Reading between events : exploring the role of machine understanding in event tracking

Abstract

Humans observe, humans understand, and then, humans act, but machines only act. The Topic Detection and Tracking (TDT) community realised, early on, that to accomplish its task, to detect and track events from the news media, it would not suffice to act without understanding. Yet the TDT community rarely sought to understand. Therefore when Twitter modernised the task, now to detect and track events from social media, researchers had no response to the new challenges: the volume and velocity, the brevity and the noise. Today, we ask more of our TDT algorithms. We demand that they detect events precisely, describe comprehensively and model formally. We demand that they meet our modern needs without answers to the questions posed by understanding. What does it mean to understand events? How can we understand events? How can understanding improve TDT? In this dissertation, we answer the three questions. We debate interpretations of understanding and adopt a structured, semantic definition of events: Who does What, Where and When. We develop DEPICT, a novel algorithm to understand Who participates in events and Where, and EVATE to understand What can happen from past events. And with understanding, we propose SEER, a novel TDT algorithm that tracks events with increased precision, coverage and sensitivity, and which drives a novel and simplified event modeller. In the end, we demonstrate that understanding remains a worthwhile ambition. Machines can observe and understand, and when they do, they act better