Reproduce “Social Bias Frames: Reasoning about Social and Power Implications of Language” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maarten Sap, Saadia Gabriel, Lianhui Qin, Dan Jurafsky, Noah A. Smith, Yejin Choi
SOCIAL BIAS FRAMES extends harmful-language analysis beyond toxicity labels by representing offensiveness, intent, targeted groups, in-group usage, and the stereotypes or power relations implied by a statement. In the 44,671-post Social Bias Inference Corpus, GPT-2-based models reached about 70–77 F1 on broad classification tasks and generated group descriptions more reliably than context-specific implied stereotypes.
Key insights
- Broad judgments are easier to recover than nuanced pragmatic meaning: GPT-2 achieved test F1 scores of 70.2 for offensiveness, 69.2 for intent to offend, 77.3 for lewdness, and 69.4 for group targeting.
- Models identified targeted groups substantially better than the associated implied stereotypes, often substituting a generic group stereotype for the specific implication supported by the post.
- In-group language remained especially difficult, with the best reported development-set F1 reaching only 16.3, highlighting both severe class imbalance and the importance of speaker identity and context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.