Understanding and Countering Stereotypes: A Computational Approach to the Stereotype Content Model

Kathleen C. Fraser, Isar Nejadgholi, Svetlana KiritchenkoACL 2021

TL;DR

This paper introduces a computational framework for understanding and categorizing stereotypes based on the psychological Stereotype Content Model (SCM). By mapping stereotypes into a 2D space of warmth and competence, the authors demonstrate a method for analyzing existing stereotypes and lay the groundwork for generating anti-stereotypes to combat biased thinking.

What Problem Does This Paper Address?

Stereotypes, even seemingly positive ones, can have harmful consequences by dictating roles and limiting individuals. While psychology has long studied stereotypes, computational approaches to understanding and actively countering them in text are nascent. Existing work often focuses on debiasing word embeddings or detecting abusive content. This paper aims to bridge this gap by computationally modeling the widely-accepted Stereotype Content Model (SCM) to analyze textual stereotypes and explore strategies for generating anti-stereotypes, a promising method for reducing biased thinking. A key challenge is the ill-defined nature of "anti-stereotypes" – are they simple antonyms, or something more nuanced?

Methods

The core of the paper is the development of a computational model for the Stereotype Content Model (SCM). This involves:

  • Mapping to Warmth-Competence Space: Using the POLAR framework, the authors project word embeddings into a 2D space defined by warmth (sociability, morality) and competence (ability, agency). This is achieved by defining "polar directions" based on seed words associated with positive and negative warmth and competence.
  • Lexicon Validation: The model's ability to accurately place words along these dimensions is validated using an extended lexicon of words annotated for warmth and competence.
  • Stereotype Analysis: Stereotypes from the StereoSet dataset are then projected into this 2D space to analyze their distribution and compare them with findings from psychological literature.
  • Anti-Stereotype Analysis: The paper analyzes human-generated anti-stereotypes from StereoSet, investigating strategies annotators used (e.g., direct antonym, opposite quadrant, flipping one dimension) and how these relate to the original stereotypes.

Datasets

  • StereoSet: This dataset, specifically the development set, was used for analyzing human-generated stereotypes and anti-stereotypes. It contains stereotype and anti-stereotype word pairs for 79 target groups across gender, race/nationality, profession, and religion.
  • Nicolas et al. (2020) Lexicon: This supplementary data, containing seed words and an extended lexicon of words associated with warmth and competence, was used for validating the computational model.

Key Insights and Ideas

  • SCM as a Computational Framework: The paper successfully demonstrates that the psychological Stereotype Content Model can be computationally modeled, allowing for the analysis of stereotypes in a quantifiable 2D space.
  • Nuance in Anti-Stereotypes: Human-generated anti-stereotypes are not simply direct antonyms of stereotypes. Annotators often employ more complex strategies, such as flipping one dimension of the stereotype or choosing words that represent a different facet of a multi-dimensional stereotype.
  • Ambivalent Stereotypes are Key: The paper highlights that ambivalent stereotypes (e.g., "X but Y") are particularly important. Countering the "Y" aspect of these stereotypes, rather than just the primary "X" aspect, can lead to more realistic and effective anti-stereotypes.
  • Computational Analysis Mirrors Psychology: The computational model's projection of target groups onto the warmth-competence plane largely aligns with findings from decades of psychological survey research, validating the approach.

Key Results

  • Model Validation: FastText embeddings achieved the highest accuracy (up to 85.8%) in predicting warmth and competence labels from the extended lexicon.
  • Stereotype Mapping: The computational approach successfully mapped 67% of stereotypes to their predicted quadrants in the warmth-competence plane, aligning with psychological literature. Deviations were often explainable upon closer examination.
  • Anti-Stereotype Strategies: While direct antonyms were used 23% of the time for individual stereotype-anti-stereotype pairs, "opposite quadrant" strategies were more common (29.6%). For group-level anti-stereotypes, 45% of means fell into the opposite quadrant.
  • Ambivalent Countering: The paper proposes that for ambivalent stereotypes, countering the negative aspect (e.g., "women are nurturing but weak" -> "women are nurturing and strong") is more effective than simply negating the primary stereotype.

Limitations and Future Work

  • Dataset Limitations: The StereoSet dataset is English-only and relies on crowd-sourced data from US workers, potentially not representing the views of all subpopulations or marginalized groups. The publicly available portion is also limited.
  • Ambivalent Stereotype Complexity: The paper acknowledges that the process of generating anti-stereotypes is complex and not easily predicted by simple rules.
  • Future Work:
    • Larger and Diverse Datasets: Expanding analysis to include data from a wider range of social and cultural groups, and potentially collecting stereotypes "in the wild" from social media.
  • Sentence-Level Analysis: Extending the framework to analyze stereotypes at the sentence level and consider larger conversational contexts.
  • Identifying Stereotype Clusters: Developing methods to identify multiple clusters of stereotypes within a single target group.
  • Formal Evaluation of Anti-Stereotypes: Conducting controlled psychological studies to evaluate the effectiveness of different anti-stereotype generation strategies in reducing bias.
  • Human-in-the-Loop Systems: Emphasizing the importance of human oversight in any technology-driven anti-stereotyping to mitigate risks of introducing new biases.

Summary updated

BibTeX
@inproceedings{fraser2021understanding,
  author = {Kathleen C. Fraser and Isar Nejadgholi and Svetlana Kiritchenko},
  title = {{Understanding and Countering Stereotypes: A Computational Approach to the Stereotype Content Model}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure