💬 NLP

NormGenesis: Multicultural Dialogue Generation via Exemplar-Guided Social Norm Modeling and Violation Recovery

The paper introduces NormGenesis, a multicultural framework that generates and annotates 10,800 socially grounded dialogues in English, Chinese, and Korean by employing exemplar-guided refinement and a novel Violation-to-Resolution (V2R) dialogue type to model norm violations and repairs, thereby significantly enhancing the pragmatic competence and cultural adaptability of dialogue systems.

Original authors: Minki Hong, Jangho Choi, Jihie Kim

Published 2026-03-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Minki Hong, Jangho Choi, Jihie Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot how to be a good friend. You don't just want the robot to speak perfect grammar; you want it to know when to be polite, how to apologize without sounding robotic, and what to say when it accidentally offends someone.

This is the challenge behind NormGenesis, a new project by researchers at Dongguk University. They built a "social school" for AI, specifically designed to teach computers how to handle conversations in three very different cultures: English (US), Chinese, and Korean.

Here is the story of how they did it, explained simply.

1. The Problem: The "Robotic" Friend

Imagine you ask a robot to apologize to your boss in Korean.

The Old Way: The robot might say, "I am sorry for my mistake." It's grammatically correct, but in Korean culture, that might sound too blunt or disrespectful because it didn't use the right honorifics (special words for showing respect to elders/superiors). It's like wearing a tuxedo to a beach party—it's the wrong "vibe."
The Result: The conversation feels awkward, cold, or even rude, even though the robot tried its best.

Existing AI models are great at English but often stumble when trying to navigate the complex social rules of Asian languages, where tone, hierarchy, and "saving face" are everything.

2. The Solution: The "Social School" (NormGenesis)

The researchers created a framework called NormGenesis. Think of this as a specialized training camp for AI. Instead of just feeding the AI millions of random chats, they taught it the rules of the game before it ever started speaking.

They used three main tricks:

A. The "Cultural Cheat Sheet" (Exemplar-Based Refinement)

Before the AI writes a single line of dialogue, the researchers give it a "cheat sheet."

The Analogy: Imagine you are writing a letter to a difficult relative. Before you write, you look at a perfect letter your wise aunt wrote to a similar relative. You study her tone, her choice of words, and how she softened the blow.
How it works: The AI looks at these "perfect examples" (exemplars) that are culturally spot-on. It uses them to rewrite its own drafts before it finalizes the conversation. This ensures the AI gets the "flavor" of the culture right from the start, rather than trying to fix it later.

B. The "Oops-to-Okay" Lesson (Violation-to-Resolution)

Most AI training only teaches how to have a perfect conversation. But real life is messy. We make mistakes.

The Analogy: Think of a dance partner. If you step on their toe, a good dancer doesn't just freeze; they apologize, adjust their step, and keep dancing smoothly.
The Innovation: The researchers created a new type of training called Violation-to-Resolution (V2R). They teach the AI:
1. The Mistake: The AI intentionally says something rude or breaks a social rule.
2. The Repair: The AI then learns exactly how to fix it (e.g., "I'm so sorry, I didn't mean to interrupt," or "Please forgive me").
  This teaches the AI how to recover from social blunders, which is a huge part of being human.

C. The "Three-Lane Highway" (Multicultural Data)

They didn't just build this for English. They built it for English, Chinese, and Korean simultaneously.

Why? Because what is polite in New York might be rude in Seoul or Shanghai.
The Result: They created a massive dataset of 10,800 conversations. Every single sentence in these conversations is tagged with: "Is this polite?" "What is the speaker feeling?" and "Did they break a rule?"

3. The Results: From Robot to Human

When they tested their new AI against older models:

Old AI: Sounded like a tourist who memorized a phrasebook but didn't understand the culture. It used the wrong honorifics or sounded too casual in serious situations.
NormGenesis AI: Sounded like a local. It knew when to be formal, when to be warm, and how to apologize in a way that actually made people feel better.

In tests, the new AI was preferred 65% to 80% of the time over existing models. It didn't just speak the language; it understood the culture.

The Big Picture

NormGenesis is like a bridge. It connects the cold logic of computers with the warm, messy, beautiful complexity of human culture.

By teaching AI not just what to say, but how to say it in a way that respects cultural norms, the researchers are helping us build digital friends that don't just talk to us, but truly understand us. Whether you are in a boardroom in Chicago, a tea house in Beijing, or a family gathering in Seoul, this AI is learning to be a good neighbor everywhere.

1. Problem Statement

Current dialogue systems struggle to generate responses that are not only syntactically correct but also socially and culturally appropriate, particularly in low-resource languages (e.g., Korean, Chinese) compared to high-resource languages (e.g., English).

Pragmatic Failures: Existing models often produce responses with incorrect honorific usage, unnatural tones, and misaligned emotional responses, failing to capture the nuances of power dynamics and relational distance.
Static Norm Modeling: Prior datasets primarily focus on classifying whether a dialogue adheres to norms or violates them. They lack the modeling of dynamic recovery, i.e., how speakers repair social interactions after a norm violation occurs.
Data Scarcity: There is a lack of high-quality, turn-level annotated datasets that cover multicultural norms, speaker intent, and emotional responses across diverse languages.

2. Methodology: The NormGenesis Framework

The authors propose NormGenesis, a multicultural framework designed to generate and refine socially grounded dialogues in English (US-centric), Chinese, and Korean. The framework operates through four core stages:

A. Norm and Style Design

Taxonomy: The authors constructed a taxonomy of 12 social norm categories (e.g., Apology, Empathy, Respect, Greeting) by extending existing work with new categories like Empathy and Respect.
Subnorms: For each category, they defined 10 culturally grounded subnorms per language (360 total).
- Korean: Derived from World Values Survey data to ensure cultural specificity.
- English/Chinese: Adapted from existing literature and aligned with Korean outputs via LLMs.
Parameters: The system defines pragmatic parameters including tone (formal/casual), honorific usage, relational distance, and emotional alignment.

B. Scenario-Situation Construction

The system generates Scenario-Situation pairs:
- Scenario: A concise real-world context.
- Situation: A 3–5 sentence elaboration specifying roles, emotional states, and stylistic constraints.
Interaction Types: Each instance is labeled as one of three types:
1. Norm Adherence: Standard compliant interaction.
2. Norm Violation: Intentional breach of social norms.
3. Violation-to-Resolution (V2R): A novel dialogue type modeling the recovery process where speakers acknowledge the violation and employ repair strategies (e.g., apology, explanation) to restore social harmony.

C. Exemplar-Based Iterative Refinement (Core Innovation)

To address the "pragmatic mismatch" in low-resource languages, the framework introduces an upstream refinement mechanism before full dialogue generation:

Mechanism: Instead of static few-shot prompting, the system retrieves semantically and structurally aligned exemplars (curated by experts) based on intent, emotional tone, and discourse patterns.
Process: The model uses these exemplars to iteratively refine the generated Scenario-Situation pairs.
Evaluation Loop: A Refinement Quality (RQ) protocol evaluates the revised output against the original. The loop continues until the quality score meets a threshold, ensuring cultural and linguistic alignment before the dialogue is generated. This mitigates errors common in "generate-first" pipelines.

D. Multi-Turn Dialogue Generation & Annotation

Refined scenarios are expanded into 5–15 turn dialogues.
Turn-Level Annotation: Every utterance is annotated with:
- Norm adherence status.
- Speaker reaction (Intent: e.g., Apology, Justification; Emotion).
- Justification for the label.
This results in a dataset of 10,800 dialogues (1,200 per language × 3 interaction types).

3. Key Contributions

Multicultural Framework: A novel pipeline for generating socially grounded dialogues in English, Chinese, and Korean, specifically addressing the pragmatic degradation in low-resource settings via exemplar-based iterative refinement.
Violation-to-Resolution (V2R): Introduction of a new dialogue category that models the dynamic recovery of social interactions, capturing how speakers repair norm violations through contextually appropriate strategies (apology, explanation, empathy).
High-Quality Dataset: Construction of a 10,800-instance dataset with fine-grained, turn-level annotations for norm adherence, speaker intent, and emotional response, grounded in dialogue act theory.
Performance Validation: Demonstration that models trained on this data significantly outperform existing baselines (NORMDIAL, SODA) in norm alignment, emotional coherence, and generalization across languages.

4. Results and Evaluation

The authors evaluated the framework using Refinement Quality (RQ), Dialogue Quality (DQ), and Generalization Quality (GQ) metrics, assessed by both humans and LLMs.

Refinement Quality (RQ):
- In Korean and Chinese, the exemplar-based refinement significantly improved Linguistic Quality (e.g., Korean: 3.59 → 4.91) and Norm Alignment (e.g., Korean: 4.58 → 4.91).
- The process reduced pragmatic inconsistencies like tone mismatches and honorific errors.
Dialogue Quality (DQ):
- Generated dialogues achieved high scores (>4.9) in Consistency, Naturalness, and Scenario Coherence.
- V2R dialogues showed superior performance in Emotional Appropriateness and Scenario Coherence compared to standard adherence dialogues, proving the model's ability to handle complex repair scenarios.
- Human evaluations showed strong agreement with LLM scores (Pearson correlation $r > 0.92$ ).
Generalization Quality (GQ):
- Models fine-tuned on NormGenesis data were preferred over baselines (NORMDIAL, SODA) in 65–79% of A/B preference tests across English and Chinese.
- V2R Impact: In ethically sensitive scenarios (tested on PROSOCIALDIALOG), models trained with V2R data were preferred 82% of the time over models trained only on standard norm data, showing improved pragmatic competence in conflict resolution.

5. Significance

Bridging the Low-Resource Gap: The paper provides a scalable methodology to overcome the lack of culturally nuanced data in languages like Korean and Chinese, moving beyond simple translation or static prompting.
Dynamic Social Modeling: By introducing V2R, the work shifts the paradigm from static norm classification to modeling the fluidity of social repair, which is critical for real-world conversational agents.
Benchmarking: NormGenesis establishes a new benchmark for culturally adaptive dialogue modeling, offering a robust foundation for developing agents that can navigate complex social hierarchies, honorifics, and emotional contexts across diverse cultures.
Ethical AI: The framework explicitly addresses the need for agents to recognize and repair social errors, contributing to the development of more socially responsible and empathetic AI systems.