Protest Event Data from Geolocated Social Media Content

Abstract

While it is understood that protester identity, violence, and emotions affect the size of protests, these concepts have proved difficult to measure at the protest-day level. Geolocated text and images from social media can improve these measurements. This advance is demonstrated on protests in Venezuela and Chile; it uncovers more protests in Venezuela and generates new measures in both countries. Moreover, the methodology generates daily city-day protest data in 107 countries containing 82.7% of the world’s population and 97.15% of its GDP. These multimodal protest event data complement existing event datasets, though countries’ population and income constrain the reach of any methodology.

Analyzing Text and Images in Digital Communication: The Case of Securitization in American White Supremacist Online Discourse

“…The former task assesses content; it detects and labels objects in images, finds people and their faces, infers humans’ attributes (e.g., ethnicity, age) and expressions (e.g., emotion, ideology), and so on. It is usually done by building on the manual coding of concepts within and across observations (e.g., Casas and Williams 2019; Trilló and Shifman 2021; van Haperen, Uitermark, and van der Zeeuw 2020), using machine learning (ML) methods (e.g., Cantú 2019; Steinert-Threlkeld, Chan, and Joo 2021; Xi et al 2020; Zhang and Pan 2019), or a combination of the two, such as by first doing manual coding to uncover concepts—and examples of these concepts—that are then used in ML-driven analysis (e.g., Steinert-Threlkeld and Joo 2020; Steinert-Threlkeld et al 2021; van Haperen et al 2020). 2 After identifying and measuring relevant dimensions of images, researchers usually then classify all the images in the database using a framework relevant to their research questions.…”

An analysis of emotions and the prominence of positivity in #BlackLivesMatter tweets

“…However, measuring emotions is nontrivial, and computational models that overestimate expressions of emotions, like “anger,” can reinforce negative stereotypes. Previous examinations of emotions and affect expressed in tweets about the Black Lives Matter movement have relied on lexicon (Linguistic Inquiry and Word Count, LIWC) scores ( 3 ), and analyses of other protest events have similarly relied on lexicon-based approaches ( 15 , 16 ). While recent research has led to the development of more powerful deep learning–based models and annotated datasets, these models nevertheless are prone to overfitting to shallow lexical cues and often perform poorly in new domains ( 17 – 19 ).…”