Passage Feedback for News Tracking

摘要

We extend the event tracking task of Topic Detection and Tracking (TDT) to create a framework in which a user can highlight relevant passages in addition to specifying the relevance of documents. A dual framework of combined document and passage feedback improves performance over a state-of-the-art system without feedback by over 70%. Although annotators vary in the content and length of the passages marked for feedback, improvements in performance are consistent, We demonstrate how events in news follow certain trends over time, making passage feedback critical to event tracking.