This study proposes BERTGRU4Sentiment, a deep learning model that combines fine-tuned Bidirectional Encoder Representations from Transformers (BERT) with a Gated Recurrent Unit (GRU) for multilingual sentiment analysis. The proposed model is evaluated on four benchmark datasets: Wongnai (Thai), Amazon Reviews (English), Twitter/X (English), and IMDB (English), covering both three-class (negative, neutral, positive) and two-class (negative, positive) classification tasks. To address class imbalance, a mixed over- and under-sampling strategy is applied to all datasets. Results demonstrate that BERTGRU4Sentiment consistently outperforms traditional baseline models including Decision Tree, Random Forest, XGBoost, and LSTM across all datasets, achieving accuracy of 69.49% on Wongnai, 93.67% on Amazon Reviews, 71.73% on Twitter, and 83.97% on IMDB. The findings confirm that BERTGRU4Sentiment enables richer contextual feature extraction, and the proposed architecture is effective for multilingual sentiment classification tasks.