Title: LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification

Abstract:

Identifying psychological defense mechanisms within conversational data presents a significant hurdle in clinical natural language processing. For the PsyDefDetect 2026 shared task, which involves nine-class utterance classification assessed by macro F1, our team, LinguIUTics, secured the 4th position among 21 participating teams. We achieved an official positive-class leaderboard macro F1-score of 0.3917. This result represents an improvement of 7.7 absolute points (a 24.4% relative increase) over the Ministral-8B task baseline, which recorded a macro F1 of 31.48.

Due to severe class imbalance, both BERT-family encoders and zero-shot large language models demonstrated poor performance on rare classes. Consequently, we adopted a QLoRA fine-tuning approach for the Qwen3-8B model. Our methodology incorporates three primary strategies: grouped stratified cross-validation to prevent data leakage, minority-class round-robin lexical augmentation, and a post-processing pipeline featuring logit bias tuning and ensemble blending. These techniques significantly enhanced minority-class recall and narrowed the discrepancy between validation and leaderboard scores. Notably, this approach transformed the performance of the "Unclear" class (Level 8), lifting its F1 score from near-zero to 0.797.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC