transformer_1376 at PsyDefDetect: A QLoRA-Based Generative Framework for Context-Aware Psychological Defense Mechanism Detection
Year: 2026
Type: conference-paper
Abstract: Psychological defense mechanisms play a crucial role in shaping human responses during emotionally charged conversations, yet remain underexplored in natural language processing.In this work, we address the PSYDEFCONV shared task, which involves classifying defense mechanisms in multi-turn dialogues using clinically grounded annotations based on the Defense Mechanism Rating Scales (DMRS).We propose a generative supervised fine-tuning framework that reformulates the task as conditional text generation.A pre-trained causal language model (Gemma-2-2B) is adapted using parameter-efficient fine-tuning (PEFT) with 4-bit quantization, enabling efficient training under limited computational resources.To handle class imbalance, we apply random oversampling, and we design a prompt-based input representation to incorporate conversational context effectively.Experimental results demonstrate that our generative approach is competitive with discriminative baselines while offering improved flexibility in modeling subtle and context-dependent defensive behaviors.The findings highlight the potential of generative large language models for psychologically grounded dialogue understanding tasks.
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: gold
OpenAlex ID: W7166811314