Inferring Event Causality in Films via Common Knowledge Corpora.

Human understanding of a movie plot is partially driven by our ability to reason about the causal relations between events. Thus, recognizing causal chains of events is a key requirement for computational models of movie understanding. In this paper, we propose to use available corpora of common-sense knowledge about human behavior for automatically inferring event causality in movie scenes. Our initial experiments with a dataset of annotated movie events and a corpus of human commonsense reasoning demonstrate that a) for 86% of movie events, there exist relevant commonsense rules and those rules can be used for predicting other movie events. b) in 70% of the cases, the consequences of the rules triggered by movie events can accurately or semi-accurately predict subsequent movie events.
These preliminary results indicate the potential of automated commonsense reasoning to detect the narrative structure in movies. Hence, the proposed method can contribute to the development of story-related video analytics tools, such as automatic video summarization and movie editing systems.