Abstract:
Protests (or movements) are a form of collective sociopolitical action in which mem- bers with similar beliefs express their objections to a cause or situation. Often, a heated debate during protests on social media, such as Twitter, may lead to divided users and divergent discourse. On the bright side, studying divergent discourse on contentious topics can help infer the collective perceptions of people in terms of their collective narratives. On the dark side, narratives shared during protests may become susceptible to various harmful influence operations, disrupting society’s peaceful fabric. This thesis aims to understand digital strategies to organize protests, identify collective narratives shared during protests, and identify harmful behaviors with potential online and offline consequences. We focus on hate speech and coordinated inauthentic behavior as prox- ies for harmful conduct during online protests. We divide the thesis into four parts: (i) Understanding strategies used for conducting online protests, (ii) Detecting and analyz- ing collective narratives shared during protests, (iii) Detecting and analyzing opposing stances during the protest, inclusive of authentic and inauthentic actors, (iv) Detecting and analyzing harmful behavior during protest. To focus on the strategies used for conducting protests on social media, we examine the protest over the cause of the death of Sushant Singh Rajput (#SSR) 1 on Twitter. Study of shared hashtags and retweets during #SSR protests reveals a combination of centralized and decentralized information aggregation strategies in retweet networks, suggesting a mix of self-motivated individuals and organized entities. Next, we pro- pose an unsupervised clustering-based framework to focus on the collective narratives shared during protests. Our findings suggest clusters of call-to-action and on-ground activities across protests under study. Next, we delve into the opposing stances formed during an online protest, using #CAA protest on Twitter as a case study. We build an unsupervised stance detection technique to identify users’ stances and analyze their content, follower networks, and inauthentic behavior (i.e, bots, suspended users, and deleted users). Our findings suggest homophily (i.e., users of the same stance follow each other on Twitter) in follower network and presence of edges between authentic and inauthentic users, suggesting their connectedness. Finally, we focus on hate speech and coordinated inauthentic behavior as proxies for harmful conduct and study their contribution during the divergent discourse on #CAA. To this end, we built a multi-task classification model with hate speech detection as the primary task and stance detection as an auxiliary task and obtained an F1 score of 0.92. Our findings suggest that more hateful users produced more tweets, received faster retweets, and held a central position in the retweet network. Regarding coordinated inauthentic behavior, our findings sug- gest that coordinated communities, which were highly inauthentic, showed the highest clustering coefficient towards a greater extent of coordination. In conclusion, this thesis examines strategies, collective narratives, and harmful
behavior within protests, comprehensively exploring the intricate facets of online ac-
tivism. To summarize, the research contributions of the thesis are: - (i) Analyze protest
hashtags and retweet communities to provide insights into protest strategies, (ii) Build
an unsupervised collective narrative detection technique, (iii) Build an unsupervised
stance detection technique for user-level stance detection for multi-lingual Twitter data,
(iii) Build automated hate speech detection method for opposing stances, (iv) Build a
framework for coordinated communities in opposing stances. Through our research, we
aim to foster a more secure digital environment for participants in online protests.