Lights, Camera, Action: Knowledge Extraction from Movie Scripts

Niket Tandon, Gerhard Weikum, Gerard de Melo, Abir De

Abstract: With the success of large knowledge graphs, research on automatically acquiring commonsense knowledge is revived. One kind of knowledge that has not received attention is that of human activities. This paper presents an information extraction pipeline for systematically distilling activity knowledge from a corpus of movie scripts. Our semantic frames capture activities together with their participating agents and their typical spatial, temporal and sequential contexts. The resulting knowledge base comprises about 250,000 activities with links to specific movie scenes where they occur.