Rag improves explanations for detecting hidden misogyny in english and italian
RAGing Against Misogyny: Improving Implicit Misogyny Explanations
Computation and Language
Summary
Understanding hidden or indirect misogyny can be difficult because it often needs extra background knowledge. The authors propose a method that brings in information from encyclopedias, slang, and feminist writings to help detect and explain this kind of hidden misogyny in both English and Italian. They found that using feminist literature was especially helpful in identifying and explaining implicit misogyny. While their approach improved the detection of implicit misogyny in one dataset, it did not improve detection in another. However, people preferred explanations generated with this additional knowledge.
What this means in practice
- •For content moderation teams: Enhance detection and explanation of subtle misogynistic content using integrated feminist and encyclopedic knowledge.
- •For social media platform developers: Develop tools that provide clearer explanations to users when implicit misogyny is flagged in posts across English and Italian languages.
Authors
Ivana Crescenzi, Arianna Muti, Maarten Sap, Tommaso Caselli, Debora Nozza
Abstract
Implicit misogyny often requires background knowledge to be understood. To address this challenge, we propose a Retrieval-Augmented Generation (RAG) framework for detecting and explaining implicit misogyny in English and Italian, leveraging encyclopedic, slang and feminist knowledge sources. Using the ImplicIT-Mis and SBIC+ datasets, we evaluate the contribution of each knowledge source to detection and explanation quality, investigating also whether RAG addresses known LLM failure modes in implicit misogyny. Results show that RAG improves recall on ImplicIT-Mis, but does not improve detection performance on SBIC+. However, human evaluation shows that RAG-based explanations are preferred over no-RAG explanations in both languages. Feminist literature emerges as the most consistently useful source for detection and explanation.