Story Sifting Using Object Detection Techniques

Wilkins Leong, Julie Porteous, John Thangarajah

Abstract: Our focus is on the problem of story sifting (or story recognition), which is the automated sifting of interesting stories that emerge from the interactions between virtual characters in virtual storyworld environments. To date, approaches to story sifting have been either: manual, with the burden of authoring sifting patterns; or automated, but of limited efficiency and scalability. In this paper, we address these shortcomings via a novel approach that recasts the problem as one of object detection. We demonstrate how an object detection model can be trained to detect story arcs of prominent story types emerging from an interactive virtual storyworld which can occur anywhere in the storyworld’s timeline. We evaluate our approach using synthetic virtual story environments that show our approach is able to: detect story arcs anywhere in the storyworld’s timeline with a high degree of accuracy and more efficiently than the state-of-the-art Arc Sift, making it scalable in real-time.