Abstract
Conversations often adhere to well-understood social norms that vary acrosscultures. For example, while "addressing parents by name" is commonplace in theWest, it is rare in most Asian cultures. Adherence or violation of such normsoften dictates the tenor of conversations. Humans are able to navigate socialsituations requiring cultural awareness quite adeptly. However, it is a hardtask for NLP models. In this paper, we tackle this problem by introducing a "Cultural ContextSchema" for conversations. It comprises (1) conversational information such asemotions, dialogue acts, etc., and (2) cultural information such as socialnorms, violations, etc. We generate ~110k social norm and violationdescriptions for ~23k conversations from Chinese culture using LLMs. We refinethem using automated verification strategies which are evaluated againstculturally aware human judgements. We organize these descriptions intomeaningful structures we call "Norm Concepts", using an interactivehuman-in-loop framework. We ground the norm concepts and the descriptions inconversations using symbolic annotation. Finally, we use the obtained datasetfor downstream tasks such as emotion, sentiment, and dialogue act detection. Weshow that it significantly improves the empirical performance.