Detecting environmental disasters in digital news archives

Automatically extracting events from large, unstructured/semi-structured textual data requires a mechanism for identifying the event, abstracting it from the text, validating the event's occurrence against some known values, and sharing the event with users effectively. Inherent in the challenge of Big Data is that it often exceeds a scale at which humans can effectively operate. In this paper, we focus on the domain of archived newspaper articles, and describe a system that generates a collection of event summaries from unstructured text, extracts a geographic marker for the event, and stores both in an on-line database that can be searched and/or visualized using an interactive map. The system relies on text mining techniques to filter out a dataset of news stories from a digital news archive source and extracts 1-2 sentences from each event to be stored in the database. We illustrate this approach using a flood database case study, automatically extracting descriptions of past flooding events occurring in Nova Scotia, Canada from a 20-year archive of regional newspaper articles. We validate our event extraction in two dimensions (identification of articles mentioning flood events; identification of accurate geographic markers from articles about flood events) using Amazon's Mechanical Turk (MTurk) to obtain human assessments at scale.