- This event has passed.
Ph.D. Proposal: Hieu Nguyen
November 2, 2022 @ 2:00 pm - 3:00 pm EDT
Title: A Unified, Comprehensive Framework to Study Radical and Offensive Content on
Twitter During the Covid-19 Pandemic
Student: Hieu Nguyen
Major Advisors: Joseph Johnson and Swapna Gokhale
Associate Advisors: Dongjin Song and Shiri Dori-Hacohen
Date/Time: Wednesday, November 2nd, 2022, 2:00 PM
Location: WebEx
Remote Access: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mbbf4faa1c12d3a60820ce0d7d36ae205
Meeting number: 2623 990 0505
Password: P9Dt4iyM6Na
Abstract:
Social media platforms have been used to spread radical and offensive content since their advent. This trend was exacerbated during the Covid-19 pandemic, when people experienced a dramatic loss of their routines and social isolation, due to the need to physically distance themselves from others to curb the spread of the coronavirus. Extreme and hateful posts can incite violence and social unrest, even more so during the turbulent and chaotic circumstances brought about by the pandemic. Identifying and demoting such content is then necessary to mitigate their damage. This cannot be undertaken manually due to the large volumes of information posted on these platforms, calling for automated detection approaches. Most of the current approaches devised to detect such content are optimized for a specific topic and a theme. However, questionable content during the Covid-19 pandemic took many forms – it spread anti-government sentiment undermining the faith in public health measures, stirred up racial and social tensions, and insulted those with differing views. To facilitate such automatic detection, and understand how this content spreads, this dissertation develops a unified approach. The framework is demonstrated on data collected and annotated around three key topics – anti-lockdown protests in the state of Michigan, Proud Boys protests in Portland, and debates surrounding the use of facemasks. From the tweet data returned by the API, we curate a rich collection of features extracted not only from the text of the tweets, but also from other metadata including the authors’ activities and biographies and parameters measuring interaction with the tweets. These features are used to train an extensive selection of conventional machine learning and deep learning models, handling class imbalance when necessary. Feature engineering methods based on importance analysis, statistical significance, and principal components are used to improve model performance. Lastly, we conduct an in-depth study of each event to understand the root, the prominent influencers, and how such information is spread using network analysis.
