The rapid growth of Twitter as a real-time communication platform has created an urgent need for effective hot topic detection. Traditional statistical and machine learning models often fail to capture contextual semantics and long -range dependencies, while deep learning approaches such as CNNs and LSTMs improve representation but face challenges in scalability, optimization, and convergence. This study proposes a novel deep learning framework that integrates Multi-Scale Conv1D for diverse n-gram feature extraction, an attention-enhanced BiLSTM for contextual learning, and a hybrid Modified Bald Eagle Optimization–Particle Swarm Optimization (MBES-PSO) strategy for robust parameter tuning. Unlike conventional models limited by fixed kernel sizes or shallow architectures, the proposed design dynamically captures both local and global semantic patterns in tweets. The hybrid optimizer balances global exploration with local exploitation, achieving faster convergence and improved stability. The framework is evaluated on a large-scale Twitter dataset from Kaggle. Experimental results show that the proposed model achieved the highest accuracy of 90.12%, significantly outperforming 13 state-of-the-art baselines across precision, recall, and F1-score. This study contributes: 1) a Multi-Scale Conv1D architecture for enriched feature extraction; 2) an attention-based BiLSTM module for improved interpretability; 3) a hybrid MBES-PSO optimizer that enhances convergence and avoids local minima; and 4) extensive comparative evaluation validating robustness on real-world Twitter data. The proposed framework offers a scalable, interpretable, and high-performing solution for real-time hot topic detection in social media analytics.