Dissertation committee:
Σπυριδούλα Βαρλοκώστα, Καθηγήτρια, Τμήμα Φιλολογίας, ΕΚΠΑ
Γεώργιος Μικρός, Πανεπιστήμιο Hamad Bid Khalifa (HBKU), Κατάρ, Καθηγητής
Γεώργιος Μαρκόπουλος, Αναπληρωτής Καθηγητής, Τμήμα Φιλολογίας, ΕΚΠΑ
Χριστίνα Αλεξανδρή, Καθηγήτρια, Τμήμα Γερμανικής Γλώσσας και Φιλολογίας, ΕΚΠΑ
Μαρία Ιακώβου, Αναπληρώτρια Καθηγήτρια , Τμήμα Φιλολοφίας, ΕΚΠΑ
Στυλιανή Μαρκαντωνάτου, Ινστιτούτο Επεξεργασίας Λόγου, Ερευνητικό Κέντρο «Αθηνά», Ερευνήτρια Α’
Αλέξανδρος Τάντος, Επίκουρος Καθηγητής, Τμήμα Φιλολοφίας, ΑΠΘ
Summary:
The topic of this PhD dissertation is the detection of clinical depression in Modern Greek by analyzing data of users belonging to the social networking platform of Twitter. According to the American Psychiatric Association, depression is a mental disorder that impacts one in fifteen adults (6.7%), with an additional 16.6% of individuals encountering depression at some point in their lives. Nevertheless, in many cases, such among adolescents aged 12 to 17 (approximately 77%, as indicated by Schiller et al., 2013), individuals are not aware of their psychological state. Traditional methods of diagnosing depression involve questionnaires and individual examinations that are time-consuming, costly and depend on the individual's willingness and awareness.
Recent studies within the Computational Linguistics and Clinical Psychology Shared Task (Coppersmith et al., 2015) have leveraged data obtained from social media users diagnosed with depression. These studies have employed diverse methodologies with the aim of developing models for depression detection. The majority of them have centered their observations on the English language, focusing on discerning distinctive traits between individuals with depression and neurotypical ones. Primarily, the research has revolved around three key detection approaches: (i) emotion detection methods (Schwartz et al., 2014.), (ii) methods employing linguistic markers such as LIWC-categories (Pennebaker et al, 1999) or n-gram models (Coppersmith et al., 2015; Mitchell et al., 2015), and (iii) topic detection techniques (Resnik et al., 2013).
The role of language in discriminating psychological states has been a topic of interest since the 1960s, as demonstrated by the Gottschalk method (Gottschalk & Gleser, 1969). According to this, lexical features can reveal the extent of various psychological dimensions, including anxiety or social alienation. Features incorporated into models can be categorized into two primary types: (i) non-linguistic features, such as post frequency, repost percentages, follower counts, user demographic information, etc., and (ii) linguistic features, such as words carrying a negative emotional weight (De Choudhury et al., 2013a). The first set of indicators attempt to capture and quantify the social interaction factor, while the linguistic indicators enable the identification of two aspects: how individuals with depression express themselves (i.e., frequent use of 1SG personal pronouns) and what they discuss, namely their topics of interest (i.e., increased interest in medical issues and religious events).
The contribution of this dissertation lies in the attempt to adapt the methods required to identify depression through the observation of linguistic features in Modern Greek language. To this end, we created a depression corpus based on user self-reports. In addition, we collected two corpora for neurotypical users in two ways: (i) based on random selection and (ii) based on topic similarity. Subsequently, we implemented and compared several Machine and Deep Learning models in order to identify both the existence of specific linguistic markers and to create the first baseline model for detecting depression in Modern Greek. We employed two distinct sets of features: LIWC and TF-IDF. In this context, we adapted the LIWC (Linguistic Inquiry and Word Count) dictionary to Greek, considering the unique linguistic characteristics of the language. Moreover, the dictionary was evaluated both in terms of adequacy by running the lexicon on parallel textual corpora, and in terms of prediction by applying it to corpora of depressed language.
In summary, our study identifies notable linguistic indicators of depression, encompassing increased usage of 1SG personal pronouns, expressions of sadness, heightened interest in health-related topics, decreased participation in work-related activities, lowered motivation accompanied by diminished expectations for success, and frequent references to events in the Present Tense. The present PhD dissertation provides a basis for investigating depression in Modern Greek. In the future, it would be possible to create a tool that could serve as a starting point for the diagnosis of people suffering from depression in the early stages. In addition, the current models could be used as a reference and evaluated against data from formally diagnosed depressed patients.
