The Crime of Being Poor: Associations between Crime and Poverty on Social Media in Eight Countries

Georgina Curto, Svetlana Kiritchenko, Kathleen Fraser, Isar NejadgholiNAACL 2024

TL;DR

This paper uses large language models (LLMs) to analyze social media discussions about the association between poverty and crime across eight countries. It finds that such discussions are significantly more prevalent in the US and Canada, often framing the poor as criminals, while other countries focus more on systemic issues and unfair laws. This highlights a critical social bias that may hinder poverty reduction efforts.

What Problem Does This Paper Address?

The paper tackles the persistent and harmful stereotype that links poverty with criminality. This "crime-poverty association" (CPA) can manifest as public bias (crime-poverty bias) or as institutional actions that criminalize poverty. Such associations can lead to discrimination, hinder effective poverty reduction policies, and perpetuate a cycle of disadvantage for vulnerable populations. While the correlation between poverty and crime is a known phenomenon, public perception often overestimates this link, leading to unfair judgments and policies. The authors aim to quantify and understand the prevalence and nature of these discussions on social media, a rich source of public discourse.

Methods

The researchers employed a multi-stage approach using Natural Language Processing (NLP) techniques:

  1. Data Collection: They collected a large corpus of English-language tweets from eight geographically diverse countries.
  2. CPA Identification: They used pre-trained Large Language Models (LLMs) – specifically Llama 2, Flan-T5, and ChatGPT – to classify tweets as either discussing a crime-poverty association (CPA) or not. This was framed as a binary text classification task.
  3. Model Evaluation: To ensure accuracy, they manually annotated a small test set of tweets and evaluated the LLMs' performance using metrics like accuracy and F1-score. ChatGPT, with a few-shot prompt, emerged as the best performer.
  4. Topic Modeling: For the tweets classified as CPA, they used BERTopic, an unsupervised topic modeling technique, to identify common themes and discourse patterns within these discussions.
  5. Contextualization: Finally, they compared their findings with existing socio-economic and crime indicators for each country to provide context.

Datasets

  • Twitter Corpus: A dataset of over 1.3 million tweet sentences collected from users in eight countries (United States, United Kingdom, Canada, India, Nigeria, Australia, South Africa, Kenya) between August and November 2022.
  • Manually Labeled Datasets:
    • 'Random' Test Set: 400 sentences randomly sampled from the corpus.
  • 'Crime' Test Set: 400 sentences specifically selected to increase the proportion of CPA examples, by including tweets with crime-related keywords.
  • Development Set: A smaller set created similarly to the 'Crime' set for model development.

These datasets were used for evaluating LLM performance and for topic modeling.

Key Insights and Ideas

  • US/Canada Discrepancy: The most striking finding is the significantly higher prevalence of CPA discussions in the US and Canada compared to other countries, even those with higher poverty and crime rates.
  • Nature of Bias: In the US and Canada, CPA discussions tend to focus on "crime-poverty bias," portraying the poor (especially the homeless) as criminals and drug addicts, and advocating for their removal.
  • Systemic vs. Individual Blame: In contrast, discussions in countries like Nigeria, Kenya, and India more frequently highlight systemic issues, such as unfair laws and policies that disproportionately affect the poor, rather than blaming individuals.
  • "Land of Opportunity" Narrative: The authors hypothesize that the "land of opportunity" narrative prevalent in the US and Canada might contribute to blaming individuals for their poverty, leading to a stronger association with criminality.
  • Interconnectedness of Bias: The paper emphasizes how public bias (crime-poverty bias) and institutional actions (criminalization of poverty) reinforce each other, creating a vicious cycle.

Key Results

  • Approximately 8% of the analyzed tweets (38,034 sentences) were classified as discussing crime-poverty associations.
  • The United States showed the highest percentage of CPA sentences (9.1%), followed by Canada (7.6%). India had the lowest (2.6%).
  • Topic modeling revealed distinct themes: US/Canadian discourse often centered on homeless people as criminals (Topic 1), homelessness being illegal (Topic 3), and calls to remove homeless people (Topic 4).
  • Discourse in other countries frequently focused on discriminatory laws and policies (Topics 5 & 6) and the unjust imprisonment of the poor (Topic 9).

Limitations and Future Work

  • Language and Platform Bias: The study is limited to English-language tweets from Twitter, which may not represent the views of the entire population or capture discussions in other languages or on different platforms. Twitter's user base also has demographic biases.
  • Preliminary Analysis: The authors acknowledge this is a preliminary analysis and call for deeper exploration into specific stereotypes and discriminatory acts.
  • Future Directions: The paper suggests extending the analysis to other languages, geographical locations, and social media platforms. Incorporating multi-modal data (images, videos) could also enrich the understanding of poverty and criminality discourse. The ultimate goal is to inform poverty mitigation strategies that address not only wealth redistribution but also the mitigation of social bias and discrimination.

Summary updated

BibTeX
@inproceedings{curto2024crime,
  author = {Georgina Curto and Svetlana Kiritchenko and Kathleen Fraser and Isar Nejadgholi},
  title = {{The Crime of Being Poor: Associations between Crime and Poverty on Social Media in Eight Countries}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure