Enabling Rapid Classification of Social Media Communications During Crises

摘要

The use of social media platforms such as Twitter by affected people during crises is considered a vital source of information for crisis response. However, rapid crisis response requires real-time analysis of online information. When a disaster happens, among other data processing techniques, supervised machine learning can help classify online information in real-time. However, scarcity of labeled data causes poor performance in machine training. Often labeled data from past event is available. Can past labeled data be reused to train classifiers? We study the usefulness of labeled data of past events. We observe the performance of our classifiers trained using different combinations of training sets obtained from past disasters. Moreover, we propose two approaches (target labeling and active learning) to boost classification performance of a learning scheme. We perform extensive experimentation on real crisis datasets and show the utility of past-labeled data to train machine learning classifiers to process sudden-onset crisis-related data in real-time.