Finding Related Events Based on Bursty Phrase Detection and Clustering

摘要

Wikipedia is known as the largest up-to-date online encyclopedia, in which articles are versioned and these edits are stored as revisions. In this paper we propose a new method to find related bursty edit events, based on detecting and clustering temporally significant phrases by their bursts over time, from revisions of articles. We discuss evaluation functions to find phrases that are semantically representative as well as temporally significant. After bursts are detected from the time series for each phrase, these phrases are clustered by their temporal similarities, using FastDTW. We evaluate how clustering quality is affected by the time resolution of FastDTW, and discuss optimum time resolution in terms of average burst duration. Experimental results show clustered phrases share similar burst patterns, which can be linked to related real-world events.