Towards Learning Action Models from Narrative Text Through Extraction and Ordering of Structured Events

Event models, in the form of scripts, frames, or precondition/effect axioms, allow for reasoning about the causal and motivational connections between events in a story, and thus are central to AI understanding and generating narratives. However, previous efforts to learn general structured event models from text have overlooked important challenges raised by the narrative text, such as the complex (nested) event arguments and inferring the order and actuality of mentioned events. We present an NLP pipeline for extracting (partially) ordered, structured event representations for use in learning general event models from three large text corpora. We address each of the challenges that we identify to some degree, but also conclude that they raise open problems for future research.