Automatic association of news items

摘要

A new media for news delivery has emerged in which print, photographs, video, and audio can be integrated into personalized multimedia news presentations. An electronic news delivery system produces personal ‘editions’ by selecting news items from various sources in a variety of media types and ordering and/or grouping these items for presentation. Clearly, if personalized and dynamic ‘editions’ of news are to be a reality then algorithms are needed that select and group items with minimal human intervention. In this paper we will examine only one of the problems involved in the automatic generation of electronic editions: the association of related items of different media type, specifically photos and stories. The goal of this research is to be able to determine to what degree any two news items refer to the same news event. This metric has several uses. First, it can be used to link multimedia items that can be shown together, such as a video, photo, and text story related to a shipwreck or state visit. Second, it can be used to form clusters of very similar items from a variety of sources so that one or two can be chosen to represent that event in an edition. In this paper we discuss the specific association of text and photo news items although our approach to the problem of defining relationships between photos and stories applies to a larger domain of news items including scripted news video clips and scripted radio broadcasts.