Semantic-frame representation for event detection on Twitter

摘要

Unsupervised methods for detecting news events from tweet streams cluster feature representations via their burstiness, and filter out more news worthy clusters as outputs. Words, segments and tweets have been used as event feature representations, with segments being state-of-the-art due to their balance of expressive power and non-sparsity. However, segments do not convey structural event information, making output clusters difficult to understand. We investigate the use of semantic frame elements instead of segments as event features, observing not only better readability, but improvements in both precision and recall thanks to the effect of noise-filtering in frame extraction.