Broadcast News Navigator (BNN) Demonstration

1 Intelligent Segmentation

Fundamental to BNN is the exploitation of text, audio, and imagery streams and associated cues which are used to detect story and commercial segment boundaries and to select media elements to use for summaries and mul­ timodal displays. We will demonstrate how certain cross media cues such as language content (e.g., frequent weather or sports terms, funding and/or copyright no­ tices), discourse cues (e.g., "coming up next"), music (e.g., characteristic jingles), silence indicating breaks to commercial, and visual cues (e.g., logos, anchor booth vs report shots) help signal program start/end, anchor/report shots, commercials, and/or story shifts. Underlying BNN is a set of machine learned, time-enhanced finite state automata modeling news structure that take into account the above cues and probabilistic, temporal models of event occurrence [Boykin and Merlino, 2000].

3 Relevancy Feedback

Users can navigate directly from the Story Skim of Fig­ ure 2 on the left to a "Story Details" display as shown on the right. "Story Details" include the key frame, one line summary (the story line containing the most named enti­ ties), all extracted named entities, and pointers to video source, transcript, and relevant stories. The user can ei­ ther review the story or engage in further query refine­ ment, accomplished with an underlying Local Context Analysis (LCA) algorithm.

4 Visualization and Discovery

As shown in Figure 3, users can visualize named entity frequencies by collection, source, or date range, visualize temporal occurrences, and display story occurrences associated with geospatial regions, animated over time. Finally, users can interactively run data mining algo­ rithms to discover relationships among named entities (e.g., people associated with locations or organizations), and automatic detection of topic clusters . In addition, users can set profiles that specify their media (e.g., text summary, image key frames, summary), and navigation preferences (e.g., overviews or not, links to related sto­ ries). We demonstrate live how BNN users can find video stories and answer questions about two to three times as fast as with original sources, with no performance loss.

Acknowledgments

BNN is the product of numerous contributors including Stanley Boykin, Andy Merlino, Warren Grieff, Jay Ponte, Chris Clifton, and Chad McHenry.