Abstract:
This paper studies multilingual emotion analysis by exploring classification and conditioned text generation within language models. We examine how mixed-language training and sequential transfer affect cross-lingual generalization in multilabel emotion classification. Focusing on English and Spanish, our experiments show that limited target data yields substantial gains, while pretraining on high-resource or related languages supports robust transfer in low-resource settings. We extend this framework to emotion-conditioned text generation, producing text via prompting and evaluating alignment using our optimized classifier as an automated proxy. Results reveal that while few-shot prompting provides marginal gains in alignment, zero-shot generation produces more diverse outputs. These findings indicate that targeted multilingual training reduces annotation bottlenecks in target languages and that classification models can serve as scalable, robust evaluators for conditioned generation.