Zero-Shot Cross-Cultural Emotion Detection Using Language-Agnostic Embeddings and Adaptive Emotion Prototypes

Emotion detection from text is a critical task in natural language processing, yet most existing models are limited by language and cultural bias, often relying on translation or large annotated datasets in high-resource languages. This paper presents a novel, translation-free framework for zero-shot cross-cultural emotion detection using language-agnostic sentence embeddings and adaptive emotion prototypes. Our method leverages pre-trained multilingual encoders (such as LaBSE) to map sentences from any language into a shared semantic space. Emotion classes are represented as centroids (prototypes) formed from culturally diverse, emotion-annotated phrases. For inference, input sentences are assigned the emotion of the nearest prototype in embedding space, enabling robust zero-shot classification across languages. Furthermore, the system supports rapid few-shot adaptation to new languages or cultural contexts by updating prototypes with a handful of local expressions. Experiments on multilingual datasets—including low-resource languages like Tamil, Farsi and more —demonstrate that our approach outperforms translation-based and monolingual baselines, especially for culturally nuanced emotions. The prototype-based architecture offers interpretability and scalability, making it suitable for global applications such as chatbots, counseling tools, and cross-cultural sentiment analytics.