Race, Ethnicity and National Origin-based Discrimination in Social Media and Hate Crimes Across 100 U.S. Cities

Kunal Relia, Zhengyi Li, Stephanie H. Cook, Rumi Chunara

Abstract

We study malicious online content via a specific type of hate speech: race, ethnicity and national-origin based discrimination in social media, alongside hate crimes motivated by those characteristics, in 100 cities across the United States. We develop a spatially-diverse training dataset and classification pipeline to delineate targeted and self-narration of discrimination on social media, accounting for language across geographies. Controlling for census parameters, we find that the proportion of discrimination that is targeted is associated with the number of hate crimes. Finally, we explore the linguistic features of discrimination Tweets in relation to hate crimes by city, features used by users who Tweet different amounts of discrimination, and features of discrimination compared to non-discrimination Tweets. Findings from this spatial study can inform future studies of how discrimination in physical and virtual worlds vary by place, or how physical and virtual world discrimination may synergize.

Introduction

While the existence of online hate speech and offline hate crimes is well-documented, research often focuses on high-profile, episodic events like elections or terrorist attacks. However, systemic discrimination is a persistent daily reality. This study addresses a critical gap: the lack of a systematic method to measure how different forms of online discrimination—specifically attacking others versus reporting personal experiences—correlate with physical violence at a local, city-wide level. By distinguishing between targeted discrimination (harassment) and self-narration (sharing personal experiences of victimhood), we provide a more nuanced tool for understanding structural discrimination.

Methodology

The researchers developed a multi-stage computational pipeline to process and analyze social media data at scale.

Computational pipeline for classifying and delineating discrimination tweets.
  1. Classification: Shallow neural networks identified "discrimination" tweets. To overcome the rarity of hate speech in general data streams, the team employed Active Learning, iteratively refining the model's decision boundary through human labeling of uncertain cases.
  2. Delineation: A specialized "Pronoun-checker" categorized tweets. Content featuring a majority of first-person pronouns (I, me, we) was classified as self-narration, while content attacking individuals or groups was classified as targeted.
  3. Linguistic Analysis: Using the EMPATH tool, the authors extracted affective and psychological states (e.g., "aggression," "sadness") and socio-economic markers (e.g., "work," "money").
  4. Statistical Modeling: Negative Binomial Regression was used to model hate crime counts, controlling for Census variables such as race, income, and population density to ensure the social media signal was not merely a proxy for demographics.

Datasets

The study integrated several large-scale datasets:

  • Twitter Data: 532 million tweets from a 1% sample of the public stream (2011–2016), filtered to 100 U.S. cities.
  • FBI Hate Crime Data: Statistics from 2011–2016 for the same 100 cities, focused on race, ethnicity, or national origin motivations.
  • U.S. Census Data: Demographic and socio-economic attributes from the American Community Survey (ACS).
  • Hatebase: Used to seed initial keyword expansion for the classification model.
  • Bot Lists: Seven public lists were used to identify and analyze the impact of automated accounts.

Key Insights

  • The Nature of Speech Matters: The total volume of hate speech is less predictive of offline violence than the ratio of targeted harassment to self-reporting. Cities where online discourse is dominated by attacks on others are more likely to experience higher rates of physical hate crimes.
  • The "Positive Emotion" Paradox: Surprisingly, "positive emotion" in discrimination tweets was a significant positive predictor of hate crimes. This suggests that hate speech often involves a sense of "bonding" or "celebration" within exclusionary groups rather than just raw anger.
  • Hyper-Local Hate: Significant geographic variation exists in discriminatory language. Specific derogatory terms related to local demographics or sports mascots appeared as top features only in certain cities, emphasizing the need for spatially-diverse training data.
  • Bot Influence: Bots are active participants in discriminatory discourse. In Washington D.C., nearly 31.6% of users discussing discrimination were identified as bots.

Key Results

The study's primary finding is that the ratio of targeted to self-narration discrimination is a significant predictor of hate crimes ($p < 0.001$), even when controlling for city-level demographics.

Distribution of the Targeted-to-Self-Narration ratio across 100 U.S. cities.

In 36 out of the 100 cities studied, targeted discrimination was more prevalent than self-narration. For instance, Woodbury, NJ exhibited the highest ratio of targeted speech (0.91), while West Jordan, UT had the lowest (0.02).

Linguistically, features like "night," "money," and "disappointment" in tweets correlated positively with higher hate crime rates, whereas "work-related" features showed a negative correlation. The classifier itself was highly effective, achieving an F1 score of 0.86 and an AUC of 0.90.

Conclusion and Future Work

While this study establishes a strong correlation, the authors note that social media discourse and hate crimes may both be manifestations of the same underlying social frictions rather than a direct causal chain. Future research could expand this work by incorporating non-English languages, analyzing emoji usage, and increasing granularity to the neighborhood level to better inform local policy and community interventions.

Summary updated

Abstract

We study malicious online content via a specific type of hate speech: race, ethnicity and national-origin based discrimination in social media, alongside hate crimes motivated by those characteristics, in 100 cities across the United States. We develop a spatially-diverse training dataset and classification pipeline to delineate targeted and self-narration of discrimination on social media, accounting for language across geographies. Controlling for census parameters, we find that the proportion of discrimination that is targeted is associated with the number of hate crimes. Finally, we explore the linguistic features of discrimination Tweets in relation to hate crimes by city, features used by users who Tweet different amounts of discrimination, and features of discrimination compared to non-discrimination Tweets. Findings from this spatial study can inform future studies of how discrimination in physical and virtual worlds vary by place, or how physical and virtual world discrimination may synergize.
BibTeX
@inproceedings{relia2019race,
  author = {Kunal Relia and Zhengyi Li and Stephanie H. Cook and Rumi Chunara},
  title = {{Race, Ethnicity and National Origin-based Discrimination in Social Media and Hate Crimes Across 100 U.S. Cities}},
  booktitle = {ICWSM},
  year = {2019},
  url = {https://openreview.net/forum?id=ysPZjDyuhFz},
}
Figure