligence, reasoning, and beliefs, we need to understand why narrative is universal and explain the function it serves.
Computational modeling is a natural method for investigating narrative. As a complex cognitive phenomenon, narrative touches on many areas that have traditionally been of interest to artificial intelligence researchers: its different facets draw on our capacities for natural language understanding and generation, commonsense reasoning, analogical reasoning, planning, physical perception (through imagination), and social cognition. Successful modeling will undoubtedly require researchers from these many perspectives and more, using a multitude of different techniques from the AI toolkit, ranging from, for example, detailed symbolic knowledge representation to large-scale statistical analyses. The relevance of AI to narrative, and vice versa, is compelling.
The Computational Models of Narrative workshop (1) had three main objectives: (1) to understand the scope and dimensions of narrative models, identifying gaps and next steps, (2) to evaluate the state of the art, and (3) to begin to build a community focused on computational narrative. The interdisciplinary group of 22 participants (see figure 1) included computer scientists, psychologists, linguists, media developers, philosophers, and storytellers. Ten speakers were selected to represent a range of views, and their presentations were organized into four groups, each followed by an extensive discussion moderated by a panel. The day after the presentations, there was a lively, morning-long extended discussion. The meeting's audio was captured and later analyzed in depth. A detailed summary of the group's conclusions at the workshop appears elsewhere (Richards, Finlayson, and Winston 2009), together with recommendations for future initiatives. (2) Regarding models of narrative, the main findings were: (1) a three-level organization of narrative representations unifies work in the area, (2) the area suffers from a deficit of investigation at the highest, most abstract level aimed at the "meaning" of the narrative, and (3) there is a need to establish a standard data bank of annotated narratives, analogous to the Penn Treebank (Marcus, Marcinkiewicz, and Santorini 1993).
A Three-Level Organization
Computational modeling requires a precise statement of the problem (or problems) to be solved. Thus, an obvious first step is to understand how narrative should be represented.
There were three common denominators among the representations presented at the workshop: (1) narratives have to do with sequences of events, (2) narratives have hierarchical structure, and (3) they are grounded in a commonsense knowledge of the world. Similarly, it was uncontroversial that narratives can be told from multiple points of view, and that all four of these characteristics were independent of whether or not a narrative was told with words. (3)
After analysis of the presentations and discussions, it became clear that all the representations considered at the workshop were subsumed within a three-level structure. The heavily investigated middle level stressed event sequences that were built on the classic logical-predicate-like representations introduced in artificial intelligence in its earliest days, exemplified by instances such as KISS (JOHN, MARY) and CAUSE(SHOOT, DIE).
Below the middle level were representations that examined the detailed structure of the narratives in question. There was quite a bit of work at this detail level, such as commonsense reasoning (Mueller 2007), discourse structures (Asher and Lascarodes 2003), argument-support hierarchies (Bex, Prakken, and Verheij 2007), or plan graphs (Young 2007). …