Automating the classification of psychological defense mechanisms is a critical yet challenging frontier in clinical natural language processing. General-purpose Large Language Models (LLMs) struggle to apply fine-grained ordinal frameworks like the Defense Mechanism Rating Scales due to the implicit nature of clinical cues and a fundamental clinical reasoning gap. These models exhibit severe extreme response bias, systematically gravitating toward the scale’s endpoints while failing to resolve nuanced, mid-level defenses. In this paper, we present our third-place system for the PsyDefDetect Shared Task at BioNLP 2026, designed specifically to overcome this failure mode. We propose a hybrid architecture that synergizes label-flattened generative retrieval with an LLM classifier fine-tuned via the distillation of supervised clinical reasoning traces. This dual approach, grounding decisions in rubric criteria while leveraging task-specific supervision, successfully mitigates the observed bias, achieving an accuracy of 67.37% and a macro-F1 of 39.56%. Our work provides empirical evidence that tightly integrating targeted clinical supervision with dynamic rubric-grounded retrieval significantly outperforms the raw parameter scale of un-tuned foundation models.

This paper presents PerceptionLab’s approach for the BLP-2025 Shared Task 1A on multiclass Bangla hate speech detection, addressing severe class imbalance and informal online discourse. We perform Domain-Adaptive Pretraining (DAPT) on BERT models using a curated corpus of over 315,000 social media comments to capture slang, non-standard spellings, and contextual nuances of online discourse. To enrich underrepresented categories, we align external resources and construct a novel Bangla sexism dataset of over 6,800 comments via weak supervision and manual verification. Two classification strategies are compared: a single-shot six-way classifier and a two-stage hierarchical model that first separates Hate from Non-hate before fine-grained categorization. Experimental results show that single-shot classification with DAPT-enhanced BUET-BERT achieves the highest micro-F1 score (0.7265), outperforming the hierarchical approach and benchmarked general-purpose Large Language Models. Error analysis reveals persistent challenges in detecting subtle sexism and context-dependent religious hate. Our findings highlight the value of domain adaptation, robust end-to-end modeling, and targeted dataset construction for improving fine-grained hate speech detection in low-resource settings.