Exploring news events using random walks

Large collections of news stories are hard to understand. When a major event occurs, there are numerous media reports describing it. When faced with multiple unstructured news reports, users may find it difficult to capture the big picture of the news story. Search engines can help retrieve information on a particular news event, but users may find it difficult to discover connections between news events and track how an event evolves over time. Previous work on story chains addresses this issue. A story chain is a sequence of news articles that connects two news events by revealing hidden relationships between them. The problem with this method is that users need to provide a start and end event (news articles) to the find relationship between them. Users who do not have a complete idea about an event do not know what they are looking for and cannot provide meaningful endpoints.;This thesis relaxes the need for an endpoint. Given a starting news event, we present users with endpoints describing all possible aspects of the event. This provides users with different perspectives on a news story and helps them to better capture the big picture of the story. Users can further choose an endpoint based on the aspect they are interested in and explore how the endpoint is connected to the starting event. This presents users with a summarized view of a news story. In this thesis we describe a random walk based algorithm to find how a news event evolves over time. Possible endpoints that cover different threads of the news story are obtained by multiple iterations of random walks. We describe methods to construct news stories that are coherent, relevant and provide good coverage of the event. Finally, we evaluate our algorithm over real news datasets constructed from the New York Times Search API and The New York Times Annotated Corpus. Experimental results show that our proposed algorithm is effective in providing users with a coherent view of a news story.