Embedded Files
For this edition, the corpus consists of simulated yet realistic therapeutic conversations in Spanish between patients and professional therapists. The conversations were created using a combination of human-authored and synthetically generated dialogues, all of which were reviewed by experts to ensure clinical plausibility.
TBU
Task 1: Early Symptom Detection in Therapeutic Conversations.
Task 1: Early Symptom Detection in Therapeutic Conversations.
Multiclass classification
Multiclass classification
This task focuses on the early identification of mental health symptoms expressed by a patient during a therapeutic dialogue.
Participants are given therapist-patient conversations that unfold chat messages. After each patient's turn, systems must decide whether there is evidence of one or more predefined mental health symptoms and transdiagnostic psychopathological verbal processes, based on the dialogue observed so far. Key characteristics:
No labeled training data are provided.
Systems must rely on zero-shot, weakly supervised, or knowledge-based approaches (e.g., pretrained language models, prompting strategies, external resources).
A taxonomy of psychopathological processes and symptoms with clear definitions will be provided prior to evaluation. Performance depends not only on correctness, but also on how early symptoms are detected, penalizing late or spurious predictions.
The symptoms will come from tests such as: PHQ, CompACT-10 or GAP.
Task 2: Therapist Response Selection.
Task 2: Therapist Response Selection.
Binary classification
Binary classification
Task 2 evaluates NLP systems as decision-support tools for therapists. Given the full conversation history up to a therapist intervention point, systems must select the most appropriate therapist response from a predefined set of candidate options. Each candidate represents a plausible therapeutic action, but only one aligns with expert-defined best practice for the given context. Key characteristics:
The full dialogue history is available at each decision point.
Responses are selected, not generated, ensuring controlled and ethically safe evaluation.
Candidate responses may include AI and non-AI responses from which one has been identified as preferable by experts.
Page updated
Google Sites
Report abuse