TL;DR
This paper quantifies "aporophobia" – negative bias against the poor – in online discourse, finding it's disproportionately associated with low-wealth contexts. Using Bayesian modeling and Structural Causal Modeling on Reddit data, they demonstrate that low-wealth contexts can exacerbate the stigmatization of groups, particularly Black people, in online discussions.
What Problem Does This Paper Address?
While toxicity detection in online text is a growing field, "aporophobia" – prejudice against the poor – has been largely overlooked. This paper argues that aporophobia is not only a distinct form of toxicity but also a potential amplifier of broader group stigmatization. Existing toxicity datasets often lack sufficient examples of aporophobia, hindering its effective modeling. The authors aim to quantify the extent of aporophobia and investigate its role as an aggravating factor in the stigmatization of specific groups, using Black people as a case study due to their history of being targets of negative bias.
Methods
The researchers employ a two-pronged approach:
- Quantifying Aporophobia: They use BERTopic to identify emergent topics within Reddit comments. They then analyze the co-occurrence of specific n-grams (short phrases) indicative of stigmatizing topics with both low-wealth and high-wealth contexts. Bayesian estimation is used to calculate the relative propensity of these stigmatizing topics appearing in each wealth context.
- Investigating Aggravation: To assess aporophobia's role as an aggravating factor, they utilize Structural Causal Modeling (SCM). This framework allows them to simulate causal relationships between variables. They propose three generative regression models representing different causal assumptions about how references to a "people group" (specifically Black people), low-wealth contexts, and stigmatizing topics influence each other. By performing interventions on these simulated models, they can quantify the causal effect of low-wealth references on the rate of stigmatizing topics, especially when Black people are also mentioned.
Datasets
The primary dataset used is a subset of 266,268,920 public comments from Reddit, collected between January 2015 and May 2015.
- Low-wealth contexts were identified using n-grams like "poor people," "homeless," and "on welfare."
- High-wealth contexts were identified using n-grams like "the rich," "millionaires," and "billionaires."
- A control sample of comments not flagged for wealth context was also created.
- Comments referencing Black people were identified using both neutral n-grams (e.g., "Black people") and derogatory n-grams (e.g., "negro," "nigger").
The authors note they differ from a previous study (Kiritchenko et al., 2023) by using a different subset of Reddit data and slightly different n-grams for wealth contexts.
Key Insights and Ideas
- Aporophobia is real and context-dependent: The study provides quantitative evidence that certain stigmatizing topics are significantly more prevalent in discussions associated with low-wealth contexts compared to high-wealth contexts.
- Low-wealth contexts amplify existing biases: The research suggests that references to low-wealth status can act as an "aggravating factor," increasing the likelihood of negative associations with specific groups, even when controlling for other factors.
- Targeted stigmatization: The analysis reveals that the effect is not just about general association but can extend to directed stigmatization, where individuals from low-wealth backgrounds are more likely to be characterized by negative attributes.
- Data sparsity is a challenge: The authors highlight that identifying relevant low-wealth and Black people n-grams in the vast Reddit corpus is rare, which can make detecting subtle causal relationships difficult. This underscores the need for more targeted datasets for studying phenomena like aporophobia.
Key Results
- Relative Propensity: For several stigmatizing topics (mental illness, drug testing, addiction, heroin association), the relative risk ratio of their occurrence in low-wealth contexts versus high-wealth contexts was significantly greater than 1, with some intervals indicating a skew of 6.8 to 14.3 for drug testing.
- Directed Stigmatization: When focusing on directed stigmatization (where the topic is explicitly applied to individuals of a certain wealth status), the risk ratios were even higher, with intervals for mental illness and drug testing reaching 14.3 to 51.0 and 14.9 to 58.0, respectively.
- Aggravating Effect on Black People: For the specific case of Black people, the SCM analysis showed a weak but detectable aggravating causal effect of low-wealth references on stigmatizing topics like "prison, jail, prisons" and "police, cops, cop, police officer" under certain causal assumptions. However, the effect was found to be weak, requiring large interventions to become non-negligible. No significant causal relationship was found under one of the proposed causal models.
Limitations and Future Work
- Proxy Limitations: The study relies on n-grams as proxies for wealth contexts. While effective for identifying strong signals, these proxies might not capture more nuanced or ambiguous language related to low-wealth status.
- Data Sparsity: The low frequency of relevant n-grams in the dataset limits the statistical power to detect subtle causal effects and may explain why aggravation was only detected for a few specific topics.
- Future Work: The authors suggest extending this research to other dataset domains and incorporating modified feature sets from future human-annotated datasets specifically designed for aporophobia. This would enable a broader and more sensitive analysis of the phenomenon.
Summary updated