Social media and its network data play an important role in modern life, particularly as a key platform and data resource for emergency response. Using machine learning models, this work aims to classify tweets as disaster-related or irrelevant, addressing the critical challenge of filtering genuine disaster information from noisy content on Twitter. The research integrates advanced feature extraction techniques, including Word2Vec and GloVe, with a hybrid approach combining Bidirectional Encoder Representations from Transformers (BERT), Support Vector Machine (SVM), and bagging techniques to enhance classification accuracy. It also compares classification performance across different feature extraction methods and machine learning classifiers. To ensure the reliability of the classification results, model performance is rigorously evaluated using a 5-fold cross-validation approach, with key model performance metrics like accuracy, precision, recall, as well as F1-score. The findings obtained from this study demonstrate that BERT as a feature extraction method effectively captures contextual information, SVM achieves strong performance in high-dimensional feature spaces, and bagging improves model robustness, resulting in superior classification performance for the data set we considered. The study show that the proposed approach outperforms other model combinations, achieving significant improvements in accuracy, recall, and F1-score. The proposed method also significantly enhances the reliability of social media information, providing emergency responders with accurate, real-time data during disasters, thus facilitating timely and effective response efforts.