Knowledge Discovery via Content Indexing of Multimedia and Text

Indexing and retrieving audio or video content presents challenges specific to the nature of these media. Two primary difficulties are the inaccuracy of speech recognition and the timed nature of streaming media — that is, the property that words and other information in audio/video are tied to times, and are not readily accessed and scanned at arbitrary positions. StreamSage's approach to resolving these problems exploits statistical properties of language to determine whether an occurrence of a query term is likely to signal a truly relevant interval, and where the segment should begin and end. We describe here the steps required to process large bodies of information in order to make them useful for intelligence analysts or other searchers. These include a set of contentspecific capabilities, such as categorization and classification, large coverage wordsense disambiguation, topic detection, and summarization.