Causality in Hundreds of Narratives of the Same Events

Empirical research supporting computational models of narrative is often constrained by the lack of large-scale corpora with deep annotation. In this paper, we report on our annotation and analysis of a dataset of 283 individual narrations of the events in two short video clips. The utterances in the narrative transcripts were annotated to align with known events in the source videos, offering a unique opportunity to study the regularities and variations in the way that different people describe the exact same set of events. We identified the causal relationships between events in the two video clips, and investigated the role that causality plays in determining whether subjects will mention a particular story event and the likelihood that these events will be told in the order that they occurred in the original videos. Analysis of Narrative Causality plays a central role in narrative. In artificial intelligence research, logical formalisms of action and planning have long been shown to provide effective ways to represent and manipulate causal structures (cf. Wilensky 1983; Riedl and Young 2010). However, a computational model of narrative requires more than just causal soundness. There are complex and idiosyncratic psychological processes that dictate the presentation choices a narrator makes, and considerable research has been done on the impact of causal structure. In the field of discourse psychology, Trabasso and van den Broek (1985) developed the causal network model, a highly influential model of the causal structure of goal-based stories. Using this model, discourse psychologists have been able to predict a wide range of observed memory behaviors, sentence reading times, recognition priming latencies, lexical decision latencies, goodness of fit judgments for story sentences, and the inferences produced during thinking aloud (van den Broek 1995; Magliano 1999). This type of empirical research can inform better computational models. However, it is difficult to obtain suitable corpora for analysis. The narratives being studied must be paired with detailed descriptions of the situations that are being described. Further, variations in situations and variations in narrative style confound one another. Ideally, researchers would be Copyright © 2011, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. able to analyze narratives from hundreds of subjects, where each is describing the same situation where the causal details are known. We have identified an existing corpus of narratives, collected as part of a separate research effort, which comes close to this ideal. Gratch et al. (2007) describe a series of experiments to study the rapport that people can develop with interactive animated virtual characters. In that research, hundreds of human subjects told narratives describing the events of the same two video clips. The transcriptions of these verbal narrations (the Rapport Corpus) present a unique opportunity to study the factors that influence narrative choices. With hundreds of narratives of the same set of (known) events, it is possible to conduct quantitative analyses of the specific ways that the causal structure of the situation (the events of the two videos) plays a role in narrative discourse. In this work, we have annotated the Rapport Corpus to link the narrative events in each transcript to the known events in the source videos. We have also identified the causal structure of those known events. Through a quantitative analysis of the resulting set of annotations, we explore the role that causality plays in determining two key aspects of narrative generation. First, we show that events with numerous causal connections to other events are more likely to be mentioned in a narration. Second, we show that the number of causal connections to and from an event impact the likelihood that that event is narrated out of the order in which it actually occurred. The next sections of this paper describe the Rapport Corpus and our annotation process. We then present the methodology and results of our statistical analysis to answer our two research questions, and conclude with discussion.