Discovering links between news events

Abstract

The analysis of news requires the identification of its constituent events. Since events are complex linguistically and ontologically, discovering their reference is difficult. In this work, a method to automatically link articles containing references to events with the news reports which originally describe those previously identified events in detail is proposed. This method consists of the following steps: detecting the background of an event, i.e. the segment which mentions events prior to the current action, by first subdividing the text into multi-paragraph units. Valuable information is then extracted from these and the articles in which the events are reported more thoroughly are associated with their parent articles based on some similarity. Extracted information includes named entities, especially temporal expressions, and relations between them. In this project, the Guardian was used as a source of news reports and to assess the performance of our system based on the manual annotations often found in the articles' text. An evaluation plan of the system is also presented together with the experimental results.