KODIS: A Multicultural Dispute Resolution Dialogue Corpus

TL;DR

KODIS introduces a large multicultural dispute-resolution dialogue corpus comprising 4,061 participants from 75+ countries to study how culture shapes dispute dynamics in text-based interaction. The framework combines a theory-driven collection design based on Dignity, Face, and Honor with pre/post measures, elicited preferences, and turn-by-turn emotion labeling using GPT-4o, including a human–AI pairing element. Findings show clear escalation patterns where anger propagates toward impasses, and emotional expressions robustly predict subjective outcomes (approximately $R^2 \approx 0.50$), with cross-cultural generalization highlighting biases in emotion interpretation across cultures. The work provides a valuable resource for social science, NLP evaluation, and AI-assisted dispute coaching, while acknowledging limitations related to the artificial scenario, labeling biases, and the need for broader language coverage and human validation.

Abstract

We present KODIS, a dyadic dispute resolution corpus containing thousands of dialogues from over 75 countries. Motivated by a theoretical model of culture and conflict, participants engage in a typical customer service dispute designed by experts to evoke strong emotions and conflict. The corpus contains a rich set of dispositional, process, and outcome measures. The initial analysis supports theories of how anger expressions lead to escalatory spirals and highlights cultural differences in emotional expression. We make this corpus and data collection framework available to the community.

Figures (12)

• Figure 1: Participants did pre / post-dispute questionnaires and interacted with their counterpart. We first try to match the participant with a human, though we match them with GPT4 if unmatched after seven minutes.
• Figure 2: This illustrates an example of a contentious dialogue from the KODIS corpus.
• Figure 3: K-means clusters of countries ($N\geq 10$).
• Figure 4: Anger by role, outcome and dialogue turn.
• Figure 5: Average GPT emotion scores for the five most common countries broken by within or cross-culture.