Automatically Learning to Tell Stories about Social Situations from the Crowd

Narrative intelligence is the use of narrative to make sense of the world and to communicate with other people. The generation of stories involving social and cultural situations (eating at a restaurant, going on a date, etc.) requires an extensive amount of experiential knowledge. While this knowledge can be encoded in the form of scripts, schemas, or frames, the manual authoring of these knowledge structures presents a significant bottleneck in the creation of systems demonstrating narrative intelligence. In this paper we describe a technique for automatically learning robust, script-like knowledge from crowdsourced narratives. Crowdsourcing, the use of anonymous human workers, provides an opportunity for rapidly acquiring a corpus of highly specialized narratives about sociocultural situations. We describe a three-stage approach to script acquisition and learning. First, we query human workers to write natural language narrative examples of a given situation. Second, we learn the set of possible events that can occur in a situation by finding semantic similarities between the narrative examples. Third, we learn the relevance of any event to the situation and extract a probable temporal ordering between events. We describe how these scripts, which we call plot graphs, can be utilized to generate believable stories about social situations.