Mapping Moral Valence of Tweets

Overview

This project analyzes over 36,000 geo-located tweets in the aftermath of George Floyd's murder to uncover patterns of moral expression within the Black Lives Matter movement.

Using FrameAxis and the extended Moral Foundations Dictionary (eMFD), tweets are embedded in a moral space to explore the moral dimensions activated in online discourse.

Motivation

• Social media can catalyze moral alignment and social change.
• Moral evaluations—especially outrage—drive group coordination and identity signaling.
• We use the MAD model (Motivation, Attention, and Design) and Moral Foundations Theory to guide analysis.

Method

The eMFD approach was developed to map language documents to moral frames, defined by a set of 5 distinct moral axes. GitHub Code Repository

Foundation (virtue/vice)	Description (Virtuous Direction)	Example Keywords
Authority/ Subversion	Desire/need for beneficial relationships with hierarchies in society.	respect, duty, obedience, rebellion, tradition
Care/ Harm	Compassion towards victims and the vulnerable, anger towards those perpetrating injustice and harm.	empathy, suffering, protect, violence, kindness
Fairness/ Cheating	Desire for cooperation and gratitude for just and trustworthy systems and people.	justice, equality, rights, fraud, fairness
Loyalty, Ingroup/ Betrayal	Desire for cohesive groups. Instantiates group pride and anger at traitors.	patriotism, unity, team, betrayal, allegiance
Sanctity, Purity/ Degradation	Relevant virtues consist of being temperamental and pious and clean.	purity, sacred, clean, disgusting, holy

Adapted from Moral Foundations Theory (Graham et al., 2013), which proposes that moral reasoning is structured by distinct psychological foundations.

Data

• 36,282 tweets geo-located to Los Angeles (Feb–Aug 2020)
• 30,223 tweets had at least one moral term
• Breakdown: 26,827 BLM | 1,578 All/Blue Lives | 1,818 Defund Police

Key Findings

• Most moral content focused on Care, Harm, and Fairness violations (in line with MAD model predictions).
• Ingroup Loyalty emerged strongly, highlighting community solidarity.
• Clustering revealed multiple moral subthemes within BLM discourse.

Discussion & Future Directions

• FrameAxis + eMFD pipeline effectively identifies nuanced moral patterns.
• Short tweets and clustering limitations present challenges.
• Future work: guided topic modeling and longer text corpora.