EMNLP 2025 paper accepted! 🎉

Excited to share that our paper, “NormGenesis: Multicultural Dialogue Generation via Exemplar-Guided Social Norm Modeling and Violation Recovery,” has been accepted to the Main Conference of EMNLP 2025! 🎉 This work addresses a key limitation of large language models (LLMs): their inability to adequately reflect social norms and generate high-quality dialogues in low-resource languages.

With NormGenesis, we: 1️⃣ Built culturally appropriate, high-quality dialogue datasets by collecting social norm data from diverse cultures and refining generated dialogues through iterative guidance from a small set of expert exemplars. 2️⃣ Introduced the Violation-to-Resolution (V2R) paradigm, the first systematic framework where conversations not only include norm violations but also the process of realizing and repairing those violations, enabling smoother and more natural dialogues. 3️⃣ Demonstrated that models trained on the NormGenesis dataset significantly outperform those trained on existing social norm or commonsense dialogue datasets, achieving state-of-the-art performance.

A camera-ready version, along with the preprint and code, will be released soon. Stay tuned!!