Learning Threshold Parameters for Event Classification in Broadcast News
Abstract
In this paper we present two methods for automatic threshold parameter estimation for an event tracking algorithm. We view the threshold as a statistic of the incoming data stream, which is assumed to contain broadcast news stories from radio, television, and newswire sources. Query bias defined in terms of threshold estimators can be identified when a word co-occurrence representation for text is used. Our results suggest that both approaches learn bias from training corpora, leading to improved classification accuracy for event tracking applications. 1 Introduction The following work describes two automatic threshold selection algorithms for the event tracking problem. This problem was defined by the Topic Detection and Tracking (TDT) research initiative, a DARPA-sponsored effort comprising research groups from several commercial and academic sites. Event tracking is a form of supervised learning in which a system formulates a classifier for broadcast news using a few relevant stori...