, , ,

The words of depression

With a prevalence of 3.8% of global population in 2019, roughly amounting to 280 million people, depressive disorders are among the mental health issues rising the highest concern worldwide. The Institute of Health Metrics and Evaluation at the University of Washington estimates that Major Depressive Disorder (MDD, i.e. the repetition of episodes of pervasive low mood and lack of interest and energy [1]) and dysthymia (a longstanding depression of mood which never reaches the severity of MDD) affect more than 15 million people in the USA, 20 million in Western Europe and 2.6 million in Italy [2].

In view of such numbers, the danger posed by such disorders clearly extends from the individual to the social sphere, urging experts to understand their causes besides their symptoms. Several theories of depression have been formulated so far, each focusing on certain aspects of the disorder, but the lack of appropriate tools hindered extensive screening. Indeed, however impressive the aforementioned estimates may appear, underdiagnosis and late diagnosis remain open issues. Adequate and extensive screening therefore plays a pivotal role both to deepen the study of the disorders and to widen the scope of the diagnosis process.

For these purposes, researchers have started experimenting with unconventional assessment methods, going beyond traditional questionnaire and interview-based screening. Most of these new approaches rely on two building blocks:

  • Social media, as a source of authentic and useful data about people’s lives, feelings and thoughts;
  • Machine learning, as a set of techniques to analyze the data and produce reliable assessments.

Naturally, the exploration of these unconventional means hasn’t come without risks, and worries for privacy and the socio-political consequences of its breach have promptly been risen. At the same time, far from dismissing these concerns, the undeniable potential of the new approaches has attracted special interest in both computer science and psychology communities. In other words, whether the processing of information spontaneously disclosed online in the form of writings[3], images[4], as well as activity and connections [5], can yield reliable and significant insights about people’s mental health has become a crucial question in the light of the aforementioned challenges.

Insofar as the goal is limited to a binary prediction between depressed and non-depressed categories, results obtained so far suggest a cautiously positive answer. Machine learning techniques have proven able to assess whether a social media user suffers from a depressive disorder in a fairly accurate, consistent and generalizable way. But another research question rises in this context: is it possible to automatically suggest not only whether a social media user suffers from a depressive disorder but also what triggers his/her suffering? Or, in more proper words: is it possible to employ automated means to discover insights about what aspects of a person’s narrative about herself and the world involve dysfunctional thinking schemas?

In the domain of textual data, which is what we are working on but also the most common mean by which a person expresses herself on the web, the aim of extracting aspects, themes, triggers or causes of depression symptoms practically translates into extracting words or spans of text. Several machine learning approaches address similar tasks in diverse domains, for instance aspect-based sentiment analysis, but the challenge here is complicated by two factors: the complex manifestation of depressive thinking schemas, and the lack of annotated data. Indeed, be it for social stigma or any other reason, people suffering from depression may not convey their feelings explicitely, using sentences such as “I am depressed”. As for data availability, while some corpora do exist in English and Chinese languages, the difficulty intrinsic in annotating them with triggers or aspects of the symptoms, also due to the lack of a standard taxonomy, makes supervised approaches less appealing for this task. What are the alternatives then? The most dated one is given by rule-based (aka syntax-based) approaches, which apply a predefined set of syntactical patterns to extract words or phrases, but they aren’t the only option. Our first experiments, whose output is displayed in Figure 1, focused on the comparison between rule-based models and another kind of models which rely on the properties of Neural Networks (in particular Recurrent Neural Networks) to extract spans of text that are deemed relevant for emitting a depressed prediction. We may call these neural models without loss of generality.

Figure 1: Most frequent words given high importance by our neural model compared with two syntax-based models.

Are these words – the words that were most often assigned high importance by the neural model – what we may call the “words of depression”, indicating by this locution a lexicon of the aspects, themes, triggers and causes that constitute one’s dysfunctional narrative? And, secondly, are they different from the ones found by syntactical rules?

The explicitly depression-related lexicon seems to confirm the first hypothesis. Notice that clues are also hidden in other terms. For instance, previous research has highlighted a tendence of people affected by depression towards absolutist thinking schemas [6], reflected in words such as adjectives terrible and horrible, verb hate as well as adverbs really and always. Insulting and swearing may be symptoms of profound anger, which – together with anxiety and guilt – is often a symptom of psychological distress. What’s more, some frequent words indicate areas of life which may contribute to the person’s distress: people hints to social life, died to grief time, girl to relationships, work to professional life, etc.

As for the second question, it is interesting to notice that rule-based approaches capture general-meaning nouns and pronouns far more often. It is the case for thing, anything, something, someone, anyone. While these techniques get many interesting terms as well, neural models better capture specificity of terms. What’s more, the high degree of language informality in writings on social media favors approaches that are not based on formal syntax. Far from providing a global and comparative rating of these models, such considerations encourage experiments with neural models, in particular for what concerns interpretability of results through aspect extraction.

Grasping what a person feels by her words alone is a difficult task even for humans, understanding what is inducing that feeling is even harder. However limited and risky machine learning techniques may be for mental health screening, aspect-oriented approaches raise at least the possibility to figure out an interpretation for a depressed prediction. How to use that interpretation is generally in the hands of the practitioner, who may in the future find himself or herself collaborating with artificial intelligence to improve screening and diagnosis capacity, and the researcher, who may use such information to spot the biases of the models and improve them accordingly. Despite being very preliminary, the results achieved so far open a research area that will have to deal with the complexity of human language as well as other sources of information if the goal is to provide substantial support in the social struggle against mental health disorders. It will have to deal with time, as writings are produced diachronically, and with articulated clauses, as meaning is rarely conveyed by single words. We will see whether state-of-the-art Machine Learning models can handle such a task.


[1] WHO, L. Zivetz, and WHO Staff. The ICD-10 Classification of Mental and Behavioural Disorders: Clinical Descriptions and Diagnostic Guidelines. ICD-10 classification of mental and behavioural disorders / World Health Organization. World Health Organization, 1992

[2] Global burden of disease study 2019 (GBD 2019). http://ghdx.healthdata.org/gbd-results-tool?params=gbd-api-2019-permalink/d780dffbe8a381b25e1416884959e88b, 2020.

[3] Jana M Havigerova, Jirı Haviger, Dalibor Kucera, and Petra Hoffmannova. Text-based detection of the risk of depression. Frontiers in psychology, 10:513, 2019.

[4] Reece, Andrew G., and Christopher M. Danforth. “Instagram photos reveal predictive markers of depression.” EPJ Data Science 6.1 (2017): 15.

[5] Munmun De Choudhury, Michael Gamon, Scott Counts, and Eric

Horvitz. Predicting depression via social media. In Seventh international AAAI conference on weblogs and social media, 2013.

[6] Mohammed Al-Mosaiwi and Tom Johnstone. In an absolute state: Elevated use of absolutist words is a marker specific to anxiety, depression, and suicidal ideation. Clinical Psychological Science, 6(4):529–542, 2018.

  • Davide Riva

    Research fellow since June 2022, Davide Riva holds a BSc in “Mathematical Engineering” and a MSc in “Data Science and Economics”. Currently, he mainly works on the NGUPP project, maintaining also a deep interest in NLP and machine learning applications for psychology.