Narrative Theme Navigation for Sitcoms Supported by Fan-generated Scripts

摘要

The following article presents a novel method to generate indexing information for the navigation of TV content and presents an implementation that extends the Joke-O-Mat sitcom navigation system, presented in [1]. The extended system enhances Joke-o-mat's capability to browse a sitcom by scene, punchline, dialog segment, and actor with word-level keyword search. The indexing is performed based on the alignment of the multimedia content with closed captions and "found" fan-generated scripts processed with speech and speaker recognition systems. This significantly reduces the amount of manual intervention required for training new episodes, and the final narrative-theme segmentation has proven indistinguishable from expert annotation. This article describes the new Joke-o-mat system, discusses problems with using fan-generated data, and presents results on episodes from the sitcom Seinfeld, showing segmentation accuracy and user satisfaction as determined by a human-subject study.