News

PsyDefDetect Shared Task at BioNLP 2026: Detecting Psychological Defense Mechanisms in Conversations

Published on June 30, 2026 • Co-organized by University of Technology Sydney, Xi'an Jiaotong-Liverpool University, ELLIS Institute Finland, etc.

The upcoming BioNLP 2026 workshop (July 3–4, 2026) co-located with ACL 2026 in San Diego, California, United States, will feature the PsyDefDetect Shared Task. Co-organized by researchers from University of Technology Sydney, Xi'an Jiaotong-Liverpool University, University of Utah, Harvard University, University of Melbourne, University of Manchester, ELLIS Institute Finland, and University of Turku, the initiative addresses the challenge of identifying psychological defense mechanisms in emotional support conversations.

In emotional support dialogues, defense mechanisms shape how help-seekers disclose their feelings and accept or resist support. Despite progress in computational empathy and dialogue strategy modeling, the defensive function of seeker utterances has remained largely unmodeled. The shared task overview paper will be presented on Friday, July 3, 2026 (12:00 - 12:15 PDT / 19:00 - 19:15 UTC) during Session 2 (Shared Tasks Overviews). The shared task aims to bridge this gap by classifying seeker utterances into specific levels based on the Defense Mechanism Rating Scales (DMRS) framework, supporting research at the intersection of psychology and natural language processing.

The PsyDefConv Dataset & Task Challenge

Systems participating in the shared task are evaluated on the newly released PsyDefConv dataset, which consists of 200 dialogues and 2,336 seeker utterances annotated under the DMRS framework (4,709 utterances in total including supporter turns).

A key technical challenge for participating systems is the pronounced class imbalance. The Level 7 High-Adaptive Defense class alone accounts for 51.8% of all annotated seeker utterances. Models that over-predict the majority class can easily attain high overall accuracy but suffer from poor recall on minority defense classes. The top-performing architectures are those that maintain a favorable balance between overall accuracy and minority-class recall.

Introducing the Participating Teams

The shared task attracted competitive solutions from across the global NLP community. Over the course of the evaluation period, the competition registered 172 participants who together produced 563 submissions to the leaderboard. By the final registration deadline,21 teams had officially submitted their results for inclusion in the ranking.

The participating teams will present their systems on Saturday, July 4, 2026 during Session 3 (14:00 - 15:30 PDT / 21:00 - 22:30 UTC). Three featured team presentations include:

Overall, the shared task results demonstrate that while multi-model ensembling and deliberative-agent architectures lead the field, substantial headroom remains for future research to achieve robust clinical defense detection in emotion support dialogues.

Workshop Presentation Schedule

The shared task overview and team systems will be presented across two days. Add these events directly to your calendar or download the combined iCal event: