Exploiting Topic Pragmatics for New Event Detection in TDT-2004

Abstract

TDT-2004 marked Stottler Henke's second year of participation in the New Event Detection (NED) track of the Topic Detection and Tracking (TDT) evaluations. Our official entry this year consisted of three "pragmatics-based" classifiers operating in a majority voting framework. The system performed well, achieving by small margins the best optimized topic- and story-weighted CFSD scores for participating NED systems. We again validated the hypothesis that ensemble collections of classifiers can outperform the individual classifiers that compose them. Performance over the new TDT5 corpus was worse relative to previous corpora and overall accuracy within the NED community remains significantly below operationally desirable levels. We present a brief summary of our second year approach and a preliminary characterization of our performance results based on the experimental runs submitted to the TDT-2004 evaluation.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jan 01, 2004
Accession Number
ADA439316

Entities

People

  • Ronald K. Braun
  • Ryan Kaneshiro

Organizations

  • Stottler Henke Associates

Tags

Communities of Interest

  • Air Platforms
  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Abstracts
  • Accidents
  • Accuracy
  • Detection
  • Elections
  • Error Analysis
  • Errors
  • Event Detection
  • False Alarms
  • Frequency
  • Governments
  • Hurricanes
  • Judgment
  • Machine Learning
  • Semantic Models
  • Test And Evaluation
  • Warning Systems

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.
  • Team-Based Human-Centered Cognitive Task Decision Making and Information Performance.