This paper describes the use of connections between named entities for summarization of broadcast news. We first extract named entities from a transcript of a news story, and find related entities nearby. In the context of a query, a link graph of relevant entities is rendered in an interactive display, allowing the user to manipulate, browse and examine the components, including the ability to play back video clips that mention with interesting relationships. An evaluation of the approach shows that completely automatic summaries from a year of broadcast news can reflect almost 50% of the entities in a manually created reference summary found on the web. Locations are the most accurate aspect of summarization.
展开▼
