Tell Me What You Know About Sexism: Expert-LLM Interaction Strategies and Co-Created Definitions for Zero-Shot Sexism Detection

Myrthe Reuver, Indira Sen, Matteo Melis, Gabriella LapesaNAACL 2025 · DOI

This paper, accepted for NAACL 2025, explores the intersection of human expertise and machine intelligence in the challenging domain of sexism detection. By pairing nine sexism researchers with Large Language Models (LLMs), the authors investigate whether "hybrid intelligence"—the collaboration between humans and AI—can produce better definitions and detection results than either could achieve alone.

TL;DR

Researchers teamed up with LLMs to co-create definitions of sexism, finding that while LLM-generated definitions surprisingly outperform concise expert-written ones in zero-shot detection, human-AI co-creation allows even LLM-novice experts to improve model performance. The study highlights that for complex social constructs, the "richness" of a definition matters more for model performance than traditional academic brevity.

The Challenge of Defining Sexism

Detecting sexism in text is notoriously difficult because it is a "complex social construct"—it is nuanced, context-dependent, and defined differently across cultures and disciplines. While LLMs are increasingly used to annotate such data, they are often treated as "black boxes."

The authors address a critical gap: How do domain experts actually interact with these models to assess their "knowledge"? And can the collaboration between a human expert (who understands the nuance) and an LLM (which understands its own internal "logic") produce better tools for social science research?

Methodology: A Four-Stage Pipeline

The study uses a novel four-component pipeline to explore this "hybrid intelligence":

  1. Survey: Nine experts (sexism researchers) reported their expertise and familiarity with LLMs.
  2. Interaction (Probing): Experts engaged in a multi-turn "chat" with GPT-3.5 to assess its knowledge of sexism using any strategy they chose (e.g., asking for examples or testing its biases).
  3. Co-Creation: Experts created three types of definitions: academic expert-written, LLM-generated, and co-created (expert-edited LLM output).
  4. Modelling: These 27 definitions were used as prompts for GPT-4o to perform zero-shot classification on 2,500 texts across five diverse English-language sexism benchmarks (CMSB, EDOS, Reddit Misogyny, EXIST, and HateCheck).
The four-stage pipeline used to explore hybrid intelligence between experts and LLMs.

Expert Interaction Strategies

The researchers developed a taxonomy of expert strategies, identifying how domain experts "stress-test" models. Rather than simple queries, experts used sophisticated methods such as "persona induction" (assigning the model an expert role) and "implicit tests" (e.g., asking the model to translate a sentence to see if it injects sexist tropes).

Taxonomy of strategies used by experts to probe LLM knowledge of sexism.

Key Insights and Ideas

  • The "Expertise Paradox": Expert-written definitions were significantly shorter (avg. 34 tokens) than LLM-generated ones (avg. 120 tokens). While brevity is a virtue in human academia, LLMs performed much better with the "wordier," more descriptive definitions they generated themselves.
  • LLMs as "Mirror Thinkers": LLM-generated definitions performed best on average (Mean F1 = 0.765). This suggests that LLMs might be better at following instructions written in their own "style" or conceptual framework.
  • Experience Doesn't Limit Collaboration: Interestingly, experts who were less experienced with LLMs were often more successful at co-creating effective definitions than their more "LLM-savvy" colleagues. This suggests that deep domain expertise is more valuable for hybrid intelligence than technical prompt engineering skills.

Key Results

  • Performance Gap: Expert-written definitions performed poorly (Mean F1 = 0.532) compared to LLM-generated (0.765) and co-created (0.762) ones.
  • Dataset Sensitivity: Results were highly dependent on the dataset. On the "CallMeSexist" dataset, all definitions performed well (>0.80 F1). However, on the "RedditGuest" dataset, performance fluctuated wildly, showing that some datasets require much more specific "expert-tuned" definitions to crack.

Limitations and Future Work

The study was limited to nine English-language experts and one primary social construct. Future work should expand this to other complex constructs like "populism" or "toxicity" and include non-Western (non-WEIRD) contexts to account for cultural nuances. Additionally, the reliance on proprietary models remains a challenge for reproducible social science.

Reference: Reuver, M., Sen, I., Melis, M., & Lapesa, G. (2025). Tell Me What You Know About Sexism: Expert-LLM Interaction Strategies and Co-Created Definitions for Zero-Shot Sexism Detection. Findings of the Association for Computational Linguistics: NAACL 2025.

Summary updated

BibTeX
@inproceedings{reuver2025tell,
  author = {Myrthe Reuver and Indira Sen and Matteo Melis and Gabriella Lapesa},
  title = {{Tell Me What You Know About Sexism: Expert-LLM Interaction Strategies and Co-Created Definitions for Zero-Shot Sexism Detection}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.470},
}
Figure