Automated Assessment of the Aftermath of Typhoons Using Social Media Texts

Abstract

Disasters are one of the major threats to economics and human societies, causing substantial losses of human lives, properties and infrastructures. It has been our persistent endeavors to understand, prevent and reduce such disasters, and the popularization of social media is offering new opportunities to enhance disaster management in a crowd-sourcing approach. However, social media data is also characterized by its undue brevity, intense noise, and informality of language. The existing literature has not completely addressed these disadvantages, otherwise vast manual efforts are devoted to tackling these problems. The major focus of this research is on constructing a holistic framework to exploit social media data in typhoon damage assessment. The scope of this research covers data collection, relevance classification, location extraction and damage assessment while assorted approaches are utilized to overcome the disadvantages of social media data. Moreover, a semi-supervised or unsupervised approach is prioritized in forming the framework to minimize manual intervention. In data collection, query expansion strategy is adopted to optimize the search recall of typhoon-relevant information retrieval. Multiple filtering strategies are developed to screen the keywords and maintain the relevance to search topics in the keyword updates. A classifier based on a convolutional neural network is presented for relevance classification, with hashtags and word clusters as extra input channels to augment the information. In location extraction, a model is constructed by integrating Bidirectional Long Short-Time Memory and Conditional Random Fields. Feature noise correction layers and label smoothing are leveraged to handle the noisy training data. Finally, a multi-instance multi-label classifier identifies the damage relations in four categories, and the damage categories of a message are integrated with the damage descriptions score to obtain damage severity score for the message. A case study is conducted to verify the effectiveness of the framework. The outcomes indicate that the approaches and models developed in this study significantly improve in the classification of social media texts especially under the framework of semi-supervised or unsupervised learning. Moreover, the results of damage assessment from social media data are remarkably consistent with the official statistics, which demonstrates the practicality of the proposed damage scoring scheme.