Document Selection for Extracting Entity and Relationship Instances of Terrorist Events

Abstract

In this chapter, we study the problem of selecting documents so as to extract terrorist event information from a collection of documents. We represent an event by its entity and relation instances. Very often, these entity and relation instances have to be extracted from multiple documents. We therefore define an information extraction (IE) task as selecting documents and extracting from which entity and relation instances relevant to a user-specified event (aka domain specific event entity and relation extraction). We adopt domain specific IE patterns to extract potentially relevant entity and relation instances from documents, and develop a number of document ranking strategies using the extracted instances to address this extraction task. Each ranking strategy (aka pattern-based document ranking strategy) assigns a score to each document, which estimates the latter's contribution to the gain in event related instances. We conducted experiments on two document collection datasets constructed using two historical terrorism events. Experiments showed that our proposed pattern-based document ranking strategies performed well on the domain specific event entity and relation extraction task for document collections of various sizes. 309 310 Chapter 15. Sun, et al.