N\"urnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for Psychological Defence Mechanism Classification

Nürnberg NLP's ensemble approach for detecting psychological defence mechanisms achieved first place in the PsyDefDetect shared task by leveraging nine independent voters across different model architectures and training methods. The strategy prioritizes error independence over single-model strength, addressing the inherent ambiguity in classifying overlapping psychological categories.

This research addresses a fundamental challenge in natural language processing: detecting subtle semantic distinctions where human annotators themselves show only moderate agreement. The PsyDefDetect task requires distinguishing eight positive defence mechanisms that share similar surface language but differ in pragmatic function, making it a particularly difficult classification problem. Nürnberg's winning approach reveals an important principle in machine learning—when dealing with genuinely ambiguous boundaries, ensemble diversity matters more than individual model performance.

The multi-axis voter ensemble strategy spans three orthogonal dimensions: class granularity (using a gatekeeper model for all nine classes and specialist models for the eight defence categories), training methodology (combining generative and discriminative approaches), and base model selection. This structural diversity ensures that errors remain uncorrelated across voters, allowing the ensemble to overcome individual model weaknesses on overlapping category boundaries. The achieved F1 score of 0.420 on the hidden test set demonstrates the effectiveness of this approach.

This work has broader implications for NLP applications involving nuanced human psychology, clinical assessment, and subjective content classification. The methodology could apply to domains like sentiment analysis with ambiguous emotional states, medical coding with overlapping symptom descriptions, or content moderation where borderline cases frustrate single-model approaches. The research demonstrates that when human consensus itself is limited, computational systems should be designed around orthogonal redundancy rather than attempting to build a single, more powerful classifier.

Future work might explore whether this ensemble principle applies to other inherently ambiguous NLP tasks and whether weighted voting mechanisms could further optimize performance by learning which voter types excel at specific distinction boundaries.

• →Ensemble diversity across orthogonal axes outperforms single-model strength on ambiguous classification tasks
• →The nine-voter system combining gatekeeper and specialist models achieved first place with F1=0.420 among 21 teams
• →Multi-axis approach spans class granularity, training methods, and base models to ensure error independence
• →The methodology addresses tasks where human inter-annotator agreement is inherently moderate due to overlapping categories
• →This strategy has potential applicability to clinical NLP, sentiment analysis, and other subjective classification domains