Participate
This is the tenth year of eRisk and the lab plans to organize
three tasks:
Task 1: Conversational Depression Detection
This task extends last year's pilot by detecting depression through conversational agents while improving access and reproducibility.
Participants will interact with LLM personas fine-tuned with diverse user histories and released on Hugging Face.
Each model will be released on a different day, and participants will have limited days before giving their predictions.
The challenge is to determine whether each persona exhibits signs of depression and, within a limited conversational window,
identify active depressive symptoms and the overall depression level. The LLM personas will reflect different severity levels
guided by the BDI-II questionnaire, allowing systems to be evaluated across a spectrum of simulated depression.
Teams will download the released persona models, conduct their interactions, and submit predictions a few days later.
Evaluation will focus on two key aspects: (i) accurate identification of depressive symptoms present in the persona (if any)
and (ii) the overall depression level of the persona, following BDI-II standards.
A limited number of runs will be accepted, with both fully automated and manual-in-the-loop variants permitted,
encouraging exploration of conversational strategies while maintaining comparability across submissions.
The proceedings of the lab will be published in the online CEUR-WS
Proceedings and on the conference website.
To have access to the collection, all participants have to fill,
sign, and send a user agreement form (follow the instructions
provided here). Once you have
submitted the signed copyright form, you can proceed to register
for the lab at
CLEF 2026 Labs Registration site.
Important Dates
Task 2: Contextualised Early Detection of Depression
This task continues last year's shift from isolated posts to full conversational contexts,
aiming to capture real interaction dynamics across multiple speakers.
Participants must process dialogues sequentially, accumulating evidence,
where a message may only become informative when interpreted alongside preceding or subsequent publications.
Training phase. For 2026 we provide a specific, fixed training set derived from last year's collection.
This dataset is released upfront to enable reproducible development and validation.
Test phase. As in 2025, the evaluation is interactive: teams connect to our server,
receive one conversational context at a time, and submit predictions as the dialogue progresses within a limited window.
Systems are expected to integrate multi-speaker context, handle evolving evidence, and decide when there is sufficient
information to issue a reliable classification.
Evaluation. We assess both accuracy and timeliness, combining early-detection measures
(e.g., ERDE, Flatency) with standard classification metrics.
This design encourages models that monitor ongoing interactions effectively while minimising delays,
bringing early risk prediction closer to realistic, conversation-centric scenarios.
The proceedings of the lab will be published in the online CEUR-WS Proceedings and on the conference website.
To have access to the collection, all participants must fill, sign, and send a user agreement form (follow the instructions provided
here). Once you have submitted the signed copyright form, you can proceed to register for the lab at
CLEF 2026 Labs Registration site.
Important Dates
Task 3: ADHD Symptom Sentence Ranking
This new task targets sentence-level retrieval for the 18 symptoms defined in the Adult ADHD Self-Report Scale (ASRS-v1.1).
Participants must rank candidate sentences by their relevance to each symptom.
A sentence is considered relevant when it conveys information about the user's state with respect to the target ADHD symptom
(irrespective of polarity or stance), encouraging models to capture clinically meaningful evidence rather than surface keywords.
We will release a sentence-tagged dataset derived from publicly available social media writings, collected to contain ADHD-related expressions.
As this is the first edition of the ADHD ranking task, no annotated training data will be provided.
The release will consist solely of the test inputs, following the formatting conventions of recent eRisk ranking tasks.
Participants will submit 18 rankings (one per ADHD symptom) ordering candidate sentences by decreasing likelihood of relevance.
Relevance assessments will be produced via top-k pooling and expert annotation, and systems will be evaluated using standard IR metrics
such as MAP, nDCG (e.g., @100), and P@10.
By extending symptom-oriented retrieval beyond depression to ADHD, this task advances interpretable, symptom-aware retrieval
and supports cross-condition generalisation at sentence granularity.
The proceedings of the lab will be published in the online CEUR-WS Proceedings and on the conference website.
To have access to the collection, all participants must fill, sign, and send a user agreement form (follow the instructions provided
here). Once you have submitted the signed copyright form, you can proceed to register for the lab at
CLEF 2026 Labs Registration site.
