Tracking Factual Information in Evolving Text: An Empirical Study

We consider the problem of tracking information over time and across sources in clusters of news stories about emergency events. While previous approaches have focused on finding new information at the document and sentence levels (eg TDT FSD and the TREC Novelty track, respectively), we are interested in following information at the factual level. Specifically, we propose to develop a “fact tracking” system, that when given a user’s factual question of interest, returns a matrix displaying the extracted answers by time and source. As a first step towards this goal, we present an empirical analysis of a corpus of breaking news stories that have been manually annotated for factual questions and answers. Our current goal is to compare extracted answers to a given question and to examine how features such as lexical similarity and source information relate to the chronological and semantic relationships between them. Our study will show that while there appears to be no direct relationship between the lexical similarity and publication time difference of a given answer pair, lexical similarity is highly correlated to whether or not the answers come from the same sources, and whether or not they express the same or different answer to the given question.