Multi-Axis Ensemble Wins PsyDefDetect Shared Task at BioNLP 2026
A 9-voter ensemble system developed by Nürnberg NLP achieved first place in the PsyDefDetect shared task at BioNLP 2026, which focuses on classifying psychological defence mechanisms in supportive conversations. The task is inherently ambiguous due to overlapping language across eight positive defence categories, with trained raters reaching only moderate inter-annotator agreement. The winning approach uses a multi-axis ensemble spanning three orthogonal axes: class granularity (a gatekeeper for all nine classes versus specialists for the eight defence classes), training method (generative and discriminative), and base model. This design leverages error independence rather than a stronger single model, achieving an F1 score of .420 on the hidden test set and placing first among 21 registered teams.
Key facts
- Nürnberg NLP developed a 9-voter ensemble for the PsyDefDetect shared task at BioNLP 2026.
- The task involves classifying eight positive psychological defence categories in supportive conversations.
- The ensemble spans three orthogonal axes: class granularity, training method, and base model.
- The system achieved an F1 score of .420 on the hidden test set.
- It placed first among 21 registered teams.
- The approach emphasizes error independence over a stronger single model.
- Trained raters reach only moderate inter-annotator agreement on the task.
- The shared task is part of BioNLP 2026.
Entities
Institutions
- Nürnberg NLP
- BioNLP 2026
- PsyDefDetect