2. Related Works
Studies such as Zaghouani [3] has looked into a big social media
corpus for identifying youth depression. The author suggests developing
a sentiment analysis-based linguistically annotated corpus to study teen
online behaviour across the MENA region. The authors want to eventually
compile a sizable user base with more precise self-reported sadness
signals. A sentiment analysis and emotion recognition-based automated
psychometric analyzer for healthcare has been looked at by Vij [4].
They asserted and came to the conclusion that the main objective of the
proposed work is to develop a self-service medical kiosk or a
psychometric analyzer with fast computational linguistics capabilities
that can produce a brief, concise summary of the patient’s emotional
health based on previous records, medications, and treatments. Almouzini
et al. [5] looked into finding Arabic Twitter users that were
depressed. They claimed to have developed a prediction model based on
the identification of depressed individuals using an Arabic sentiment
analysis employing supervised learning to assess whether a user’s tweet
is depressed or not. They found that sad people are more socially
isolated. Priya et al. [6] have suggested ML algorithms for
predicting stress, depression, and anxiety in contemporary life. They
claimed that in this study, ML algorithms were employed to evaluate five
different levels of stress, depression, and anxiety. They discovered
that random forest has the highest accuracy (91% and 89%). A research
has been done by Feuston et al. [7] on how mental illness is
expressed on Instagram. They explained how their individual histories,
viewpoints, and experiences with mental illness and health had an impact
on how they understood the findings. Murnane et al. [8] have
suggested designing technology to support long-term mental health
management social ecologies. They paid close attention to the patient’s
perspective as well as the many viewpoints and experiences. A new class
of collaborative informatics infrastructures and interfaces aimed at
enabling the social ecologies of personal data activities will be built
using the widely applicable design principles they gave. Pater et al.
[9] have looked at a study of a case involving eating disorder
patients who used digital self-harm indicators. Future studies,
according to the report’s authors, might examine post-intervention data
and contrast it with pre-intervention data to assess changes in
patients’ online identity presentations. Among methods for identifying
depression in college students, Xu et al. [10] suggested employing
contextually-filtered characteristics and routing behaviour. In this
paper, they present a unique association rule mining-based technique for
automatically producing contextually filtered features that performs
better than existing feature selection techniques for a depression
diagnosis. Psycholinguistic patterns in social media texts have been
presented by Trifan et al. [11], which aid in our understanding of
depression. They are eager to discuss other psycholinguistic components
with those who can shed light on them through clinical papers in a
subsequent investigation. In a work that Mathur et al. [12]
suggested, suicidal intent was estimated using temporal psycholinguistic
clues. This study fills a gap by combining qualitative and quantitative
approaches to examine the effects of enhancing text-based suicidal
ideation identification.
There aren’t many statistics and publications about depression despite
the fact that it’s a serious mental health problem. Problems with NLP
are common today. The reason for this is likely a general lack of
interest in the subject. Another problem is that because the topic is
quite subjective, classifying such specific behavioural patterns may be
challenging. The work by Losada and Crestani [2] offers excellent
insight into this issue. Their dataset is the first to be used in
research on language use and depression. The details of this dataset,
which was also used in this investigation, are described in Section 3.
An Early Risk Detection Error (ERDE) measure was established by Losada
and Crestani [2] as a fresh evaluation statistic for their
methodology. This metric is concerned with the speed at which
affirmative circumstances can be found and the accuracy of assessments.
In a fascinating study, Wang et al. [13] employed sentiment analysis
to assess whether or not a user was depressed. It is advised that word
and artificial regulations be used to determine each micro-depressive
blog’s propensity. After that, a framework for detecting depression is
created using the suggested approach and ten psychologically confirmed
traits of depressed people. Since social networks have a lot of text
information, many researchers are seeking to build models based on ever
expanding data. In the years to come, using NLP with such a benefit to
address the growing depression problem may be adequate to delve further
into melancholy and provide doctors with fresh and intriguing
information. Another excellent method for assessing the mental health
and suicide risk of a community was provided by Benton et al. [14].
It has been demonstrated that gender modelling improves accuracy in
tasks involving social media text. For 10 prediction tasks, the authors
of these developed neural Multi-Task Learning (MTL) models. The outcomes
of their model demonstrated that choosing the MTL thinks that employing
the appropriate selection of auxiliary activities for a certain mental
state might result in a significantly better model. For situations with
the fewest data points, the model dramatically improves. The most
significant finding for our purposes was that gender prediction does not
adhere to the two aforementioned rules but rather improves performance
as a measure of a supplemental task.