Examples of sarcasm-related errors.

Sentiment analysis across multiple languages remains a challenging problem due to linguistic diversity, domain-specific expressions, and the limited explainability of existing models. This study aims to address these issues by proposing the Explainable Multilingual Sentiment Analyzer (EMSA), a novel framework that integrates large language models with prompt engineering. EMSA employs a two-stage process, first generating sentiment reasoning through chain-of-thought prompts, and then producing sentiment classification with explicit interpretability. We evaluate EMSA on both the GubaSenti dataset (Chinese financial domain) and the SST dataset (English benchmark). Experimental results demonstrate that EMSA consistently outperforms pre-trained language models such as RoBERTa, XLNet, and ALBERT, while providing transparent reasoning steps that enhance user trust. These findings suggest that EMSA not only improves multilingual sentiment classification performance but also contributes to the development of more interpretable and practical sentiment analysis systems.