A Dataset for Misogyny Detection in Polish Social Media
Description
Harmful language detection is commonly evaluated on broadly defined hate speech and toxicity benchmarks, yet it remains unclear whether models that successfully detect harmful language in general can transfer this capability to more specific forms of hate. We investigate this problem for misogyny detection in Polish, an area for which dedicated resources remain scarce. We introduce a new dataset of Polish social media comments annotated for misogyny, gender-directed hateful speech. The dataset was constructed to capture explicit and implicit manifestations of misogynistic discourse in contemporary Polish social media. To this end, we collected user comments published under Facebook posts related to a highly publicized media controversy involving Anna Lewandowska, a prominent Polish public figure and influencer. The selected discussion threads generated substantial public engagement and contained a wide spectrum of reactions ranging from support to overt hostility, contempt, and stereotype-based criticism directed at women. The choice of this domain was motivated by two considerations. First, discussions concerning female public figures frequently trigger gender-oriented evaluations that extend beyond criticism of specific actions and instead invoke broader assumptions regarding women's social roles, expected behavior, and perceived responsibilities. Second, such discussions provide numerous examples of implicit forms of hostility that would be unlikely to be captured by traditional keyword-based approaches to hate speech detection. Following data collection, comments were anonymized and prepared for annotation. Each comment was assigned one of three mutually exclusive labels. Misogyny involves expressions of hostility, contempt, hatred, or deeply rooted prejudice directed at women as a social group or at an individual explicitly on the basis of her gender. Stereotyping refers to statements attributing generalized characteristics, behaviours, competencies, or social roles to women as a group; unlike misogyny, stereotype-based comments do not contain overt hostility. Neutral category comprises comments that do not contain misogynistic content, gender stereotypes, or other forms of gender-based derogation. To increase annotation reliability while reducing the costs associated with large-scale manual coding, the dataset was annotated independently by two frontier LLMs: GPT-5.5 and Grok-4.3. Both models received identical annotation guidelines and were instructed to assign one of the three predefined labels to each comment. Cases in which the models produced conflicting labels were subjected to expert adjudication who assigned the final label.
Files
Files
(70.3 kB)
| Name | Size | |
|---|---|---|
|
md5:139eecd0e669391085a33bd931f63ed2
|
70.3 kB | Download |
Additional details
Funding
- National Science Centre
- Decoding Persuasion 2025/57/N/HS1/00480