Sentiment analysis using text classification on social media post

Abstract

Our research introduces an advanced and robust framework in sentiment analysis for mental health detection on social media platforms using both standalone and hybrid deep learning models. Addressing the common challenges posed by noisy, informal, and semantically complex user generated content here in this research our study systematically investigates and compares the performance of various deep learning models such as : LSTM, RNN, MLP, and CNN architectures against a hybrid model that integrates the Universal Sentence Encoder (USE) with Gated Recurrent Units (GRU) deep learning model. By cleaning, and annotating a large-scale the multi-source dataset validated by our choosen psychologist expert review’s our work demonstrates that conventional deep learning approaches often suffer from class imbalance and limited generalization, especially on real-world social data. On the other hand our hybrid (USE+GRU) model achieves a substantial improvement with an accuracy of 83% and significantly better precision-recall tradeoffs across sentiment classes, outperforming all baselines.These research findings decisively establish the superior potential of combining transfer learning with sequential modeling for nuanced sentiment detection and early mental health risk assessment in digital environments. The results not only extend the state of the art in text-based emotion and mental health analytics but also provide a scalable and more robust foundation for intelligent, ethical, and context aware intervention tools that can empower online platforms and researchers in monitoring well-being and mitigating the growing mental health crisis signaled through social media. Hence our research work sets a new benchmark for the practical deployment of deep learning in digital mental health and opens the way for the next generation of adaptive data focused solutions for social listening and psychological support in the near future.