Zero-Shot and Few-Shot Sentiment Analysis for Low-Resource Code-Mixed Languages

Code-mixed text is prevalent in multilingual communities, especially on social media. However, sentiment analysis in low-resource code-mixed languages (Konkani-English, English-Konkani-Marathi, Marathi-English, Tulu-English) remains challenging due to the lack of labeled data. This paper presents a zero-shot and few-shot sentiment classification framework leveraging mT5 with cross-lingual transfer learning. Our model is pre-trained on high-resource code-mixed data (Hindi-English, Tamil-English) and fine-tuned on small-scale datasets from low-resource languages. To improve performance, we integrate cross-lingual sentiment-aware embeddings and hybrid tokenization. Experimental results demonstrate state-of-the-art performance across real-world social media datasets, with an extensive ablation study validating the effectiveness of transfer learning, fine-tuning strategies, and tokenization techniques. These findings suggest that zero-shot and few-shot learning can significantly reduce data dependency for sentiment analysis in low-resource code-mixed languages.