Blunt Giants, Sharp Specialists: Decoding Stress and Triggers Beyond General-Purpose LLMs

Abstract

The United Nations sustainable development goals emphasize mental health as essential for inclusive and sustainable development, making stress, an important factor affecting well-being and societal participation, a critical area of study. Leveraging social media data, we propose the emotion-aware stress cause analysis framework (ESCAF), which jointly models stress detection (SD) and stress cause identification (SCI). ESCAF employs a contextual attention block (CAB) to prioritize stress-relevant context and incorporates sentence-level emotion information for enriched contextual understanding. A key contribution is the stress and cause recognition dataset, derived from the Dreaddit corpus, providing utterance-level annotations for emotion and stress-triggering sentences within each post. ESCAF surpasses in-domain state-of-the-art methods with average F1 gains of +5.23% for SD and +6.57% for SCI. For comprehensive evaluation, we also assess fine-tuned and zero-shot performances of leading large language models (LLMs); ESCAF outperforms larger models including Llama, Mistral, Qwen, and Phi, while maintaining a smaller parameter footprint. These findings highlight the need for task-specific architectures requiring nuanced emotional understanding and caution against over-reliance on generic LLMs in sensitive applications.