A Framework for the Computational Linguistic Analysis of Dehumanization

Julia Mendelsohn, Yulia Tsvetkov, Dan Jurafsky

Abstract

Dehumanization is a pernicious psychological process that often leads to extreme intergroup bias, hate speech, and violence aimed at targeted social groups. Despite these serious consequences and the wealth of available data, dehumanization has not yet been computationally studied on a large scale. Drawing upon social psychology research, we create a computational linguistic framework for analyzing dehumanizing language by identifying linguistic correlates of salient components of dehumanization. We then apply this framework to analyze discussions of LGBTQ people in the New York Times from 1986 to 2015. Overall, we find increasingly humanizing descriptions of LGBTQ people over time. However, we find that the label homosexual has emerged to be much more strongly associated with dehumanizing attitudes than other labels, such as gay. Our proposed techniques highlight processes of linguistic variation and change in discourses surrounding marginalized groups. Furthermore, the ability to analyze dehumanizing language at a large scale has implications for automatically detecting and understanding media bias as well as abusive language online.

Introduction

Dehumanization—the psychological process of perceiving others as "less than human"—is a precursor to some of society's worst outcomes, from systemic bias to genocide. While social psychologists have studied this for decades, quantifying it at scale has remained a challenge. Traditional sentiment analysis is often too blunt for this task; a sentence can have negative sentiment without dehumanizing its subject. Dehumanization is embedded in how we attribute (or deny) feelings, actions, and moral worth to groups. This paper bridges the gap between social science and Natural Language Processing (NLP), providing a robust toolkit to measure how the media "others" marginalized groups.

A Computational Framework for Dehumanization

The core contribution of this work is a framework that operationalizes four pillars of dehumanization from social psychology into measurable linguistic correlates. By mapping psychological concepts to NLP metrics, we can quantify multidimensional, subjective processes using large-scale automated methods.

Operationalization of dehumanization elements into NLP metrics.

The framework identifies four key elements:

  1. Negative Evaluation: Measured via paragraph-level valence (using the NRC VAD lexicon), connotation frames (tracking writer perspective), and word embedding neighbor valence.
  2. Denial of Agency: Measured using connotation frames for agency and "dominance" scores of a word's nearest neighbors.
  3. Moral Disgust: Quantified by calculating the cosine similarity between a group label's vector and a "Moral Disgust" concept vector (e.g., pervert, sin).
  4. Vermin Metaphors: Quantified by the cosine similarity between a group label and a "Vermin" concept vector (e.g., rat, infestation).

The researchers utilized diachronic word embeddings (Word2Vec) trained on yearly slices of data to observe how these associations shifted over three decades.

Case Study: LGBTQ Representation in the New York Times

We applied this framework to a corpus of New York Times articles from 1986 to 2015, totaling approximately 94,000 paragraphs and 7.36 million tokens. To ensure findings were not artifacts of general language change, "American" and "Americans" were used as an in-group control baseline.

The Pejoration of "Homosexual"

The most striking finding is the "stark divergence" between the labels gay and homosexual. While they denote the same group, homosexual has become an index for dehumanizing attitudes. In the NYT, it is significantly more associated with lower agency, lower valence, and higher similarity to "disgust" and "vermin" vectors than the word gay. Wilcoxon signed-rank tests confirmed that gay occurs in significantly more positive and high-agency contexts than homosexual ($p < 0.0001$).

Comparison of top associated words for 'gay' and 'homosexual' across three decades.

As shown in the table above, the linguistic neighbors for "homosexual" increasingly shifted toward clinical or pejorative terms like "deviant," "immoral," and "pedophilia," whereas "gay" moved toward terms associated with civil rights and social integration, such as "marriage," "equality," and "couples."

From 1986 to 2015, LGBTQ terms generally moved further away from "disgust" and "vermin" vectors and toward higher valence contexts, indicating a broad trend of humanization. However, this progress was not linear. In 1999, a spike in negative associations occurred. Qualitative analysis reveals this was due to intense reporting on the murder of Matthew Shepard. The media’s focus on the "rodent-like" or "homicidal" nature of the crime inadvertently embedded LGBTQ terms in more dehumanizing linguistic contexts.

Conclusion

This research demonstrates that dehumanization can be quantified at scale using computational methods. Human validation via Amazon Mechanical Turk confirmed that the model's proxy measures align with human intuition regarding dehumanizing language. By distinguishing between labels like gay and homosexual, we reveal how subtle lexical choices reflect and reinforce social biases. While the study is limited to a single prestigious newspaper, it provides a foundation for future work on social media and other marginalized groups. Understanding these linguistic patterns is crucial for detecting media bias and mitigating the impact of dehumanizing rhetoric in public discourse.

Summary updated

Abstract

Dehumanization is a pernicious psychological process that often leads to extreme intergroup bias, hate speech, and violence aimed at targeted social groups. Despite these serious consequences and the wealth of available data, dehumanization has not yet been computationally studied on a large scale. Drawing upon social psychology research, we create a computational linguistic framework for analyzing dehumanizing language by identifying linguistic correlates of salient components of dehumanization. We then apply this framework to analyze discussions of LGBTQ people in the New York Times from 1986 to 2015. Overall, we find increasingly humanizing descriptions of LGBTQ people over time. However, we find that the label homosexual has emerged to be much more strongly associated with dehumanizing attitudes than other labels, such as gay. Our proposed techniques highlight processes of linguistic variation and change in discourses surrounding marginalized groups. Furthermore, the ability to analyze dehumanizing language at a large scale has implications for automatically detecting and understanding media bias as well as abusive language online.
BibTeX
@inproceedings{mendelsohn2020framework,
  author = {Julia Mendelsohn and Yulia Tsvetkov and Dan Jurafsky},
  title = {{A Framework for the Computational Linguistic Analysis of Dehumanization}},
  booktitle = {Frontiers Artif. Intell},
  year = {2020},
  url = {https://openreview.net/forum?id=IXVRCxGeUEX},
}
Figure