We describe the Eraserhead system submitted to the PsyDefDetect shared task at BioNLP 2026, which frames psychological defense level detection as a nine-class utterance classification problem over supportive dialogue. Our system is based on Qwen3-14B and combines clinically informed prompt design, per-label oversampling, and careful inference settings for stable prediction. A central challenge of the task is strong class imbalance, with High-Adaptive responses appearing far more often than several minority classes. This makes it easy for models to favor the majority class and achieve reasonable accuracy while performing poorly on rarer categories. To address this, we iteratively adjusted oversampling targets based on error analysis and predicted label distributions across submission rounds. Our final system achieved an official macro F1 of 0.3418 on Leaderboard 1 and 0.3947 on Leaderboard 2, ranking 7th among the 21 registered teams on both leaderboards. We further analyze the main failure modes of the system, especially the difficulty of distinguishing Minor Image Distorting defenses from High-Adaptive responses and the persistent tendency to over-predict the majority class. These findings highlight the broader difficulty of modeling psychological function from text alone.

Misogynistic memes spread harmful stereotypes and toxic content across social media platforms, often combining sarcastic text and offensive visuals that make them difficult to detect using traditional methods. Our research has been part of the the Shared Task on Misogyny Meme Detection - LT- EDI@LDK 2025, identifying misogynistic memes using deep learning-based multimodal approach that leverages both textual and visual information for accurate classification of such memes. We experiment with various models including CharBERT, BiLSTM, and CLIP for text and image encoding, and explore fusion strategies like early and gated fusion. Our best performing model, CharBERT + BiLSTM + CLIP with gated fusion, achieves strong results, showing the effectiveness of combining features from both modalities. To address challenges like language mixing and class imbalance, we apply preprocessing techniques (e.g., Romanizing Chinese text) and data augmentation (e.g., image transformations, text back-translation). The results demonstrate significant improvements over unimodal baselines, highlighting the value of multimodal learning in detecting subtle and harmful content online.