Reproduce “Social Chemistry 101: Learning to Reason about Social and Moral Norms” on OpenResearch

About this paper

Maxwell Forbes, Jena D. Hwang, Vered Shwartz, Maarten Sap, Yejin Choi

Researchers from the University of Washington and Allen Institute for AI introduce "SOCIALCHEMISTRY," a conceptual formalism for social and moral norms, and SOCIAL-CHEM-101, a dataset of 292k "Rules-of-Thumb" with over 4.5 million detailed annotations. This work provides a resource for AI systems to understand and reason about human social situations, demonstrating that current language models can learn to generate relevant norms and enabling analysis of societal biases in real-world text.

Key insights

• The dataset effectively captures a wide spectrum of social and moral judgments, demonstrating that a single situation can evoke multiple, sometimes conflicting, Rules-of-Thumb (RoTs).
• The multi-dimensional annotation scheme allows for distinguishing between social judgment, anticipated agreement, and cultural pressure, highlighting the nuance between personal morality and societal expectations.
• Analysis of annotator demographics revealed correlations between characteristics (e.g., gender, urbanness) and how agreement or judgment was ascribed, underscoring the subjective nature of social norms.

Summary by alphaXiv.