This project highlights the importance of analyzing social networks in the digital transition that must take place in the administration in order to adapt to the new sources of information. The creation of the mental health observatory proposed could help hospitals to: receive a global image of the social perception and its changing dynamics as a source of social information, manage their human resources by anticipating work peaks based on the analysis of trends found in social networks, consult demographic profiles, obtain an analysis of temporal or seasonal characteristics, and discover relationships. This project makes technological contributions in: the generation and availability to the community of i) linguistic resources; and ii) tools to carry out natural language understanding adapted to Spanish and the terminology of these networks. The main challenge of the project is to convert unstructured textual information into interpretable knowledge and story-telling.
1. Introduction tims of trafic accidents [1]. The correct identification According to the World Health Organization, 80 peoplteechnology in mental health provides an opportunity to die by suicide every hour, being suicide among the 10create breakthrough solutions that improve mental health most common causes of death globally. In Spain, in 2022,and well-being outcomes on a greater scale than ever be2015 people died due to suicide and self-harm injuriesfore”. It is not new the fact that artificial intelligence while 761 were victims of trafic accidents, that is, lastleverages valuable knowledge from collective behaviors year the number of suicides almost tripled that of vaics-it is the case of the social observatories in polit6i]c.s [ Our hypothesis rests on the fact that collective emotors 2[], job occupation, age (young 3[] and elder people of signals that could indicate suicide risk is a core, yettional and behavioral reflections in the social media entail undisclosed, cornerstone. Some risks and factors conin-herent psycho-linguistic features connected to general nected to suicide ideation or commitment include: mewn-ell-being. In our case, we will pay special attention tal health disorders, loneliness or lack of connection wittho risk factors such as self-harm, loneliness, depression significant ones, depression, bullying, consumption of and anxiety. Moreover, changes in these features might substances, lack of adherence to treatment, self-harm ertecv.eal changes that can light pathways to aid healthcare There are documented diferences in the population notpractitioners envisaging a punctual social picture related only by health condition but also by socio-economic fact-o mental illnesses. The work carried out by Fine e[t7a]l.
and [8] is a proof of concept of how publicly available so[4]), gender and other social and demographic factors.cial media data might be used to assess population-level In the foreword of the “Global Governance Toolkit fomrental health. The analysis of data outside the health
tions ∗Corresponding author. nEvelop-O
0000-0003-2638-9598 (A. Pérez);0000-0001-9097-6047 care system, and in particular social media data, may lead to the development of automatic systems for assessing suicide risk based on what people share on their social networks. This kind of information allows looking into a person’s life in a more frequent and daily basis, when the potential risk of committing suicide can be monitored much more eficiently [ 9]. We find, however, a gap for this type of approaches involving deep natural language understanding in Spanish.
Obser-Menh project1 involves advanced computa- data is the core of our project on which rely the protional semantic technologies and makes a technologicaplosed artificial intelligence approaches. Annotating data contribution generating tools to carry out natural alaidne-d by human experts provides extra possibilities for guage understanding adapted to Spanish and the termtih- e algorithms to learn in semi-supervised fashions or nology of these networks with the added value of rendefro-llowing few-shot learning schemes. However, general ing unstructured textual information into interpretpaubrlepose tools to process natural language are still to be knowledge. adapted to the nuances of language employed in social media. A contribution of this project rests on the retraining or adaptation of tools to analyse and process 2. Objectives social-media texts. Specific lexica will be created to suit domain adapted language models that are employed by In brief, a review to the antecedents outlined the im
deep neural language understanding approaches such as portance of detecting negation, pronouns, variations in
BERT and related transformers. These tools are crucial to the emotional state, abstract adjectives, social factors
gain insights beyond static linguistic lists and represent etc. Thus, tools should be adapted to Spanish in this
a step ahead in semantic relatedness. Indeed, it is a deep domain. All in all, we find that there is a gap in Spanish language understanding the approach fostering the novpsico-linguistic analysis. We miss annotated data andelty of our approach, an observatory displaying psychoalso the application of medical entity recognition (MER)
linguistic features and revealing chronological dynamics prior to polarity discovery approaches. Having detected
in the embedded topics, discovering latent demographic the strenghts and weaknesses of the related work, we
profiles of anxiety, loneliness, depression, self-harm risk focus on the gaps, and thus define the objectives. Generaland suicide ideation. The novelty rests, not only on the objectives include:
information extracted from unstructured media but also • Development of tools for information retrievainl the purpose itself aiming at providing insights from in Spanish focused on the analysis of data fromlanguage to experts in Mental eHealth and bridging the social networks related to mental health for tghaep between users and specialised health facilities and early detection of changing trends in psycho‐linas-sociations. In fact, the project shall contribute with guistic characteristics. suited negation and speculation detection tools, medi• Analysis of existing public data on mental healthc,al entity recognition and emotional language discovery. development of annotation schemes for the proTb-he valuable resources developed in this project shall be lems addressed in the sub‐projects (including genr-eported and left to the research community. der data). In our project, the added value in terms of return to • Creation and adaptation of NLP tools to the dot-he society rests on the formulation of the time compomain of mental health in social media. nent within an observatory in order to keep track and, • Generation of demographic profiles focusing on particularly, focus on changes across time in social perthe psycho‐linguistic characteristics expressedcineption, emotions and behaviour latent in language and building demographic profiles of latent linguistic features. the texts and inherent to potential mental health risks for the benefit of individuals, taking genderMoreover, we envisage to render unstructured linguistic dimension into account. data into human-friendly graphical data to aid decision support by health practitioners. • Analysis of the relationships among the difer- The expected findings would be in the research line of ent problems addressed in the sub-projects thacto,mputational social psychology, particularly, in psycothrough artificial intelligence, provide dynamismlinguistics. In brief, user generated content is a valuable for the benefit of better healthcare. means of collective intelligence and subjective perception of reality, it is dynamic and fast. Connective action 3. Novelty and added value and post sharing leads to under-used complex though valuable big data in terms of natural language.
A careful review to antecedents disclosed that eHealth in social media has been addressed in English. Our project4. Methodological approach will bridge the gap to adapt deep natural language understanding tools to Spanish and has an added value tTohe methodology proposed to cope witOhbser-Menh alternative Spanish variants (e.g. Argentina, Chile, Meixs-broadly depicted in Figur1e. In brief, in order to cope ico etc.) with the particular analysis of each variantwbyith the aforementioned goals, the design encompasses means of the demographic-profile analysis. Social media the following modules and tasks: 1Obser-Menh project:http://nlp.uned.es/obser-menh-project • Data: this module involves two important tasks: 1. Crawling a large data‐set from social netc-hanging emotions and behaviors potentially connected works (e.g. Twitter) in Spanish. with substantial changes in mental health conditions. 2. Health‐expert aided manual and semi‐automatic annotation of relevant linguistic cues e.g. drugs, key phrases, specific lexi- 5. Research groups cons (e.g. chapter F of ICD‐10 or, alternaT-he Obser-Menh project is arranged as two coordinated tively, Snomed-ct) and gender indicators. • NLP&IR group at UNED (http://nlp.uned.es)/ References
Has a long trajectory in Intelligent Access to Information and Knowledge Representation. The[1] Instituto Nacional de Estadística, Defunciones por group counts on several research lines within the causa de muerte,https://www.ine.es/jaxi/Tabla. clinical domain and demographic cue extraction htm?tpx=55863&L=,02022. Accessed: 2023-03-06. from text (of particular relevance in the segmen[-2] F. Ferretti, A. Coluccia, Socio-economic factors and tation study underlying mental health). NLP&IR suicide rates in European Union countries, Legal group at UNED is the leader of the aforemen- Medicine 11 (2009) S92–S94.
tioned GELP sub-project. [3] D. Wasserman, Q. Cheng, G.-X. Jiang, Global sui• The psychiatry service of thHeospital Clínico cide rates among young people aged 15-19, World San Carlos in Madrid, cooperates with NLP&IR psychiatry 4 (2005) 114.
group at UNED in the GELP project. [4] M. Waern, E. Rubenowitz, B. Runeson, I. Skoog, • HiTZ (http://www.hitz.e u)sHiTZ is a reference K. Wilhelmson, P. Allebeck, Burden of illness and center on Language Technologies aiming at the suicide in elderly people: case-control study, Bmj promotion of research, training, technological 324 (2002) 1355. transfer and innovation in Artificial Intelligence [5] World Economic Forum, Deloitte, Global goverfocused on language and speech technologies. nance toolkit for digital mental health: Building The subset of researchers from HiTZ involved trust in disruptive technology for mental health, in this project are all within Ixa research group 2021. URL: https://www3.weforum.org/docs/WEF_ (http://www.ixa.eu)sa consolidated group with Global_Governance_Toolkit_for_Digital_Mental_ beyond than 35 years of existence in which a Health_2021.pd.f research line started more than 12 years ago is[6] S. Caton, M. Hall, C. Weinhardt, How do politicians devoted to NLP in the clinical domain. HiTZ is use Facebook? An applied social observatory, Big the leader of LOTU sub-project. Data & Society 2 (2015) 2053951715612822. • Osakidetza, the Public Health System within the[7] A. Fine, P. Crutchley, J. Blase, J. Carroll, G. CopBasque Country and, specifically, experts within persmith, Assessing population-level symptoms of the Network of Mental Health in Biscay involved anxiety, depression, and suicide risk in real time in BioCruces (the Health Research Center of Bis- using NLP applied to social media data, in: Proceedcay) cooperate with HiTZ in the LOTU project. ings of the fourth workshop on natural language processing and computational social science, 2020, pp. 50–54. 6. Conclusions [8] H. Moradian, M. A. Lau, A. Miki, E. D. Klonsky, A. L. Chapman, Identifying suicide ideation in menThe creation of the mental health observatory presented tal health application posts: A random forest algoin this project could help to: (1) receive a global image rithm, Death Studies (2022) 1–9. of the social perception and its changing dynamics as a[9] P. Resnik, A. Foreman, M. Kuchuk, K. Musacsource of social information, (2) manage their human re- chio Schafer, B. Pinkham, Naturally occurring lansources by anticipating work peaks based on the analysis guage as a source of evidence in suicide prevention, of trends found in social networks, (3) consult demo- Suicide and Life-Threatening Behavior 51 (2021) graphic profiles, (4) obtain an analysis of temporal or 88–96. seasonal characteristics, and (5) infer latent relationsh[i1p0s] M. K. Nock, E. M. Kleiman, M. Abraham, K. H. Bentbetween diferent problems. ley, D. A. Brent, R. J. Buonopane, F. Castro-Ramirez, C. B. Cha, W. Dempsey, J. Draper, et al., Consensus statement on ethical & safety practices for conduct7. Acknowledgements ing digital monitoring studies with people at risk of suicide and related behaviors, Psychiatric research and clinical practice 3 (2021) 57–66.
130398B-C21) and LOTU (TED2021-130398B-C22) are funded by MCIN/AEI/10.13039/501100011033 and by the European Union “NextGenerationEU”/PRTR. In addition, this work was partially funded by the Spanish Ministry of Science and Innovation (DOTT-HEALTH/PAT-MED PID2019-106942RB-C31 and INDICA-MED PID2019106942RB-C32) by the Basque Government (IXA IT-157022); and by EXTEPA within Misiones Euskampus2.0.