Improving broadcast news segmentation processing

Abstract

We discuss techniques and results for automatically segmenting multimedia sources, with specific emphasis on broadcast news. We describe techniques that detect news events (e.g., start and stop, of stories or advertisements) using simple cues within and across multimedia streams (e.g., audio, video, text). In addition, we discuss a process for an evaluation and modification of our model using precision and recall metrics. Guided by our findings, we isolate multimedia cues that we consider critical in news story and advertisement segmentation. We also discuss ideas for improving segmentation beyond the methods described. Throughout the paper, we reference MITRE's Broadcast News Navigator (BNN) used to perform event segmentation.