Event detection based on on-line news clustering

Abstract

In this dissertation, we propose a semi-supervised learning method to solve the disaster event detection task. This method achieved lower miss rate and false rate than the unsupervised learning method on disaster event detection task. The semi-supervised learning method contains a multi-class classification model and a single-pass clustering model. A hierarchical single-pass clustering algorithm is also developed to overcome the deficiency of traditional single-pass clustering algorithm. For the text representation learning, a pre-trained BERT model is fine-tuned on our customized dataset, and achieves great performance on the classification problem. A NER model is introduced to extract the location and time features to help the clustering algorithm detect new events and track known events.