Detecting events in a million New York times articles


We present a demonstration of a newly developed text stream event detection method on over a million articles from the New York Times corpus. The event detection is designed to operate in a predominantly on-line fashion, reporting new events within a specified timeframe. The event detection is achieved by detecting significant changes in the statistical properties of the text where those properties are efficiently stored and updated in a suffix tree. This particular demonstration shows how our method is effective at discovering both short-and long-term events (which are often denoted topics), and how it automatically copes with topic drift on a corpus of 1 035 263 articles.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)