Abstract
In recent years, the dominance of Large Language Models (LLMs) such as BERT and RoBERTa has led to remarkable improvements in NLP tasks, including mental illness detection from social media text. However, these models are often computationally intensive, requiring significant processing time and resources, which limits their applicability in real-time or resource-constrained environments. This paper proposes a lightweight yet effective hybrid model that integrates a 6-layer RoBERTa encoder with a capsule network architecture to balance performance, interpretability, and computational efficiency. The contextual embeddings generated by RoBERTa are transformed into primary capsules, and dynamic routing is employed to generate class capsule outputs that capture high-level abstractions. To validate performance and explainability, we employ LIME (Local Interpretable Model-Agnostic Explanations) to provide insights into feature contributions and model decisions. Experimental results on benchmark mental health datasets demonstrate that our approach achieves high accuracy while significantly reducing inference time, making it suitable for deployment in real-world mental health monitoring systems.1.To design a computationally efficient architecture for mental illness detection using a lightweight RoBERTa encoder integrated with capsule networks.2.To perform a detailed time complexity analysis highlighting the trade-offs between performance and efficiency.3.To enhance model interpretability through LIME-based feature attribution, supporting transparent and ex- plainable predictions.