--- title: "Nürnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for" description: "The paper describes the system that won the BioNLP 2026 PsyDefDetect shared task. 12 references, cited by 2 papers here." authors: ["Philipp Steigerwald", "Eric Rudolph", "Jens Albrecht"] year: 2026 venue: "Proceedings of the BioNLP 2026 (Shared Tasks)" source: "https://aclanthology.org/2026.bionlp-2.9/" doi: "10.18653/v1/2026.bionlp-2.9" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/steigerwald-2026-nurnberg-nlp-psydefdetect-multi-axis/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/steigerwald-2026-nurnberg-nlp-psydefdetect-multi-axis/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) / Nürnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for Psychological Defence Mechanism Classification # Nürnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for Psychological Defence Mechanism Classification **Authors:** [Philipp Steigerwald](https://md.research.elamur.ai/free/en/authors/philipp-steigerwald/), [Eric Rudolph](https://md.research.elamur.ai/free/en/authors/eric-rudolph/) and [Jens Albrecht](https://md.research.elamur.ai/free/en/authors/jens-albrecht/) · **Year:** 2026 · **Venue:** [Proceedings of the BioNLP 2026 (Shared Tasks)](https://md.research.elamur.ai/free/en/venues/bionlp/) ([BioNLP 2026](https://md.research.elamur.ai/free/en/venues/bionlp/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.bionlp-2.9/) · [PDF](https://aclanthology.org/2026.bionlp-2.9.pdf) · [DOI 10.18653/v1/2026.bionlp-2.9](https://doi.org/10.18653/v1/2026.bionlp-2.9) · [OpenAlex](https://openalex.org/W7160968435) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/steigerwald-2026-nurnberg-nlp-psydefdetect-multi-axis/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper describes the system that won the BioNLP 2026 PsyDefDetect shared task. The task is to label each help-seeker utterance in emotional-support dialogues (PSYDEFCONV, built on ESConv and annotated with DMRS) with one of eight defence levels or "No Defence". The data is hard: classes overlap semantically, human agreement is only moderate, and the labels are strongly skewed. The authors argue that diverse voters whose mistakes do not coincide help more than a single stronger classifier. Their ensemble has nine voters and varies three things: - **Label set:** a 9-class gatekeeper versus 8-class defence specialists. - **Training style:** generative QLoRA SFT versus logistic regression on frozen hidden states. - **Backbone:** Ministral-8B versus Phi-4-14B. A two-stage vote first lets the gatekeepers decide "No Defence". Otherwise, all nine voters settle the defence class by majority. Rare classes were enriched with GPT-5.2 synthetic dialogues. The system reaches test macro-F1 of .420 and ranks first of 21 teams. The authors also analyse why mid-hierarchy defences remain hard to separate. ## Key points - **Data:** in PSYDEFCONV, C7 High-Adaptive covers 52% of samples, and Cohen's κ between annotators is .639. A t-SNE of the adapted Ministral-8B hidden states shows that only C0 forms a clearly separate cluster. - **Ensemble:** three top-ranked CV5 fold models each from Min-SFT 9c, Min-LR 8c and Phi4-LR 8c. Phi4-LR was selected because its per-fold scores were the most anti-correlated with Min-LR (r=−.544). Ties go to C7. - **Score progression (F1 test):** - organisers' baseline: .315 - 5-fold voting: .373 - 6-voter ensemble: .391 - final 9-voter ensemble: .420 (+33.4% over the baseline) - **Augmentation:** 738 GPT-5.2 synthetic dialogues were added and are released for replication. They slightly lower a single model's score (.307 vs .315) but raise the 9-voter ensemble by .042. - **Main errors:** C6 and C7 are confused with each other on 28 test samples, and 54% of C5 Neurotic samples are predicted as C7. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [Georg Simon Ohm University of Applied Sciences Nuremberg](https://md.research.elamur.ai/free/en/universities/georg-simon-ohm-university-of-applied-sciences-nuremberg/) — University, Nuremberg, Germany · 3 works · 2 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,552 works · [💭 Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) · 870 works · [💭 Social Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/social-psychology/) · 306 works ## Cite Cite as: Philipp Steigerwald, Eric Rudolph and Jens Albrecht. 2026. Nürnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for Psychological Defence Mechanism Classification. Proceedings of the BioNLP 2026 (Shared Tasks). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{steigerwald2026nurnberg, title = {{Nürnberg NLP at PsyDefDetect: Multi-Axis Voter Ensembles for Psychological Defence Mechanism Classification}}, author = {Philipp Steigerwald, Eric Rudolph and Jens Albrecht}, booktitle = {Proceedings of the BioNLP 2026 (Shared Tasks)}, year = {2026}, doi = {10.18653/v1/2026.bionlp-2.9}, url = {https://doi.org/10.18653/v1/2026.bionlp-2.9} } ``` - [← previous NUS-IDS at AMIYA/VarDial 2026: Improving Arabic Dialectness in LLMs with Reinforcement Learning](https://md.research.elamur.ai/free/en/papers/gollapalli-2026-nus-ids-amiya-vardial-2026/) - [next → OjaKV: Context-Aware Online Low-Rank KV Cache Compression](https://md.research.elamur.ai/free/en/papers/zhu-2026-ojakv-context-aware-online-low/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) ## References 12 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Thomas G Dietterich. 2000. Ensemble methods in machine learning. In Multiple Classifier Systems, volume 1857 of MCS 2000, Lecture Notes in Computer Science, pages 1–15. Springer. \#\# Ethics Statement Hongbin Na, Zimu Wang, Zhaoming Chen, Yining Hua, Rena Gao, Kailai Yang, Ling Chen, Wei Wang, Shaoxio… [find](https://scholar.google.com/scholar?q=Thomas+G+Dietterich%2E+2000%2E+Ensemble+methods+in+machine+learning%2E+In+Multiple+Classifier+Systems%2C+volume+1857+of+MCS+2000%2C+Lecture+Notes+in+Computer+Science%2C+pages+1%E2%80%9315%2E+Springer%2E+%5C%23%5C%23+Ethics+Statement) 2. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. [Focal loss for dense object detection](https://md.research.elamur.ai/free/en/papers/lin-2017-focal-loss-dense-object-detection/). In Proceedings of the IEEE International Conference on Computer Vision, pages 2980–2988. 3. Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. [Towards emotional support dialog systems](https://md.research.elamur.ai/free/en/papers/liu-2021-emotional-support-dialog-systems/). In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics, pages 3469–3483. 4. Hongbin Na, Yining Hua, Zimu Wang, Tao Shen, Beibei Yu, Lilin Wang, Wei Wang, John Torous, and Ling Chen. 2025. A survey of large language models in psychotherapy: Current landscape and future directions. In Findings of the Association for Computational Linguistics: ACL 2025, pages 7362–7376, Vienna… [find](https://scholar.google.com/scholar?q=Hongbin+Na%2C+Yining+Hua%2C+Zimu+Wang%2C+Tao+Shen%2C+Beibei+Yu%2C+Lilin+Wang%2C+Wei+Wang%2C+John+Torous%2C+and+Ling+Chen%2E+2025%2E+A+survey+of+large+language+models+in+psychotherapy%3A+Current+landscape+and+future+directi) 5. At F 1 test =.420 the system misclassifies the majority of defence-bearing utterances with an adaptive-skew bias that under-flags exactly the patients who most warrant clinical attention. Assigning defence labels to a person’s utterances is itself a psychological intervention and should not occur ou… [find](https://scholar.google.com/scholar?q=At+F+1+test+%3D%2E420+the+system+misclassifies+the+majority+of+defence%2Dbearing+utterances+with+an+adaptive%2Dskew+bias+that+under%2Dflags+exactly+the+patients+who+most+warrant+clinical+attention%2E+Assigning+de) 6. Hongbin Na, Zimu Wang, Zhaoming Chen, Peilin Zhou, Yining Hua, Grace Ziqi Zhou, Haiyang Zhang, Tao Shen, Wei Wang, John Torous, Shaoxiong Ji, and Ling Chen. 2026b. [You never know a person, you only know their defenses: Detecting levels of psychological defense mechanisms in supportive conversations](https://md.research.elamur.ai/free/en/papers/na-2026-you-never-know-person-you/).… 7. J Christopher Perry. 1990. Defense Mechanism Rating Scales (DMRS), 5th edition. Cambridge, MA. [find](https://scholar.google.com/scholar?q=J+Christopher+Perry%2E+1990%2E+Defense+Mechanism+Rating+Scales+%28DMRS%29%2C+5th+edition%2E+Cambridge%2C+MA%2E) 8. J Christopher Perry. 2014. Anomalies and specific functions in the clinical identification of defense mechanisms. Journal of Clinical Psychology, 70(5):406– 418. [find](https://scholar.google.com/scholar?q=J+Christopher+Perry%2E+2014%2E+Anomalies+and+specific+functions+in+the+clinical+identification+of+defense+mechanisms%2E+Journal+of+Clinical+Psychology%2C+70%285%29%3A406%E2%80%93+418%2E) 9. Philipp Steigerwald and Jens Albrecht. 2026. From “Help” to helpful: A hierarchical assessment of LLMs in mental e-health applications. [find](https://scholar.google.com/scholar?q=Philipp+Steigerwald+and+Jens+Albrecht%2E+2026%2E+From+%E2%80%9CHelp%E2%80%9D+to+helpful%3A+A+hierarchical+assessment+of+LLMs+in+mental+e%2Dhealth+applications%2E) 10. Data Availability The 738 GPT-5.2 synthetic dialogues, generation prompt and parameters are released under CC BY-NC 4.0 at [https://github.com/th-nuernberg/](https://github.com/th-nuernberg/) nuernberg-nlp-psydefdetect. [↗](https://github.com/th-nuernberg/) 11. Philipp Steigerwald, Nico Bienlein, Jennifer Burghardt, Mara Stieler, Robert Lehmann, and Jens Albrecht. 2025. CAIA in practice: Field evaluation of an AI-assisted support system for text-based online counselling. In 2025 IEEE 37th International Conference on Tools with Artificial Intelligence (ICTA… [find](https://scholar.google.com/scholar?q=Philipp+Steigerwald%2C+Nico+Bienlein%2C+Jennifer+Burghardt%2C+Mara+Stieler%2C+Robert+Lehmann%2C+and+Jens+Albrecht%2E+2025%2E+CAIA+in+practice%3A+Field+evaluation+of+an+AI%2Dassisted+support+system+for+text%2Dbased+online) 12. Philipp Steigerwald, Jennifer Burghardt, Eric Rudolph, and Jens Albrecht. 2026. AI systems in text-based online counselling: Ethical considerations across three implementation approaches. [find](https://scholar.google.com/scholar?q=Philipp+Steigerwald%2C+Jennifer+Burghardt%2C+Eric+Rudolph%2C+and+Jens+Albrecht%2E+2026%2E+AI+systems+in+text%2Dbased+online+counselling%3A+Ethical+considerations+across+three+implementation+approaches%2E) *Updated 1 November 2026*