2026-09-24
Zero-shot emotion scores plus symbolic regression yield seven short rules: 68% accuracy on Moral Stories (one feature), 82% on SOCIAL-CHEM-101 (seven).
Social norms are numerous and culture-specific, so teaching a classifier what counts as a violation is harder than it looks. Showing up to the office drunk is a violation in most workplaces; turning that judgment into a measurable feature has not been done in a theoretically grounded way. Prior automatic work is scarce and usually targets community-specific rules on sites like Reddit. DARPA's computational cultural understanding program funded this study, aiming at tools such as digital interpreters for cross-cultural interaction.
The bet is narrow: the catalog of norms is huge, the catalog of social emotions that fire when a norm is broken is small. Shame, embarrassment, and guilt are treated as portable signals of violation.
Yair Neuman and Yochai Cohen at Ben-Gurion University run a four-stage pipeline.
Seven distinct formulas survive. Almost all of them use social emotions. The ten hand-built norm scores do not appear. Typical forms: a weighted sum of shame and anger, shame and guilt, or guilt and embarrassment; one formula is pure negative sentiment; one is the good/bad gap.
On 25 short scenarios from a psychology study, the zero-shot classifier's top emotion among pride, gratitude, guilt, anger, and sadness matched the human top label 64% of the time. Chance is 20%; binomial p ≈ 0.0002.
On 8,352 EmpatheticDialogues situations with the emotion word stripped, pride/gratitude cases scored higher than guilt/shame/embarrassment on all ten norms (Kruskal-Wallis all significant). Violation cases mostly scored below zero, except on caring. Symbolic classification on this dataset's test split averaged 94% accuracy and 96% precision.
Held-out datasets are the real test.
| Setting | Accuracy | Precision | Recall |
| Moral Stories, single-formula CRT, 50% baseline | 68% | 71% | 68% |
| Moral Stories, 22-feature CRT | 78% | 77% | 81% |
| SOCIAL-CHEM-101, seven-formula CRT | 82% | 80% | 87% |
| SOCIAL-CHEM-101, 22-feature CRT | 89% | 87% | 84% |
Moral Stories: 12,000 normative actions vs 12,000 divergent ones. Cronbach's Alpha across the seven formulas is only 0.73. The best single formula is the good/bad gap at 73% accuracy. A backward logistic regression with 18 of the 22 features also lands at 78% accuracy, 75% precision, 83% recall: the short rules beat a 50% baseline, but they do not match the full feature set. SOCIAL-CHEM-101: 5,934 illegal actions matched to 5,934 legal ones. Alpha rises to 0.93. Best single formula is shame plus guilt at 80%. Feeding all seven formula scores into one CRT reaches 82%.
A follow-up asks which moral foundation was broken. On SOCIAL-CHEM-101 cases tagged with a single foundation (care, fairness, authority, sanctity; loyalty dropped at n=30), a boosting classifier with five-fold CV beats prevalence guessing by 34 points of recall on average: care 90% (baseline 51%), fairness 71% (27%), authority 42% (13%), sanctity 31% (9%). The five most influential features are trust, caring, loyalty, and conformity scores, plus anger.
For content moderation, cross-cultural assistants, or moral-reasoning evals, the cheap path is: skip building a full norm ontology, score shame, guilt, embarrassment, and anger, add a sentiment polarity, and you already beat chance. 94% on EmpatheticDialogues looks strong; a single formula falls to 68% on Moral Stories. The rules are brittle. The 22-feature models still reach 78% and 89%.
The more useful finding is inverted. The ten GPT-3-derived norms barely enter the "was a norm broken?" formulas. Social emotions do the work. Norm scores start to matter only when the question is which kind of violation.
This is 2023-era GPT-3 plus bart-large-mnli. Methodologically incremental. The contribution is turning a psychology claim (violations travel with social emotions) into a reproducible classification pipeline, and insisting on short formulas instead of a black-box leaderboard run.
The authors already flag several: only "general" social norms, not group- or person-level; results are tied to the chosen datasets, theory, and models; conversations, multimodality, and culture-specific norms are left for later.
Several claims sit on thin ice. Training labels are proxies: shame/guilt/embarrassment mean violation, pride/gratitude mean adherence. The paper notes that not every negative emotion marks a violation, then trains as if those two emotion buckets were the classes. Experiment 2 treats legal/illegal as social-norm violation; law and etiquette are different. The 64% emotion result uses 25 scenarios, a thin warrant for "the classifier works." Both validation sets are English, crowd-sourced, and US-leaning, which sits awkwardly next to a computational-cultural-understanding grant. Symbolic regression is capped at depth 5; the surviving formulas are two- or three-term linear combinations. "Distilling natural laws from data" oversells the complexity that was found.