Commonsense-driven symbolic ReAct-NLI prompting (CSR-NLI) for causal analysis of mental health issues in workspace communication; advancements in the CAMS dataset

Abstract

Mental health issues expressed on workspace communication often involve complex causal relationships that require deeper analysis beyond sentiment detection. Traditional NLP models face challenges in interpretability and causal inference, limiting their ability to accurately identify stressors and triggers. To address this, Commonsense-Driven Symbolic ReAct-NLI (CSR-NLI) is introduced as a neuro-symbolic prompting framework that integrates commonsense reasoning with Natural Language Inference (NLI) for structured causal analysis. By leveraging symbolic validation, iterative refinement, and neural inference, CSR-NLI ensures that causal explanations align with real-world commonsense knowledge, enhancing the robustness of AI-driven mental health analysis. Additionally, advancements in the Causal and Abductive Mental State (CAMS) dataset [1] improve causal annotations, contributing to better reasoning consistency and classification accuracy. The proposed framework was evaluated using GPT-3.5-Turbo, GPT-4-Turbo, Allam-2-7B, LLaMA-3.3-70B, and DeepSeek-R1, demonstrating superior causal reasoning performance compared to existing approaches. CSRNLI effectively enhances explainability and reliability in mental health AI applications, offering a scalable solution for analyzing stressors and triggers in online discourse.