Request Identification on Twitter during Mass Convergence

Abstract

Catastrophic events create uncertain environments in which it becomes very difficult to locate affected people and provide them aids. People turn to Twitter during disasters for requesting help/providing relief to others than their friends and family. Some applications exist to facilitate people in such events, however; they follow very rigid formats and templates for posting tweets. Both affected citizens and volunteers mostly do not strictly follow these formats for seeking and providing help. A huge number of posts issued online for seeking help could not properly be detected and remained concealed. A proportion of these posts, such as a request for seeking help and relief, need instant attention. Affected citizens make requests for different types of things, e.g., medical, volunteer, cloth, food, shelter, money. This research aims to delve into tweets for extracting different types of request tweets (so-called rweets) in order to provide humanitarian relief to both seekers and responders. In this work, the problem of automatic rweet identification and categorization on Twitter is addressed. For rweet identification, both rulebased and machine learning approaches are utilized while for rweet categorization, only machine learning approach is employed. An empirical study comprising of a large number of tests is performed to assess the problem from different aspects. For rweet identification, precision of 99.72% achieved using rule-based approach and F-measure of 82.35% using logistic regression in machine learning approach. Logistic regression also outperformed in rweet categorization by gaining excellent F-measure of 94.95%. This research also provides the practical guidance to machine learning practitioners and professionals while solving and assessing classification problems considering different types of aspects. This study is limited to the use of tweets contents as features. Although it is very powerful approach for extracting information, but highly dependent upon manually labelled dataset for a specific disaster.