Social Chemistry 101: Learning to Reason about Social and Moral Norms

Understanding social norms is fundamental to human society, influencing behaviors ranging from politeness to moral judgments. Early research by Haidt and Kohlberg established moral foundations and developmental stages, emphasizing cultural variability. With the rise of NLP, efforts like Malle et al.'s descriptive ethics models attempted to quantify moral judgments, but lacked large-scale, multi-dimensional datasets. Recent advances in deep learning, especially pre-trained transformers like GPT and T5, enable modeling complex language patterns. However, applying these models to social norms remains challenging due to the nuanced, multi-faceted nature of normative judgments. This study bridges this gap by creating a comprehensive dataset and a specialized model, aiming to facilitate automated social norm reasoning across diverse contexts.

The core challenge is enabling AI to understand and reason about social norms in a manner comparable to humans, especially in complex, conflicting, or culturally sensitive situations. Existing models often rely on rule-based systems or single-dimension analysis, which cannot capture the richness and variability of real-world norms. The difficulty lies in modeling multi-faceted judgments—moral, legal, cultural—and integrating them into a unified reasoning framework. Additionally, data scarcity and lack of multi-dimensional annotations hinder progress. Addressing these issues is crucial for developing AI systems that can interact ethically and adaptively in human environments.

The main innovations include: 1) constructing SOCIAL-CHEM-101, a large, multi-dimensional social norm dataset with 12 judgment attributes, capturing diverse perspectives; 2) designing the Neural Norm Transformer, a multi-task conditional language model that jointly learns rule generation and attribute prediction, enabling nuanced reasoning; 3) introducing a multi-angle attribute fusion mechanism, improving generalization to unseen scenarios. These advancements differ from prior work by integrating rich attribute annotations with deep generative models, moving beyond rule-based or single-dimension approaches, and providing a scalable framework for ethical AI.

• �� Data collection: Gathered 104k real-life situations from online forums, advice columns, and stories. • Crowdsourcing: Annotators wrote 292k rules-of-thumb (RoTs) inspired by situations, ensuring diversity and relevance. • Structured annotation: Each RoT was broken down into actions, characters, and 12 judgment attributes, including moral foundations, legality, and cultural pressure. • Model design: Utilized pre-trained transformers (GPT, T5) with multi-task objectives—RoT generation conditioned on attributes, and attribute prediction conditioned on situations. • Training: Employed joint optimization with cross-entropy loss, multi-task balancing, and permutation-based data shuffling for robustness. • Evaluation: Used accuracy, F1-score, and human judgment to assess rule relevance, diversity, and ethical alignment.

Models trained on SOCIAL-CHEM-101 achieved 78% RoT generation accuracy, outperforming baselines by 15%. Zero-shot reasoning on unseen situations reached 65%, indicating strong generalization. Ablation studies showed multi-attribute fusion improved performance by over 20%. The model effectively captured cultural differences, aligning with moral foundation theory. Hyperparameters included a learning rate of 1e-5, batch size 32, trained over two days on high-performance GPUs. Human evaluations confirmed the relevance and diversity of generated rules, with qualitative analysis revealing the model's capacity to produce contextually appropriate and novel norms.

The model's high accuracy demonstrates its ability to understand and generate social norms across diverse contexts. Its generalization to unseen scenarios indicates robustness, while ablation results highlight the importance of multi-dimensional attribute integration. Cross-cultural experiments confirmed sensitivity to normative differences, validating the framework's versatility. These findings suggest that deep neural models can effectively learn complex social expectations, paving the way for practical applications in AI ethics and social computing.

Potential applications include automated content moderation, social media analysis, and AI-driven ethical decision support. The model can assist platforms in detecting norm violations, understanding cultural differences, and providing context-aware ethical guidance. In industry, it enables more trustworthy AI systems capable of navigating complex social landscapes, improving user experience and societal trust. Additionally, it supports research in social sciences by providing computational tools for analyzing normative variations across cultures and communities.

The dataset's cultural scope is limited to North American English sources, restricting cross-cultural generalization. The model struggles with ambiguous or conflicting norms, especially in extreme cases. Computational costs are high, limiting real-time deployment. Ethical biases present in training data may influence outputs, necessitating further bias mitigation. Future work should focus on multilingual datasets, interpretability, and bias control to enhance robustness and fairness.