<p>Social media platforms have become central to modern communication; however, they often serve as breeding grounds for biased content that can subtly or overtly reinforce harmful societal stereotypes. Detecting such bias, including political, gender-based, racial, or ideological bias, remains a challenging task due to the contextual, implicit, and dynamic nature of language. To address this challenge, this study presents a multi-phase deep learning framework that integrates statistical and contextual representations for bias detection in online content. The process begins with preprocessing to clean and normalize raw text, enabling consistent downstream analysis. The architecture employs a dual-path feature extraction strategy: the first path derives statistical features using an Enhanced Term Frequency–Inverse Document Frequency (TF–IDF) approach, while the second path captures semantic and contextual representations using a Transformer model with self-attention to learn deep linguistic dependencies. The features from both paths are concatenated and passed to a feature selection stage, where the Adaptive Lyrebird Optimization Algorithm (A_LyOA) is used to identify the most relevant and non-redundant features. Finally, classification is performed using a multi-attention-based AlexNet integrated with a Multi-Layer Perceptron (MLP), where attention modules are incorporated within convolutional layers to improve contextual feature representation. The proposed framework was evaluated using comprehensive performance metrics, achieving high accuracy (0.98), precision (0.97), a low false negative rate (FNR = 0.13), and a low false positive rate (FPR = 0.08), demonstrating strong performance compared with existing bias detection approaches.</p>