NormGenesis: Multicultural Norm Dialogue Framework

• NormGenesis is a multicultural, multilingual framework designed for generating socially grounded dialogues, featuring norm adherence, violation, and resolution through its novel V2R model.
• It employs an exemplar-based iterative refinement process to enhance pragmatic consistency and ensure cultural authenticity in English, Chinese, and Korean dialogues.
• The framework benchmarks a dataset of 10,800 multi-turn dialogues annotated for norm adherence, speaker intent, and emotional responses with rigorous evaluation metrics.

NormGenesis is a multicultural, multilingual framework for generating and annotating socially grounded dialogues across English, Chinese, and Korean. It is designed for social norm modeling in dialogue systems, with an emphasis on responses that are not only coherent but also socially acceptable. The framework introduces a novel dialogue type, Violation-to-Resolution (V2R), to model conversations following norm violations through recognition and socially appropriate repair, and it incorporates exemplar-based iterative refinement to improve pragmatic consistency, especially in underrepresented languages. Using this framework, the associated work constructs a dataset of 10,800 multi-turn dialogues annotated at the turn level for norm adherence, speaker intent, and emotional response; human and LLM-based evaluations report gains in refinement quality, dialogue naturalness, and generalization performance, and show improved pragmatic competence in ethically sensitive contexts for models trained on V2R-augmented data (Hong et al., 22 Sep 2025).

1. Conceptual scope and norm taxonomy

NormGenesis is presented as a multicultural, multilingual framework for generating and studying socially normed dialogues across English (American cultural context), Chinese, and Korean. Its norm modeling is organized around 12 major conversational norm categories—such as Apology, Empathy, and Compliment—with 10 culturally-specific subnorms per language, totaling 360 subnorms. The subnorms are derived from sociocultural surveys, including the World Values Survey for Korean, and are linguistically and culturally validated by native experts for realism (Hong et al., 22 Sep 2025).

The framework is situated against approaches that treat social norms primarily as a static classification problem. Its central premise is that conversational norms are not exhausted by adherence or violation labels alone; rather, they also govern the dynamics of recognition, response, and repair. In that sense, NormGenesis formalizes social norms as interactional structures rather than isolated tags. This suggests a shift from norm detection toward norm-conditioned dialogue generation, especially in settings where cultural variation, emotional response, and pragmatic repair are integral to system behavior.

2. Dialogue types and scenario–situation design

NormGenesis organizes dialogue synthesis around three scenario types for every subnorm: Norm Adherence, Norm Violation, and Violation-to-Resolution (V2R). Each scenario is an abstract description, while each situation expands the scenario with roles, emotions, and contextual cues that tightly constrain dialogue generation. Pragmatic features such as tone, formal or casual register, honorifics, and relational distance are explicitly included in the situation design (Hong et al., 22 Sep 2025).

Dialogue type	Function in the framework	Typical conversational role
Norm Adherence	Models socially appropriate conduct	Proper norm enactment
Norm Violation	Models socially inappropriate conduct	Detectable norm breach
Violation-to-Resolution (V2R)	Models progression after a violation	Recognition and repair

This scenario–situation distinction is structurally important. It separates abstract social configuration from its linguistically realized form, allowing the system to encode cultural and pragmatic constraints before utterance generation. A plausible implication is that this decomposition helps preserve cross-lingual comparability while still allowing language-specific realizations of hierarchy, politeness, and emotional stance.

3. Exemplar-based iterative refinement and generation pipeline

The framework uses a four-stage pipeline: Norm and Style Design, Scenario–Situation Construction, Exemplar-Based Iterative Refinement, and Multi-Turn Dialogue Generation with Annotation. In the refinement stage, the paper reports that initial model generations, especially in Korean and Chinese, often lack pragmatic fluency or appropriateness. To address this, small sets of manually-refined, high-quality exemplars are used to dynamically retrieve similar scenarios for an iterative rewrite process, in which the model revises candidate scenario–situation pairs using the exemplar as a guide (Hong et al., 22 Sep 2025).

The refinement loop is evaluated through Refinement Quality (RQ), which scores NormAlignment, LanguageQuality, and SemanticFidelity on Likert 1–5 scales: $$\text{RQ}(\text{orig}, \text{refined}) = \text{NormAlignment} + \text{LanguageQuality} + \text{SemanticFidelity}$$ Each component is scored separately and then averaged. Revision continues until quality exceeds a predefined threshold.

This procedure is described as an upstream data quality improvement process that introduces alignment with linguistic, emotional, and sociocultural expectations before full dialogue generation begins. In the paper’s formulation, the refinement stage is particularly important for low-resource languages, where surface fluency can coexist with pragmatic inconsistency, including honorific misuse, tone mismatch, and unnatural register mixing. The method is therefore not simply a post-hoc filter; it is a structural component of the generation regime.

4. Violation-to-Resolution dynamics

The most distinctive dialogue type in NormGenesis is Violation-to-Resolution (V2R). V2R dialogues capture not just the act of violating a social norm, but the process of recognizing the violation, by either party, and generating a culturally and contextually appropriate resolution. Dialogues begin with a norm violation—intentional or accidental—and then proceed through reactions such as apology, justification, empathy, explanation, or compensation. V2R threads can be multi-turn and are intended to reflect realistic recovery, which often requires more than one repair act (Hong et al., 22 Sep 2025).

The paper identifies several common repair strategies: Apology (A), Explanation (X), Empathy (E), Compensation (C), and Humor (H). Apology is reported as the most common and near-universal strategy, while empathy is especially central in Korean. One prevalent repair sequence given in the data is Empathy $\rightarrow$ Apology $\rightarrow$ Explanation for Korean (Hong et al., 22 Sep 2025).

Within the framework’s interpretation, V2R serves two purposes. First, it operationalizes repair as a dialogue phenomenon rather than as a binary label attached to a single utterance. Second, it provides training data for models that must generate socially appropriate recovery behavior in ethically sensitive contexts. This suggests that norm competence, as represented in NormGenesis, is fundamentally processual: a model is evaluated not only on whether it avoids a violation, but also on whether it can recognize and repair one.

5. Dataset structure, annotation scheme, and evaluation

NormGenesis constructs a dataset of 10,800 dialogues, corresponding to 3 languages $\times$ 3 types (Adherence/Violation/V2R) $\times$ 1,200 per type per language. The resource is described as consisting of multi-turn 5–15 turn conversations, while Appendix A reports average $\sim 12$ turns (ranges 5–16, V2R longer) (Hong et al., 22 Sep 2025).

The annotation schema is turn-level and includes three main dimensions. First, norm-level adherence is labeled as Adherence, Violation, V2R (for recovery), or Not Relevant. Second, speaker reaction labels (dialogue acts) include tags such as APO (Apology), THX (Thanks), EMP (Empathy), JUS (Justification), ACK (Acknowledgement), AGR (Agreement), DIS (Disagreement), SUG (Suggestion), QUE (Question), CRT (Criticism), and N/A. Third, each turn includes a justification in free-form explanation, and the abstract also characterizes the corpus as annotated for emotional response (Hong et al., 22 Sep 2025).

Evaluation is organized around Refinement Quality (RQ), Dialogue Quality (DQ), and Generalization Quality (GQ). DQ uses criteria including Consistency, Naturalness, Relevance, Emotional Appropriateness, Social Norm Appropriateness, and Scenario Coherence. Human evaluation is conducted with native speakers in every language on Likert-5 scales, with Krippendorff’s alpha reported as 0.81 for Korean and 0.72 for Chinese. LLMs are also used as dataset auditors through custom prompts for each criterion. The reported findings include improvement in refinement fluency and norm fit, with Korean linguistic quality rising from $\sim 3.6$ to $4.9/5$ and norm alignment from $4.6$ to $4.9+$. For Korean dialogues, sample DQ results report Consistency = 4.98 for Adherence, 4.98 for V2R, and 2.59 for Violation, while Norm Appropriateness = 4.70 for Adherence, 3.81 for V2R, and 1.61 for Violation (Hong et al., 22 Sep 2025).

6. Benchmark position and cross-disciplinary ambiguity of the term

The paper positions NormGenesis as a new benchmark for culturally adaptive dialogue modeling and describes it as the only open dataset with turn-level annotation of norm, intent, and emotional responses; multicultural subnorm coverage validated by native speakers; dynamic V2R repair modeling and explicit repair strategy analysis; and structured, upstream exemplar-based refinement for pragmatic consistency in multiple typologically distinct languages. In downstream evaluation, models including GPT-4o-mini, LLaMA-3, and Qwen are trained or fine-tuned on NormGenesis, NormDial, and SODA, then tested out of domain on DailyDialog for English and LCCC for Chinese. In blind A/B testing, responses from models trained on NormGenesis are reported as preferred in up to 79\% of cases in Chinese and 68\% in Korean, with human and LLM judgments showing strong agreement at $\rightarrow$0–$\rightarrow$1 (Hong et al., 22 Sep 2025).

The term NormGenesis also appears in a separate cosmology literature with a different meaning. In "Constraints on Genesis Cosmology from the Smeared Null Energy Condition", Case II ($\rightarrow$2) is called “NormGenesis”, where scale-invariance arises from the vacuum fluctuations of the Galileon/Genesis field itself, and the model is studied under SNEC constraints together with the gradient-stability bound $\rightarrow$3 (Yu et al., 4 Dec 2025). This suggests that the label is polysemous across research areas: in dialogue modeling it denotes a multicultural framework for social norm generation and repair, whereas in early-universe cosmology it denotes a specific Genesis scenario within generalized Galileon theories.