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—perceiving others as "less than human"—is a psychological precursor to intergroup bias, hate speech, and even systemic violence. While social psychologists have studied this for decades, their work is typically small-scale and qualitative. The challenge for the NLP community has been quantifying such a subjective, multidimensional process. Most "bias detection" tools focus on simple sentiment (positive vs. negative), but dehumanization is more complex; it involves stripping people of their ability to plan (agency), associating them with filth (disgust), or comparing them to non-human entities (metaphors). This research bridges the gap between social psychology and computational linguistics to measure these subtle signals across millions of words.

Methodology

The framework operationalizes four key components of dehumanization identified in social psychology literature. These components and their corresponding computational methods are summarized in Figure 1.

Operationalization of dehumanization elements.
  1. Negative Evaluation: Measured via paragraph-level valence (using the NRC VAD lexicon) and word embedding neighbor valence.
  2. Denial of Agency: Measured using "Connotation Frames" (which identify if a verb implies the subject has control) and embedding neighbor dominance.
  3. Moral Disgust: Quantified by calculating the cosine similarity between a group's word vector and a "Moral Disgust" concept vector (built from words like disgust, sin, pervert).
  4. Vermin Metaphors: Quantified by cosine similarity to a "Vermin" concept vector (built from words like rodent, cockroach, flea).

The researchers used diachronic word embeddings (word2vec models trained for each year) to track how these associations shifted from 1986 to 2015. They also utilized a "control" group—the term American(s)—to ensure findings weren't just artifacts of general language change in the newspaper.

Data and Experimental Setup

The study utilized the LGBTQ Corpus, a subset of New York Times articles (1986–2015) containing 93,977 paragraphs and 7.36 million tokens, filtered for specific LGBTQ-related terms. Lexicons used included the NRC VAD (Valence, Arousal, Dominance) Lexicon, Rashkin et al.’s Connotation Frames, and the Moral Foundations Dictionary. To validate the framework, 120 paragraphs were annotated by workers on Amazon Mechanical Turk to ensure vector-based measures aligned with human perception of dehumanization.

Results

The analysis revealed several key insights:

  • General Humanization: Overall, LGBTQ groups became more associated with positive valence and higher dominance/agency over the 30-year period.
  • The "Homosexual" vs. "Gay" Divergence: Despite being denotationally similar, the two labels have moved in opposite directions. Gay has become more associated with civil rights and identity, while homosexual has become increasingly "pejorated," appearing in contexts involving sexual deviancy, crime, and moral disgust. In 2015, the nearest neighbors for homosexual included pedophilia, bestiality, infanticide, and abhorrent, whereas gay was associated with marriage, equality, and unmarried.
  • Agency Gap: LGBTQ groups consistently experienced greater "denial of agency" compared to the American(s) control group across the entire 30-year span.
  • Validation: Human evaluators confirmed the model's findings: paragraphs flagged as "nearest to the Vermin/Disgust vectors" were indeed rated as significantly more dehumanizing.

Discussion and Future Work

The study demonstrates that dehumanization is not just "negative sentiment." A paragraph can contain positive words but still be dehumanizing if it denies agency or treats a group as a passive object of debate. Interestingly, dehumanizing language in the NYT was often driven by quoted sources (e.g., politicians) rather than journalists, though providing a platform for such views still shapes the media environment.

Future work should address current limitations, such as the reliance on word-level signals rather than complex syntax or discourse structures. Moving from static word2vec embeddings to contextualized embeddings like BERT could better capture nuances. Additionally, expanding the study beyond the New York Times and the English language would provide a more global understanding of dehumanization.

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