Analysis of Social Media Messages for Disasters via Semi Supervised Learning

Abstract

Automated analysis of social media messages about social disturbances and natural disasters is important for managing relief and rescue work. This paper proposes a new method that uses semi supervised training approach to analyze social media messages about disasters. Compared to fully supervised methods, the approach needs a smaller number of messages to be hand labeled. The social media messages are analyzed with term frequency vectors that are later fed to SVM and logistic regression based machine learning methods. The training dataset is grouped into online and offline messages that makes the semi supervised learning even more effective. The experiments performed on the Twitter messages provided promising validation data towards the employment of the system in practical applications. The current work is applied only to earthquake messages but it can be extended for other types of disasters and social disturbances.