Biasly: An Expert-Annotated Dataset for Subtle Misogyny Detection and Mitigation

Using novel approaches to dataset develop-001 ment, the Biasly dataset captures the nuance 002 and subtlety of misogyny in ways that are 003 unique within the literature. Built in collab-004 oration with multi-disciplinary experts and an-005 notators themselves, the dataset contains anno-006 tations of movie subtitles, capturing colloquial 007 expressions of misogyny in North American 008 film. The open-source dataset can be used for 009 a range of NLP tasks, including binary and 010 multi-label classification, severity score regres-011 sion, and text generation for rewrites. In this 012 paper, we discuss the methodology used, an-013 alyze the annotations obtained, provide base-014 lines for each task using common NLP algo-015 rithms, and furnish error analyses to give in-016 sight into model behaviour when fine-tuned on 017 the Biasly dataset. 018 Content Warning: To illustrate examples from 019 our dataset, misogynistic language is used 020 which may be offensive or upsetting. 021 1 Introduction and Related Work 022 When using language models (LMs) to perform 023 sensitive, subjective, and socially impactful tasks 024 like misogyny detection, hate speech mitigation, or