Comparing Multilingual Emoji Enhanced Product Reviews: A Transformer Based Approach for Language Pair and Emotion Detection

This paper presents a multilingual sentiment analysis pipeline leveraging two transformer-based architectures—XLM-RoBERTa (base) and BERT-based multilingual cased—to classify sentiment across four language pairs (English–Spanish, English–French, English–Hindi, and English–Italian). We fine-tune XLM-RoBERTa by unfreezing only its last three layers to adapt the model to domain-specific sentiment cues while preserving its robust cross-lingual representations. Training over ten epochs yields a best validation accuracy of 0.9579 and a test accuracy of 0.975, with an average F1-score around 0.92–0.97 across the four language pairs. The BERT-based multilingual cased model achieves a slightly higher test accuracy of 0.98, demonstrating comparable or improved performance in capturing sentiment nuances. These results confirm that selectively fine-tuning large-scale multilingual encoders is an effective strategy for cross-lingual sentiment classification, achieving high accuracy and strong generalization.