Indexter: A Computational Model of the Event-Indexing Situation Model for Characterizing Narratives Rogelio E. Cardona-Rivera, Bradley A. Cassell, Stephen G. Ware, R. Michael Young Liquid Narrative Research Group North Carolina State University {recardon, bacassel, sgware}@ncsu.edu, young@csc.ncsu.edu Abstract Previous approaches to computational models of narrative have successfully considered the internal coherence of the narrative’s structure. However, narratives are also externally focused and authors often design their stories to affect users in specific ways. In order to better characterize the audience in the process of modeling narrative, we introduce Indexter: a computational model of the Event-Indexing Situation Model, a cognitive framework which predicts the salience of previously experienced events in memory based on the current event the audience is experiencing. We approach computational models of narrative from a foundational perspective, and feel that salience is at the core of comprehension. If a particular narrative phenomenon can be expressed in terms of salience in a person’s memory, the phenomenon, in principle, is representable in our model. This paper provides the fundamental bases of our approach as a springboard for future work which will use this model to reason about the audience’s mental state, and to generate narrative fabula and discourse intended to achieve a specific narrative effect. Keywords: Narrative understanding and generation, representations, retrieval and indexing, artificial intelligence, cognitive psychology 1. Introduction Historically, computational models of narrative have fo- cused on representation of the diverse structural properties of narratives (Lebowitz, 1985; Cavazza et al., 2001; Riedl et al., 2003; Szilas, 2003). These models consider only the internal properties of the narrative. Authors, however, intentionally design stories to affect their audience in specific ways (Bordwell, 1989; Holland, 1989). As Szilas (2010) has suggested, a computational model of narrative must go beyond simple story structure and account for how the experiencer receives the narrative. In this paper, we provide initial steps toward a compu- tational model that accounts for a user’s comprehension process during the experience of a narrative. This model, which we call Indexter, explicitly reasons about the salience of narrative events in a person’s memory as they experience an unfolding story. The salience of a narrative event indicates how recallable the event is in a person’s mind. An author’s manipulation of the salience of events during a narrative experience is a key means used to affect a reader’s comprehension of the story’s structure. Salience enables the drawing of connections between new material and earlier parts of the story. Salience prompts expectations about upcoming action. Lack of salience obscures predictions and facilitates surprise or misdirection. A model of narrative that accounts for salience could be linked to existing models that build off of salience to account for a reader’s inference-making process (Niehaus and Young, 2010), her feelings of suspense (Cheong and Young, 2006), and her level of surprise (Bae and Young, 2009), along with many other narrative phenomena. Though our current model focuses on the manipulation of salience in narrative, salience alone is not sufficient for the modeling or creation of effective stories. A story’s internal structure clearly plays a role in how a reader understands it (Graesser et al., 2002). Thus, the computational model that we present extends an existing planning-based approach to narrative (Young, 2007), which models coherent story structure (Riedl and Young, 2010). We augment this plan-based approach with information that allows us to model the updates being made to a reader’s mental model of the story during online comprehension, that is, during the process of experiencing the narrative. To do this, we incorporate elements into the planning model drawn from an empirically verified cognitive model of online comprehension called the Event-Indexing Situation Model (Zwaan et al., 1995a; Zwaan and Radvansky, 1998). While we are basing our work on a planning-based knowledge representation previously developed to generate stories, our discussion here does not describe a system that uses this representation in a generative fashion. The work we describe here is preliminary. It is the first step of a four- part research agenda involving: 1. Development of a plan-based knowledge representa- tion for narratives and an algorithm that characterizes the reader’s construction of event-indexing situation models. 2. Validation of the predictive power of the algorithm and representation. 3. Integration of the computational model into a genera- tive system. 4. Validation of the generative system in an online comprehension scenario. A generative system which uses a computational model that characterizes both the internal structure of a narrative and its effects on a reader during online comprehension will lead to the creation of more engaging, effective and understandable stories.