Using psychological theory to ground guidelines for the annotation of misogynistic language

TL;DR

This paper develops a psychologically grounded misogyny annotation scheme, creates a new dataset, and evaluates the performance of Large Language Models in replicating human annotations, highlighting the importance of theory-informed detection methods.

Contribution

It introduces a novel, psychology-based misogyny annotation guideline, a new annotated dataset, and a comparative analysis of LLMs versus human annotators.

Findings

Our guideline scheme outperforms existing coding schemes in classifying misogynistic texts.

LLMs struggle to replicate human annotations, reflecting mainstream views of misogyny.

The dataset achieves substantial inter-rater agreement (kappa = 0.68).

Abstract

Detecting misogynistic hate speech is a difficult algorithmic task. The task is made more difficult when decision criteria for what constitutes misogynistic speech are ungrounded in established literatures in psychology and philosophy, both of which have described in great detail the forms explicit and subtle misogynistic attitudes can take. In particular, the literature on algorithmic detection of misogynistic speech often rely on guidelines that are insufficiently robust or inappropriately justified -- they often fail to include various misogynistic phenomena or misrepresent their importance when they do. As a result, current misogyny detection coding schemes and datasets fail to capture the ways women experience misogyny online. This is of pressing importance: misogyny is on the rise both online and offline. Thus, the scientific community needs to have a systematic, theory informed coding scheme of misogyny detection and a corresponding dataset to train and test models of misogyny detection. To this end, we developed (1) a misogyny annotation guideline scheme informed by theoretical and empirical psychological research, (2) annotated a new dataset achieving substantial inter-rater agreement (kappa = 0.68) and (3) present a case study using Large Language Models (LLMs) to compare our coding scheme to a self-described "expert" misogyny annotation scheme in the literature. Our findings indicate that our guideline scheme surpasses the other coding scheme in the classification of misogynistic texts across 3 datasets. Additionally, we find that LLMs struggle to replicate our human annotator labels, attributable in large part to how LLMs reflect mainstream views of misogyny. We discuss implications for the use of LLMs for the purposes of misogyny detection.