Mental health issues such as depression, stress, anxiety, and personality disorders are increasingly prevalent, particularly within online communities. This study proposes a lightweight and efficient multi-class classification framework to identify five mental health conditions using Reddit user-generated posts. While previous studies predominantly rely on conventional CNNs or standard machine learning techniques for binary classification, our work introduces a novel Bidirectional Long Short-Term Memory (BiLSTM) model integrated with an attention mechanism. The architecture is further enhanced by synonym-based data augmentation using the WordNet lexical database, which improves semantic diversity and enhances model robustness, particularly for underrepresented classes. Unlike prior works that focus narrowly on binary classification or employ transformer-based models with high computational demands, our model offers a lightweight, high-performance architecture optimized for multi-class detection and real-world deployment. Experimental results demonstrate that the proposed model achieves a peak validation accuracy of 95.02%, along with precision 95.08%, recall 95.02%, and F1-scores of 95.03%. These findings support the advancement of efficient AI-driven diagnostic systems in mental health analytics and lay the groundwork for future integration into mobile or resource-constrained platforms.