We focus on the problem of learning semantics from multimedia data associated with broadcast video documents. In this paper we propose to learn semantic concepts from multimodal sources based on style and context detectors, in combination with statistical classifier ensembles. As a case study we present our method for detecting the concept of news subject monologues. This approach had the best average precision performance amongst 26 submissions in the 2003 video track of the Text Retrieval Conference benchmark. Experiments were conducted with respect to individual detector contribution, ensemble size, and ranking mechanism. It was found that the combination of detectors is decisive for the final result, although some detectors might appear useless in isolation. Moreover, by using a probabilistic ranking, in combination with a large classifier ensemble, results can be improved even further.

23 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu