Making Sense of Social Media Text and the Spread of Rumours in Online Social Networks - An Interdisciplinary Approach

Abstract

As the spread of rumours in online social networks (OSNs) has grown at an alarming pace, there is a growing need to better understand the social and technological processes behind this trend. This research proposes an interdisciplinary approach to study the effects of rumours in OSNs, with the end goal of developing and validating a set of interactive visualization models that will help researchers as well as members of various OSNs to detect and prevent the rapid spread of rumours in these networks. The strength of the proposed research is that it adopts an interdisciplinary approach to study the phenomenon by integrating valuable insights from different established disciplines, such as sociology, psychology, information science, and computer science, to create a holistic view and understanding of how rumours are spread in OSNs.

The thesis first studies the impact of short and noisy nature of social media text, which could significantly affect the performance of Natural Language Processing (NLP) systems. We introduce a new terabyte-scale corpus that is created from Reddit comments from Oct 2007 to Aug 2016 and propose a novel approach to compute the semantic similarity between social media texts. The proposed semantic similarity algorithm will alleviate the inherent limitation of social media texts and improve the results of NLP systems using social media data. Then, we propose a visual framework to detect and cluster memes in OSNs. Our algorithms could conclusively identify the emerging and trending memes in OSNs. After discovering memes, we propose a visualization framework for collecting, analyzing, and visualizing memes and rumours in OSNs using theories rooted in psychology, sociology, information science, and computer science. This framework allows end users to collect data about a specific rumour and see its spread pattern, topics over time, sentiment analysis, and user interaction graphs. Using established psychological theories, we classify users based on how they interact in a rumour. Finally, we try to detect the truth of rumours based on selective feature sets that are derived from the proposed visualization tool and established social science theories.