Big Hug: Artificial intelligence for the protection of digital societies

Abstract

In this paper, we present the Big Hug Project, which aims to claim protect vulnerable citizens and help them and their families to feel more confident when using social media communication platforms. To this end, it proposes activities for building quality data, research in new algorithms to adapt current solutions to the changing nature of colloquial and informal communication, the evaluation of techniques and methods and the development of demonstrators. This project presents an interdisciplinary approach to early detection of young people at high-risk emotional problems. The involvement of colleagues from the Clinical Psychology and Corpus Linguistics fields, furthermore, provides the project with the necessary interdisciplinary to obtain robust results which may be significant to society.

Keywords

• eol>Natural Language Processing
• NLP
• sentiment analysis
• Clinical Psychology
• early detection

1. Introduction

Human language is the main transmission medium involved in social interaction. There are revolutionary Natural Language Processing (NLP) algorithms that can provide means to prevent and predict risky interactions, protecting the most fragile members of our digital societies. Children and adolescents have been identified by the World Health Organization as being at particular risk of psychological distress in these media1.

Human Language Technologies (HLT) can help us build more confident environments. Thanks to NLP, artificial intelligence solutions are able to model human language and use learned models to extract information and understand the meaning of text lfowing through social networks. The combination of deep learning algorithms with linguistic resources and tools, enable the construction of monitoring systems for the early detection of signs of misbehaviours like eating disorders, depression, bullying or suicide tendencies over social media[ 1, 2 ].

To this end, the project proposes two years of ac1https://www.who.int/news-room/fact-sheets/detail/ adolescent-mental-health tivities for building quality data, research in new al- disorder, which also caused anxiety, self-harming gorithms to adapt current solutions to the changing and a high risk of suicide. May studies have tackled nature of colloquial and informal communication, this fact from psychometrics, but better tools for the evaluation of techniques and methods and the modeling the language used would help [ 7 ], even development of demonstrators to leverage human- more when eating disorders are rising all around centered solutions that will protect vulnerable citi- the world. Emotional disorders, like depression and zens and help them and their families to feel more anxiety, afect a quarter of our population during confident when using social media communication their lifetime [ 8 ]. Depression can be studied and platforms. Besides, this project presents an inter- identified by monitoring users’ posts and activity disciplinary approach to early detection of young [ 1 ]. people at high-risk emotional problems. By indi- In Spain there are 10 suicides a day, twice as cated prevention, scientific community has agreed many people die by suicide as by trafic accidents, to name to high-risk individuals who are identified 11 times more than by homicide and 80 times more as having some detectable symptoms of emotional than by gender violence. A very complete overview disorders but who do not meet criteria or a diagnosis on how computers and algorithms can help in preat the current time. The collaboration of colleagues venting or detecting suicide risk is the one recently from the Clinical Psychology and Corpus Linguis- published by Ji [ 9 ]. Recent studies have found that tics fields, furthermore, provides the project with automatic processing of social media communicathe necessary interdisciplinary approach to obtain tions is an efective way to detect suicidal ideation robust results which may be significant to society. by applying emotion and sentiment analysis over

Joint eforts of NLP with Corpus Linguistics and textual messages [ 10 ].

Clinical Psychology are sought in this project with a NLP techniques are being applied to the analysis two-fold purpose: a) to analyse the results obtained of social media textual data to face new problems from the linguistic point of view to fine-tune and like fake-news detection [ 11 ], ofensive language idencomplement the NLP findings; and b) to contrast tification [ 12 ], sentiment analysis [ 13 ], opinion minthe results with the scientific literature on these ing and emotion detection [ 14 ]. Social Big Textual disorders in Clinical Psychology. Data is challenging, because language varies across time and space, language register is informal, colloquial and full of idioms compared to formal forms 2. Participants and project funding of text. Artificial Intelligence has gained a lot of popularity in recent years thanks to advent of Deep The project brings together 3 partners from Uni- Learning techniques [ 15 ]. Nevertheless, many of the versity of Jaén: SINAI group from Advanced Stud- applications and problems overcome where already ies Center in ICT (CEATIC), Department of Psy- attempted with traditional algorithms in machine chology and Department of English Studies. This learning, heuristic approaches or knowledge-based project has been supported by the grant P20_00956 systems. The big diference to previous approaches (PAIDI 2020) funded by the Andalusian Regional is that current proposals are data-driven: they are Government. able to learn from large amounts of data and build models to perform diferent tasks with a level of 3. State of the art success never reached by other solutions. This shift has been especially dramatic for NLP.

It is estimated 24 million children and young people Linguistic-based methods have been surpassed by in the EU sufer from bullying every year, which end-to-end architectures, where no prior knowledge means that 7 out of 10 sufer some form of ha- on language is needed [ 16 ], but massive amounts rassment or intimidation, whether verbal, physi- of data are required. During the last two years cal or through new communication technologies [ 3 ]. we have witnessed the birth of amazing models Navarro-Gómez [ 4 ] stated that social networks allow like BERT [ 17 ], GPT-2 [ 18 ] or Transformer-XL [ 19 ], the viral difusion of degrading contents. Cyber- with impressive results in many diferent tasks. New bullying or electronic aggression has already been models seem to learn language linguistic nature from designated as a serious public health threat and data. has elicited warnings to the general public from the The gross research on NLP is turning towards Centers for Disease Control and Prevention (CDC) Transformer based models and exploring how far [ 5 ]. these architectures are able to learn and perform

In another study [ 6 ], approximately 1 out of 10 in human related tasks, being sentiment analysis, people were found to develop some sort of eating emotion detection and hate-speech identification, among them. project avoids the problems of fragmentation by

There are previous projects in the pursuit of sim- co-ordinating and developing joint activities related ilar goals, like the STOP project [ 20 ] or MENHIR to early identification in order to coordinate high [ 21 ]. The Big Hug project is not only focused in quality transnational research. The diferent perexploring algorithm and models for early detection spectives and especially the diferent qualifications of disorders, but also in finding efective ways to of mental-health, applied linguistics and Informatransfer these systems to real world applications. tion and Communication of Technologies (ICT) specialists working in academia could stimulate the discovery of new and creative solutions. Apart from 4. Objectives of the project multidisciplinarity, there are relevant transversal aspects in the project.

The main objective is clear: a multidisciplinary project for the research on methods and algorithms to analyse textual streams across time and discover patterns for an early detection of potential harmful situations or behaviours. This global goal can be divided into the following sub-objectives: 1. To identify valid technologies for “listening”

the interactions in digital environments. 2. To model diferent forms of aggressive com

munication or risky situations. 3. To identify young people at high risk, but by the very first time, via a screening of altogether big data, psychological, linguistic variables. 4. To facilitate the replication of the screening protocol based on a well-defined methodology and analysis plan, if the previous objective is met. 5. To enhancement of our capabilities to feed these artificial intelligences with quality data by means of new techniques and methods to process informal language or colloquial expressions. 6. To adapt human language technologies also to the specific one that is usually used to make apologia of those scenarios. 7. To explore practical solutions which may be

integrated in the real world.

5. Conclusion

Dispositions for eating, anxiety and depressive disorders, are multifactorial. Big Hug represents a novel approach for mental disorders, integrating mental health, big data and linguistics measures as predictive measures for early diagnosis.

Research on mental health, for the early diagnosis and treatment of emotional mental health problems in the young is fragmented as researchers have traditionally worked in isolation and few studies examined the same or more than a limited set of risk factors, neglecting novel stratification strategies and development of algorithms. The Big Hug

References

[1] D. E. Losada , F. Crestani , J. Parapar , Overview of erisk 2019 early risk prediction on the internet , in: International Conference of the CLEF for European Languages , Springer, 2019 , pp. 340 - 357 .

[2] J. Parapar , P. Martín-Rodilla , D. E. Losada , F. Crestani , eRisk 2021 : pathological gambling, self-harm and depression challenges , in: ECIR , Springer, 2021 , pp. 650 - 656 .

[3] E. Cross , R. Piggin , T. Douglas , J. VonkaenelFlatt , Virtual violence ii: Progress and challenges in the fight against cyberbullying , London: Beatbullying ( 2012 ).

[4] N. Navarro-Gómez , El suicidio en jóvenes en españa: cifras y posibles causas . análisis de los últimos datos disponibles, Clínica y Salud 28 ( 2017 ) 25 - 31 .

[5] E. Aboujaoude , M. W. Savage , V. Starcevic , W. O. Salame , Cyberbullying: Review of an old problem gone viral , Journal of adolescent health 57 ( 2015 ) 10 - 18 .

[6] E. Stice , M. J. Van Ryzin , A prospective test of the temporal sequencing of risk factor emergence in the dual pathway model of eating disorders ., Journal of Abnormal Psychology 128 ( 2019 ) 119 .

[7] T. Wang , M. Brede , A. Ianni , E. Mentzakis, Detecting and characterizing eating-disorder communities on social media , in: Proceedings of the Tenth ACM International conference on web search and data mining , 2017 , pp. 91 - 100 .

[8] J. Wang , X. Wu , W. Lai , E. Long , X. Zhang , W. Li , Y. Zhu , C. Chen , X. Zhong , Z. Liu , et al., Prevalence of depression and depressive symptoms among outpatients: a systematic review and meta-analysis , BMJ open 7 ( 2017 ) e017173 .

[9] S. Ji , S. Pan , X. Li , E. Cambria , G. Long , Z. Huang , Suicidal ideation detection: A review of machine learning methods and applications , IEEE Transactions on Computational Social Systems 8 ( 2020 ) 214 - 226 .

[10] J. J. Glenn , A. L. Nobles , L. E. Barnes , B. A. Teachman , Can text messages identify suicide risk in real time? a within-subjects pilot examination of temporally sensitive markers of suicide risk , Clinical Psychological Science 8 ( 2020 ) 704 - 722 .

[11] F. Monti , F. Frasca , D. Eynard , D. Mannion , M. M. Bronstein , Fake news detection on social media using geometric deep learning , arXiv preprint arXiv: 1902 . 06673 ( 2019 ).

[12] M. Zampieri , S. Malmasi , P. Nakov , S. Rosenthal , N. Farra , R. Kumar , Semeval -2019 task 6: Identifying and categorizing ofensive language in social media (oefnseval ), arXiv preprint arXiv: 1903 . 08983 ( 2019 ).

[13] E. Martínez-Cámara , M. T. Martín-Valdivia , L. A. Urena-López , A. R. Montejo-Ráez , Sentiment analysis in twitter , Natural Language Engineering 20 ( 2014 ) 1 - 28 .

[14] F. M. Plaza-del Arco , M. T. Martín-Valdivia , L. A. Ureña-López , R. Mitkov , Improved emotion recognition in spanish social media through incorporation of lexical knowledge , Future Generation Computer Systems 110 ( 2020 ) 1000 - 1008 .

[15] J. Dean , D. Patterson , C. Young , A new golden age in computer architecture: Empowering the machine-learning revolution , IEEE Micro 38 ( 2018 ) 21 - 29 .

[16] T. Young , D. Hazarika , S. Poria , E. Cambria, Recent trends in deep learning based natural language processing , ieee Computational intelligenCe magazine 13 ( 2018 ) 55 - 75 .

[17] J. Devlin , M.- W. Chang , K. Lee , K. Toutanova , Bert: Pre-training of deep bidirectional transformers for language understanding , arXiv preprint arXiv: 1810 . 04805 ( 2018 ).

[18] A. Radford , J. Wu , R. Child , D. Luan , D. Amodei , I. Sutskever , et al., Language models are unsupervised multitask learners , OpenAI blog 1 ( 2019 ) 9 .

[19] Z. Dai , Z. Yang , Y. Yang , J. Carbonell, Q. V. Le , R. Salakhutdinov , Transformer-xl: Attentive language models beyond a fixed-length context , arXiv preprint arXiv: 1901 . 02860 ( 2019 ).

[20] D. Ramírez-Cifuentes , A. Freire , R. BaezaYates, J. Puntí, P. Medina-Bravo , D. A. Velazquez , J. M. Gonfaus , J. Gonzàlez , et al., Detection of suicidal ideation on social media: multimodal, relational, and behavioral analysis , Journal of medical internet research 22 ( 2020 ) e17758 .

[21] M. Kraus , P. Seldschopf , W. Minker, Towards the Development of a Trustworthy Chatbot for Mental Health Applications , in: MultiMedia Modeling, Springer, 2021 , pp. 354 - 366 .