This collection contains messages written by real Internet users. By requesting access you declare that you are a person engaged in research and development, and you agree to the following understandings, terms and conditions, which apply equally to all or part of the information contained in the collection.

Permitted uses

The information may only be used for research purposes.
Summaries, analyses and interpretations of the linguistic properties of the information may be derived and published, provided it is impossible to reconstruct the information from these summaries.
You may not identify the individuals whose texts are included in this dataset. You may not cross-reference individuals in the dataset against any other dataset or collection of data. You may not establish any kind of contact with the individuals in this dataset.
You are not permitted to publish any portion of the dataset (e.g. example posts) other than summary statistics, nor share it with anyone else.
You may not otherwise make unauthorised commercial use of, reproduce, prepare derivative works from, distribute copies of, perform, or publicly display the collection or parts of it.
You may present research findings obtained using the collection within the limits of this agreement. Any scientific publication derived from the use of this collection must explicitly cite the MentalRiskES paper: Mármol-Romero, A. M., Moreno-Muñoz, A., Plaza-del-Arco, F. M., Molina-González, M. D., Martín-Valdivia, M. T., Ureña-López, L. A., & Montejo-Ráez, A. (2024). MentalRiskES: A New Corpus for Early Detection of Mental Disorders in Spanish. LREC-COLING 2024, pp. 11204-11214.
You shall not use results obtained through the use of the collection for profitable purposes, including advertisement, or for defamatory or slanderous purposes.
If the organizers or the copyright holders request you to discontinue the use of the collection, or your use is deemed to be in violation of this agreement, you shall immediately discontinue use and promptly delete the collection and all data obtained by processing it from any computer or media onto which it has been copied.
You must keep the data on infrastructure under your control and protect it against unauthorised access. You may not upload it to third-party services or APIs (including external large language model APIs) that are not under your control, and you must notify the organizers of any security incident affecting the data.

Copyright. The copyright holders retain ownership and reserve all rights to the use and distribution of the information. Except as specifically permitted above, the display, reproduction, transmission, distribution or publication of the information is strictly prohibited. Violations of these restrictions may result in legal liability. Copyright holders of the information contained in the collection include a wide variety of online Internet users.

MentalRiskES

MentalRiskES is a corpus of mental disorders in Spanish, built from real conversations in public Telegram groups. It was created for the early detection of mental disorder risk and covers three disorders, each one available as a configuration of this dataset:

Configuration	Disorder	Subjects
`anxiety`	Anxiety	500
`depression`	Depression	499
`eating_disorder`	Eating disorders	335

Each configuration comprises a set of subjects, their message threads and the labels obtained from crowdsourced annotation. In total the corpus contains over 1,300 subjects and more than 45,000 messages.

⚠️ Access is restricted. The data comes from real users, so access is granted only after reviewing an individual application (see Access). All findings derived from this data are intended strictly for non-clinical research. Individuals in need of professional help are advised to consult licensed psychiatrists or clinicians.

Access

Log in to Hugging Face and click Request access at the top of this page.
Fill in the application form. It corresponds to the Individual Application to use the MentalRiskES Research Collection: your details, your organization and the acceptance of the terms shown in the form.
The organizers review each application manually. You will be notified when it is approved.

Please use your institutional email address.

How to load the data

After your access request has been approved, log in (hf auth login or the HF_TOKEN environment variable) and load the configuration you need:

from datasets import load_dataset

ds = load_dataset("amarmol/MentalRiskES", "depression")
print(ds)                          # a single split: train (the whole corpus)
subject = ds["train"][0]
print(subject["subject"], subject["bs"], len(subject["messages"]))

The corpus has no official partitions, so each configuration contains a single split (train) with all the subjects. The partitions used in the MentalRiskES shared task at IberLEF (2023, 2024 and 2025) are distributed separately; see MentalRiskES evaluation campaign.

Data format

Each row is one subject:

{
  "subject": "subject1",
  "bs": 1, "bc": 0, "rbs": 0.9, "rbc": 0.1,
  "messages": [
    {
      "id_message": 79865009803,
      "date": "2020-04-07 05:00:02",
      "message": "text with emojis written in words",
      "message_raw": "text with the original emojis"
    }
  ]
}

subject: anonymous identifier of the user.
messages: the messages of the subject, in their original order.
- message: the processed text, with emojis converted to text (and other normalizations).
- message_raw: the raw text, with emojis in their original format. Both versions are aligned message by message.
Label columns: see below. They differ between configurations.

Labels

The annotation was carried out by 10 different annotators. The most predominant label among the annotators is taken, with 1 meaning suffers and 0 meaning does not suffer. In case of a tie, the subject is taken as suffering.

The labels available for each disorder are:

Eating disorders: suffer (s), control (c)
Depression: suffer + in favour (sf), suffer + against (sa), suffer + other (so), control (c)
Anxiety: suffer (s), control (c)

Columns of the `anxiety` configuration

bs (binary suffer): 1 if the subject suffers and 0 if not, according to the frequency of the labels (in case of a tie, it is marked as suffers).
bc (binary control): 1 if the subject does not suffer and 0 if they do, according to the frequency of the labels (in case of a tie, it is marked as suffers).
rbs (regression binary suffer): number of times the subject has been marked as suffering among the total number of annotators, i.e. 10.
rbc (regression binary control): number of times the subject has been marked as not suffering among the total number of annotators, i.e. 10.

Columns of the `depression` and `eating_disorder` configurations

bs (binary suffer): 1 if the subject suffers and 0 if not, according to the frequency of the labels (in case of a tie, it is marked as suffers).
bsf (binary suffer favour): 1 if the subject suffers and is in favour, 0 if not.
bsa (binary suffer against): 1 if the subject suffers and is against, 0 if not.
bso (binary suffer other): 1 if the subject suffers and is neither in favour nor against, 0 if not.
bc (binary control): 1 if the subject does not suffer and 0 if they do (in case of a tie, it is marked as suffers).
rbs (regression binary suffer): number of times the subject has been marked as suffering among the total number of annotators, i.e. 10.
rbc (regression binary control): number of times the subject has been marked as not suffering among the total number of annotators, i.e. 10.
rsf (regression suffer favour): number of times the subject has been marked as suffering and in favour among the total number of annotators, i.e. 10.
rsa (regression suffer against): number of times the subject has been marked as suffering and against among the total number of annotators, i.e. 10.
rso (regression suffer other): number of times the subject has been marked as suffering and neither in favour nor against among the total number of annotators, i.e. 10.
rc (regression control): number of times the subject has been marked as not suffering among the total number of annotators, i.e. 10 (equal to rbc).

The labels rbs and rbc must sum to 1, and so must rsf, rsa, rso and rc.

How is the corpus constructed? 🛠️

Public groups on the Telegram social network were accessed, and conversations were extracted from them. The data was processed and only text messages were retained, excluding images, audio and other media.

To carry out the annotation, a subset of messages was extracted from each subject. This message thread was annotated by 10 different annotators through the Prolific platform, using the Doccano annotation platform.

Annotation guides 📄

Each corpus has an annotation guide adapted to its case, accompanied by a user manual for the annotation platform:

MentalRiskES evaluation campaign 🚀

MentalRiskES is a shared task organized at IberLEF that aims to promote the early detection of mental disorder risk in Spanish. MentalRiskES is also the name of a shared task on the early detection of mental disorder risk in Spanish, organized at IberLEF. The two should not be confused: the shared task is an evaluation campaign with its own data in each edition. This corpus was the data source of the first two editions (2023 and 2024); the other editions use different data and are unrelated to this corpus.

📌 Looking for the shared task partitions? This Hugging Face dataset contains only the corpus (all subjects of each disorder in a single split). The collections used in each edition of the shared task (2023, 2024, 2025,...), with their official train, trial and test splits for every task, are distributed through the MentalRiskES GitHub repository. Please follow the access instructions given there to obtain them.

If you use the shared task collections, please cite the corresponding overview papers:

@article{marmol2023overview,
  title={Overview of MentalriskES at IberLEF 2023: Early Detection of Mental Disorders Risk in Spanish},
  author={M{\'a}rmol-Romero, Alba Mar{\'\i}a and Moreno-Mu{\~n}oz, Adri{\'a}n and Plaza-del-Arco, Flor Miriam and Molina-Gonz{\'a}lez, Mar{\'\i}a Dolores and Mart{\'\i}n-Valdivia, Maria Teresa and Ure{\~n}a-L{\'o}pez, Luis Alfonso and Montejo-Ra{\'e}z, Arturo},
  journal={Procesamiento del Lenguaje Natural},
  volume={71},
  pages={329--350},
  year={2023}
}
 
@article{marmol2024overview,
  title={Overview of MentalriskES at IberLEF 2024: Early Detection of Mental Disorders Risk in Spanish},
  author={M{\'a}rmol-Romero, Alba Mar{\'\i}a and Moreno-Mu{\~n}oz, Adri{\'a}n and Plaza-del-Arco, Flor Miriam and Molina-Gonz{\'a}lez, Mar{\'\i}a Dolores and Mart{\'\i}n-Valdivia, Maria Teresa and Ure{\~n}a-L{\'o}pez, Luis Alfonso and Montejo-Ra{\'e}z, Arturo},
  journal={Procesamiento del Lenguaje Natural},
  volume={73},
  pages={435--448},
  year={2024}
}

Citation of the corpus ❗

If you use this corpus, please cite:

Mármol-Romero, A. M., Moreno-Muñoz, A., Plaza-del-Arco, F. M., Molina-González, M. D., Martín-Valdivia, M. T., Ureña-López, L. A., & Montejo-Ráez, A. (2024). MentalRiskES: A new corpus for early detection of mental disorders in Spanish. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 11204–11214). ELRA and ICCL. https://aclanthology.org/2024.lrec-main.978

@inproceedings{marmol-romero-etal-2024-mentalriskes-new,
    title = "{M}ental{R}isk{ES}: A New Corpus for Early Detection of Mental Disorders in {S}panish",
    author = "M{\'a}rmol-Romero, Alba Mar{\'\i}a  and
      Moreno-Mu{\~n}oz, Adri{\'a}n  and
      Plaza-del-Arco, Flor Miriam  and
      Molina-Gonz{\'a}lez, M. Dolores  and
      Mart{\'\i}n-Valdivia, Mar{\'\i}a Teresa  and
      Ure{\~n}a-L{\'o}pez, L. Alfonso  and
      Montejo-R{\'a}ez, Arturo",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.978",
    pages = "11204--11214",
}

The corpus is also archived on Zenodo: 10.5281/zenodo.8055604.

Ethics statement ⭐

The main purpose of creating this dataset is to provide a tool for developing systems with scientific aims and advancing artificial intelligence to detect mental disorders among young people on social networks at an early stage. The data has been labelled using measures to ensure gender diversity and avoid discrimination through our annotation platform. All findings from this dataset are intended strictly for non-clinical research. Individuals in need of professional help are advised to consult licensed psychiatrists or clinicians.

Contact 📩

Questions about the corpus or the access process:

Alba M. Mármol Romero (amarmol@ujaen.es)
Arturo Montejo Ráez (amontejo@ujaen.es)

Acknowledgements

This work has been partially supported by projects CONSENSO (PID2021-122263OB-C21), MODERATES (TED2021-130145B-I00) and SocialTOX (PDC2022-133146-C21), funded by Plan Nacional I+D+i from the Spanish Government; project PRECOM (SUBV-00016), funded by the Spanish Ministry of Consumer Affairs; and the Big Hug project (P20_00956, PAIDI 2020), funded by the Andalusian Regional Government.

Downloads last month: 51

Total file size:

11.6 MB

You need to agree to share your Hugging Face information to access this dataset

Individual Application to use the MentalRiskES Research Collection