Deep Learning CNN-LSTM Approach for Identifying Twitter Users
Suffering from Paranoid Personality Disorder
Mourad Ellouze, Seifeddine Mechti and Lamia Hadrich Belguith
ANLP Group MIRACL Laboratory, FSEGS, University of Sfax-Tunisia, Tunisia
Paranoid Personality Disorder Detection, Deep Learning Architecture, Symptoms and Disease Detection,
Text Mining, Twitter.
In this paper, we propose an approach based on artificial intelligence (AI) and text mining techniques for
measuring the degrees of appearance of symptoms related to paranoid disease in Twitter users. This operation
will then help in the detection of people suffering from paranoid personality disorder in a manner that provides
justifiable and explainable results by answering the question: What factors lead us to believe that this person
suffers from paranoid personality disorder? These challenges were achieved using a deep neural approach,
including: (i) CNN layers for features extraction step from the textual part, (ii) BiLSTM layer to classify
the intensity of symptoms by preserving long-term dependencies, (iii) an SVM classifier to detect users with
paranoid personality disorder based on the degree of symptoms obtained from the previous layer. According to
this approach, we get an F-measure rate equivalent to 71% for the average measurement of the degree of each
symptom and 65% for detecting paranoid people. The results achieved motivate and encourage researchers to
improve them in view of the relevance and importance of this research area.
Paranoid personality disorder (PPD) is a psychologi-
cal disease marked by widespread and persistent in-
terpersonal distrust, in which others’ acts are mis-
understood as spiteful and malicious (Bernstein and
Useda, 2007). These side effects might cause in-
appropriate and unwanted behaviors (such as reck-
lessness, social isolation, insecurity and moodiness),
putting the patient in a situation of conflict with soci-
ety. The worst thing is that many complications are
associated with the therapy of this disease, as psy-
chiatric ailments are diagnosed differently than other
diseases. This distinction is due to the fact that symp-
toms are not tangible as the difficulty of breathing and
the feeling of oppression for people having Coron-
All these effects have contributed to the appear-
ance of several dangerous consequences existing fre-
quently in our era such as suicide, terrorism, etc.
Despite the danger of these diseases, we notice that
the number of people having psychological problems
is increasing, especially in less-developed countries
olves et al., 2006) since there is negligence about
different problems such as economic, social, etc. In
this context, the World Health Organization (WHO)
has declared that one for every four adults in the world
suffers from mental problems and in half of the coun-
tries of the world, there is one psychiatrist per 100,000
people. Furthermore, 40 percent of countries have
fewer than one hospital bed for mental diseases per
10,000 people (Organization, 2001). As a result, new
approaches based on artificial intelligence (AI) have
been increasingly used in recent years to automate
the work of identifying people with psychiatric issues
from raw data.
In this context, (Baumgartl et al., 2020) worked
on Electroencephalographic data, while other works
based their processing on speech data (Wang et al.,
2021). Despite their relevance, these works are lim-
ited since to ensure the proper functioning of their
systems it is necessary to have sophisticated equip-
ment (MRI, sensors, etc.) which makes the task of
the detection extremely challenging.
In this era, social media represents one of the most
conducive environments that allows their users to in-
teract and express themselves freely about everything
that happens in the world. In recent years, a signifi-
cant number of researchers have based their works on
data collected from social media. In fact the progress
and impressive development of computer technolo-
gies and tools has made the processing of the huge
Ellouze, M., Mechti, S. and Hadrich Belguith, L.
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder.
DOI: 10.5220/0011322300003266
In Proceedings of the 17th International Conference on Software Technologies (ICSOFT 2022), pages 612-621
ISBN: 978-989-758-588-3; ISSN: 2184-2833
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
amount of this data more attainable.
For that, our challenge in this work is to detect
people having a personality disorder by analyzing
their textual production on social media. To achieve
our goal, we divided this paper into two objectives:
1. Measure the appearance degree of each symp-
tom of paranoid disease for each user profile from
their textual data. That will ensure the reliability of
our final results by providing an explanation and jus-
tification for our decision.
2. Detect people having paranoid disease by tak-
ing advantage of the result of (1).
This work offers to Twitter the possibility to di-
agnose the state of their users in order to ensure their
well-being by detecting hidden information. In matter
of fact, users may not be aware of their mental state.
In addition, this work can allow for Twitter the possi-
bility of tracking the progress states of their users in
real time (for example Twitter can compare the degree
of symptoms related to a specific person in different
periods of time).
Starting with a state of the art in which we present
some studies done in this field and their limitations.
Then, we detail our methods with the different tools
used. Next, we discuss the results achieved. Finally,
we conclude our work with a conclusion and some
There are many obstacles related to the treatment of
data obtained from social media, since there are cer-
tain criteria that may intervene and influence user-
generated data. These criteria can include the age of
the person, country, level of education, etc. In ad-
dition, many users did not consider social media as
a formal framework for that, several users used in
their writing style irony, sarcasm, etc., which can dis-
rupt the treatment afterward. Moreover, we note a
violation of the language rules (punctuation, capital
letters, using terms that do not belong in a particu-
lar language’s lexicon, composing a sentence in more
than one language, etc.). However, many researchers
opt for text extracted from social networks for their
research works. In this context, several researchers
have used the data of social networks to detect violent
and extremist people (Rekik et al., 2019; Ahmad and
Siddique, 2017). Despite the difficulty found in the
processing of data extracted from social networks, the
objective of these studies remains achievable because
there is no concealed information, therefore we can
detect the distinct classes using a lexical technique
based on keyword searches. On the other hand, de-
tecting hidden information such as age, personality
traits and psychological problems is different.
For that in this case, we have to process a huge
volume of varied data. In this context, (Varshney
et al., 2017; Pramodh and Vijayalata, 2016; An et al.,
2018) worked on data obtained from several sources
and having different types to ensure the variation in
the data. Other researchers worked on the diversity of
the characteristics selected (Bleidorn and Hopwood,
2019; Gonz
alez-Gallardo et al., 2015; Celli and Lepri,
2018), for example some of them combined linguis-
tic criteria as morphological analysis, etc., meta-data
of Twitter such as number of friends and different in-
formation related to the tweet like number of words
or number of hashtag. Generally, the result of hid-
den information detection system contains an impor-
tant degree of uncertainty. For this reason, there
are a lot of researchers who used the statistical ap-
proach (Pramodh and Vijayalata, 2016; Ellouze et al.,
2020) instead of classical machine learning technique
(Stankevich et al., 2019; Mbarek et al., 2019) and
deep learning technique (Wang et al., 2019b; An et al.,
2018; Wang et al., 2019a) in order to guarantee the
notion of fuzzy logic. Among the drawbacks of dif-
ferent machine learning and statistical techniques is
that their results are not explanatory, for this reason
we found a lot of works that use the rule-based tech-
nique (Umar and Qamar, 2019; Muhammad et al.,
2019). Although the results of these rules are explana-
tory since they are based on cause and effect links, the
construction of these rules is very time consuming.
Other researchers focused on extracting use-
ful knowledge for doctors by detecting linguistic
specificities from the textual production of peo-
ple having psychological problems (Hall and Caton,
2017),(Schwartz et al., 2013). Among the results
found it by (Schwartz et al., 2013): (i) the extro-
vert people used more terms related to the lexicon of
friends and family. Besides, they used terms show-
ing positive feelings. Thus, (Baik et al., 2016) pro-
posed an approach to extract the relevant writing style
of each personality trait by associating for each trait
some categories of the most used subjects. This work
help authors to conclude that extrovert people are very
interesting in sports, shopping, hotels. Whereas intro-
verted people are more interesting in gaming.
After the analysis of the different mentioned pa-
pers, we note that most of authors have focused on the
detection of the consequences of psychological dis-
eases such as violence, terrorism or suicide (Rekik
et al., 2019; Ahmad and Siddique, 2017; Mbarek
et al., 2019), and only a few researchers who worked
on detecting personality disorders types (Haz et al.,
2022; Ellouze et al., 2021b; Ellouze et al., 2021a). In
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder
fact, we did not find any paper which treats paranoid
disease on social networks. Moreover, we note an ex-
cessive use of the English language, despite the ex-
istence of other languages with the same importance.
Besides, there are some problems related to the use of
the lexical approach (Salem et al., 2019). Generally,
this technique is based on the research of keywords
from the corpus (lexicons related to each class), so
among problems related to this technique is the diffi-
culty of finding a training corpus which includes all
lexicons related to a specific class. We note also that
results obtained (Stankevich et al., 2019; An et al.,
2018; Lin et al., 2017) are very abstracts and difficult
to interpret besides they need more explanation.
In this study, we propose an approach illustrated in
figure 1 that allows Twitter to analyse in real time
the textual production of their users in order to apply
the process of diagnosis committed by the psychiatrist
(listen to the patient, identify the different symptoms,
detect the disease).
This was done using a novel deep learning model
containing a set of convolution layers CNN for auto-
matic features extraction task, since we do not know
the criteria of distinction between classes. Next, BiL-
STM to make the classification of the degree of each
symptom of paranoid disease from the textual part
since it highlights the long-distance dependencies of
the textual part. Finally, SVM in order to detect para-
noid disease based on the degree of each symptom
since SVM is among the most configurable learning
algorithms (see figure 2).
In addition, our approach treats other problems at
the same time such as: (1) corpus imbalanced by us-
ing synthetic data generation step, (2) the lexical ap-
proach, by the use of sentence embedding technique
in order to detect the meaning of the word in the sen-
3.1 Preprocessing
In this step, we focused on preparing our corpus by
deleting unnecessary elements that do not distinguish
between classes in order to avoid negatively affect-
ing the subsequent processing, especially that our
work is based on data extracted from Web 2.0 (Petz
et al., 2015). This task was done by following these
steps: For the first time, we eliminated the various
stop words including articulatory words such as and,
also, therefore, etc. These words are used by any per-
son, so they do not help to make the distinction be-
tween the different classes. Next, we removed from
our corpus the different symbols used for expressing
money, time, number, etc. Then, we converted cap-
ital letters to lowercase letters and abbreviated terms
to their ordinary form to normalize our corpus using
the resource Google Graph Knowledge like AI to Ar-
tificial Intelligence. Finally, we converted the inflec-
tional forms of words to a common root to behave
similarly to words having the same common root as
transform, transformation, transforming, etc.. This
step was done using the library NLTK (Natural Lan-
guage Toolkit).
3.2 Features Generation
This step involves converting the textual data into
numerical vectors that can be handled by machine
learning algorithms. Based on our review of several
works, there are many ways to achieve this transfor-
mation such as Word Embedding (Bakarov, 2018).
However, the major problem of this technique is that
it does not preserve the meaning of the whole sen-
tence which makes it difficult for the algorithm to
measure the intention and the nuance existing in the
text. For that, we choose to work with sentence em-
bedding techniques as Universal Sentence Encoder
(USE) (Cer et al., 2018), InferSent (Reimers and
Gurevych, 2019), Sentence Bert (Feng et al., 2020),
etc. After an empirical study, we choose to work
with the Sentence Bert technique since this model is
trained on a large amount of data also it has a specific
architecture that allows it to learn deep bi-directional
representations, it accepts a large number of parame-
ters, making it more adaptable (Eke et al., 2021). This
technique is based on the calculation of the similarity
between sentences by applying pooling layers in order
to keep only important descriptors. In addition, this
technique provides as a result a set of standardized
vectors while settling many recognized issues related
to the size of the data set and the assortment of vocab-
ularies in the corpus. In our work, we have only relied
on textual data and we do not use other types of infor-
mation such as the number of retweets per user, or the
number of retweets per tweet since we work on data
that is obtained in a streaming manner, therefore at
the beginning, the value of these attributes is zero. In
addition, our approach is based on a deep learning ap-
proach so its specific architecture offers assistance to
distinguish automatically the relevant attributes from
the raw data.
ICSOFT 2022 - 17th International Conference on Software Technologies
Figure 1: Proposed approach for symptoms and paranoid disease detection.
Figure 2: Proposed Deep CNN–BiLSTM for symptoms and paranoid disease detection.
3.3 Features Extraction
The architecture of the Convolutional Neural Net-
work differs from the classic architecture of the
MLP (Multi Layers Perceptron) model, this differ-
ence mainly revolves around the convolutional part.
The purpose of this section is to automatically ex-
tract the characteristics and reduce the gross size of
the entry form to highlight the relevant characteris-
tics. For this reason, we used in our work the CNN
architecture since it performed well on different tasks
of natural language processing for capturing the syn-
tactic and semantic aspects (Ombabi et al., 2020). The
execution of this task was done by flowing the input
(tweets) through a succession of filters, the output of
these filters is called convolution maps. The result-
ing convolution maps are concatenated into a feature
vector called CNN code.
3.4 Synthetic Generation of Data
In this step, we aim to maximize the number of in-
stances, considering the difficulty of annotating the
data and the difficulty of getting balanced data. For
this purpose, there are numerous ways like Multi-
objective Genetic Sampling for Imbalanced Classi-
fication (E-MOSAIC) (Fernandes et al., 2019), Ex-
ploratory Data Analysis (EDA) for handling dupli-
cate records, Synthetic Minority Over-sampling Tech-
nique (SMOTE) (Chawla et al., 2002), etc. After
an empirical study, we choose to work with SMOTE
technique, since it has shown a great deal of success
in various applications and fields (Quan et al., 2021;
Ishaq et al., 2021) and our corpus is not linked to a
particular field. This technique uses the nearest neigh-
bors algorithm to produce new and synthetic data.
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder
3.5 Symptoms Classification
LSTM is an extension of RNN architecture that was
created to address RNN’s problem of explosion and
vanishing gradient since we may find in a situation of
lags of unknown duration between the different events
of a time series. Therefore, the LSTM network is
adapted to the classification, processing and the re-
alization of predictions based on time series data. In
our work, we choose to use BiLSTM in order to keep
the dependency links between the lexicon. For that,
we concatenated the output of the convolutional layer
to two LSTM layers in order to measure the depen-
dence between terms. The last layer is composed of
6 neurons (from 0 to 5) which measures the degree
of each symptom of the disease such as exaggerated
mistrust, negative interpretation of the gestures of oth-
ers, incessant doubt, etc. The reason for considering
the challenge as a classification rather than a regres-
sion is based on the low instances of certain degrees.
This classification step has been repeated nine times
(the most appeared number of symptoms of paranoid
disease), given that in our case it is a multi-label is-
sue, a person may simultaneously have more than one
3.6 Paranoid Classification
The task of disease detection is very sensitive, this
sensitivity is due to the absence of specific rules that
allow taking decisions. For example, it is not neces-
sary to have all symptoms in a person to affirm that
he has the disease. Moreover, it is difficult to find a
person with incessant doubt and self-estimation at the
same time. In addition, there are a huge number of
combinations of symptoms degrees. For that, in this
step, we took advantage of the degree of all symp-
toms detected in the previous step in order to build
a vector (represent the list of symptoms related to an
individual). Then, we passed this vector to the SVM
layer in order to make the detection of the disease. In
this part, the number of features is reduced for that
we limited ourselves by the classical classification al-
gorithms such as SVM, Naive Bayes, decision tree,
etc. After an empirical study we choose to work with
SVM algorithm.
This section presents the different details about our
dataset, LSTM settings for ”negative interpretation of
others’ actions” classification and an extract of our
results. This work has been implemented using the
python programming language which integrates the
Tensorflow framework.
4.1 Corpus
We applied our approach to data composed of a set of
tweets that included a vocabulary linked to the dis-
ease’s negative effects ”personality disorders” such
as ”I congratulate myself”, ”I am wary”, ”I am in
the confusion of”, etc.. This data was obtained us-
ing Apache Spark Streaming tool for tweets in French
language from 01-03-2020 to 30-05-2020.
Two psychiatrists were requested to doubly anno-
tate this corpus based on their knowledge and expe-
riences. The annotation process began with an em-
pirical study of a 10% part of the corpus in order to
better grasp the nuances related to the language of
social networks and to develop a manual of annota-
tion. After that, each annotator separately annotated
the 90% of the corpus. The annotations of both types
of classification are done independently, which means
for each user profile (20 tweets) each annotator gives:
(i) the degree of each symptom (a number between
0 and 5 where 0 indicates the absence of the symp-
tom and 5 indicates the high degree of the presence of
a symptom), (ii) their decision about the state of the
person ”paranoid person” or ”normal person”. We
consider a person with a paranoid personality disor-
der if in their last 20 tweets there is a redundancy of
linguistic indicators that show the symptoms of this
disease such as the semantic information indicating
terrible disturbance and fear as for example the fol-
lowing expressions ”my hair is standing on the end”,
”I can hardly breathe”, ”my throat gets knotted”, etc.
We set a limit of 20 tweets per user because we aim to
develop an approach able of recognizing people with
PD by the fewest number of tweets possible in or-
der to ensure early prevention while guaranteeing the
credibility of the results.
The selection of paranoid symptoms is based on
the most well-known symptoms of paranoid disorder
that we have chosen to present at levels 1-5 to be pre-
cise. After the annotation phase committed by the 2
experts in order to annotate the degree of presence of
symptoms as well as the existence of the disease, we
proceed to calculate the rate of agreement between
these 2 experts by using Cohen’s Kappa measure. In
this context, we obtained a value of 0.9 for the detec-
tion of the disease and 0.73 for the detection of the
degree of symptoms. Conflicting cases are mainly re-
lated to the misinterpretation of cases (misinterpreta-
tion in measuring the degree of the intensity of symp-
toms as well as between missing information or dili-
gence). For that, we asked our experts to meet again
ICSOFT 2022 - 17th International Conference on Software Technologies
and choose between (agreement or removing) con-
flicting cases. Tables 1 and 2 show in more detail the
distribution of tweets per class.
Note 1: In table 2 we present the number of para-
noid symptoms of each user presented in table 1,
which means one person that exists in table 1 can ap-
pear up to 9 times in table 2.
Note 2: It should be noted that in some cases we
encountered difficulties in the collection of data. For
example, for people who believe that they are always
right or that they are isolated, they do not need to talk
to others and try to persuade them. Particularly for the
first case, which have a tendency to be self-centered.
While in the case of ”reading hidden meanings in the
innocent remarks” and ”recurrent suspicions”, any-
thing can be a trigger and an incentive for these peo-
ple to write and show what is not expressed (hidden
Table 1: The distribution of instances for paranoid classifi-
Paranoid YES NO
Number of in-
280 users
(5600 tweets)
450 users
(9000 tweets)
Table 2: The distribution of instances for symptoms classi-
Symptoms Number of instances
aggressiveness 121 users (2420 tweets)
perceives attacks 163 users (3260 tweets)
recurrent suspicions 282 users (5640 tweets)
isolation 46 users (920 tweets)
believing they are always
76 users (1520 tweets)
read hidden meanings in the
innocent remarks
227 users (4540 tweets)
poor relationships with oth-
273 users (5460 tweets)
doubt the commitment 193 users (3860 tweets)
unforgiving and hold
187 users (3740 tweets)
4.2 Results
For the various settings applied to each layer in our
model, we applied three convolution layers, we ac-
cord for each of them 320 feature maps and an activa-
tion function ”Relu”. Moreover, three pooling layers
with a pool size of (1,9). Then, we employed two
LSTM layers composed of 250 neurons for the first
layer and 150 neurons for the second layer combined
with a hidden layer using ”softmax” as an activation
function and next with an output layer composed of
6 neurons (representing the degree of this symptom).
We repeated the execution of this task 9 times (num-
ber of symptoms) and in each case we predicted for
each symptom a value which represents the degree of
this symptom. The model of CNN input and output
with multiple parameters is presented in the table 3.
Table 3: Model parameter structure.
Layer type Output
Input Layer (768,1)
conv1d (Conv1D) (768, 320) 3200
max pooling1d (233, 320) 0
dropout (Dropout) (233, 320) 0
conv1d 1 (Conv1D) (233, 320) 921920
max pooling1d 1 (85, 320) 0
dropout 1 (Dropout) (85, 320) 0
conv1d 2 (Conv1D) (85, 320) 921920
max pooling1d 2 (9, 320) 0
dropout 2 (Dropout) (9, 320) 0
time distributed (1, 8000) 0
lstm (LSTM) (250) 1211000
lstm 1 (LSTM) (100) 240600
classification layer 6 906
Next, we passed the vector composed of the de-
gree of the nine symptoms obtained from the previous
step to a SVM layer in order to make the classification
of paranoid disease. We used SVM layer with a linear
kernel and scale gamma since our instances are lin-
early separable. We employ the Python programming
language to handle these various layers with their set-
tings. The following table 4 shows an excerpt of our
results for paranoid’s symptoms degrees detection.
4.3 Evaluation
We used the classical metrics of recall, precision, and
F-measure to assess the performance of each type of
classification (symptoms, disease). For the classifica-
tion of symptoms degrees, we calculated the stated
mentioned criteria for the results of the first out-
put layer of our model (degree of symptom i) which
means for each symptom of paranoid disease. For the
evaluation of the classification of paranoid disease,
we applied the mentioned criteria to the results of the
last layer of our model.
Note 1: The evaluation of the classification of
paranoid disease contains also the error rate figured
on the detection of symptoms degrees.
The two tables 5 and 6 display with more details
the evaluation of our approach. We compared the re-
sults of our work with a Baseline architecture com-
posed of CNN, BiLSTM, SVM which consists in pre-
dicting the disease directly without going through the
symptom prediction stage. This Baseline is inspired
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder
Table 4: Extract of results (translate to English) of paranoid’s symptoms degrees detection.
user’s tweets doubt the
unforgiving and
hold grudges
1. Il y a des gens vous
etes vrmt pourri jusqu’
a l’
ame. (There are
people you are really rotten to the core.)
2. Imagine t’aimes pas provoquer ? quelle vie fade. (Imagine you
don’t like to provoke? what a bland life.)
3. Depuis que je suis ma seule priorit
e jsuis beaucoup plus heureuse,
c¸a paye pas de faire passer les autres avant soi. (Since I’m my only
priority I’m much happier, it doesn’t pay to put others before your-
4. Cette nuit j’aurais pr
e ne pas r
ever (Tonight I would have pre-
ferred not to dream.)
5. Y’a d comportement que jsupporte plus (There is a behavior that I
can’t stand anymore.)
6. C’est quoi cette nouvelle mode des meufs de se vanter de parler
1939101 mecs c’est pas une fiert
e et c¸a le sera jamais. (What is this
new trend for girls to brag about talking to 1939101 guys? It’s not a
pride and it will never be.)
7. Maintenant
etre fid
ele c’est devenue une qualit
e alors que c¸a de-
etre normal. (Now being faithful has become a quality when it
should be normal.)
8. Y’en a vous
etes culott
e c’est incroyable. (There are some of you
who are cheeky, it’s incredible.)
9. T’as qu’a imaginer que c’
etait un r
eve et que tous ce qu’on a v
etait pas vrai. (You just have to imagine that it was a dream and that
everything we lived was not true.)
10. L’ingratitude est la pire des choses sah tu donnes tout
a des gens
qui se foutent de ta gueule.(Ingratitude is the worst thing sah you give
everything to people who make fun of you.)
11. Rencontrer quelqu’un avec le m
etat d’esprit que vous est
rare. (Meeting someone with the same mindset as you is rare.)
12. Le d
ut sa aide beaucoup a oublier. (Disgust helps a lot to
13. Arr
ete de croire que tout le monde te consid
ere comme tu les
erer ! (Stop believing that everyone considers you the way you
consider them!)
14. C’
etait des grandes paroles en l’air. (It was all talk and no action.)
3 2 2
by (Ombabi et al., 2020) in which the authors ad-
dressed the issue of classifying textual data from so-
cial networks. The purpose of this comparison is to
demonstrate the impact of the layer allowing the de-
tection of symptoms before the disease on the results.
The results of this comparison are illustrated in ta-
ble 7.
Note 2: For the evaluation task of our approach,
we applied K-fold cross-validation technique. It
should be noted that each time we switch between the
training folds and the test fold. This is due that, we ap-
plied the SMOTE technique only to the training folds
in order to not influence the evaluation results of our
Note 3: We were tolerated in the evaluation of
symptoms degrees classification results, this is at the
level of accepting the difference of +1 between the
real value and the predicted value (the opposite di-
rection is not accepted). However, this remains valid
except for cases where the value is between 1 and 5
which means our system has committed an error in
the choice of the degree and not in the existence of
the symptom.
This paper presented an intelligent approach based on
machine learning and text mining techniques. The ob-
jective of this approach is to measure the presence
degrees of symptoms in order to detect afterwards
paranoid disease among people using social networks.
This work meets the limitations presented in the re-
lated work section at the level that we respected the
logical passage to detect the disease (detect the symp-
toms then the disease) which makes our results pre-
cise, reliable and interpretable. In addition, we used
the full process of the deep learning approach (fea-
tures extraction and classification techniques) since
we do not know precisely what are relevant features
ICSOFT 2022 - 17th International Conference on Software Technologies
Table 5: Variation of Recall, Precision and F-measure ac-
cording to the model CNN+BiLSTM for symptoms classi-
Symptoms Recall
aggressiveness 80 73 76
perceives attacks 68 68 68
recurrent suspicions 61 59 60
isolation 86 77 81
believing they are al-
ways right
85 78 81
read hidden meanings
in the innocent re-
69 65 67
poor relationships
with others
58 56 57
doubt the commit-
72 66 69
unforgiving and hold
80 75 77
Table 6: Variation of Recall, Precision and F-measure ac-
cording to the selected classifier for paranoid disease detec-
tion based on symptoms classification results.
Softmax 60 61 60
60 57 58
KNN 62 62 62
AdaBoosting 65 60 62
Random Forest 61 60 60
SVM 66 64 65
Table 7: Recall, Precision and F-measure comparison of our
results with baseline results for Paranoid classification.
et al., 2020)
59 53 56
Our architec-
66 64 65
Improvement 7 11 9
offering assistance in distinguishing between differ-
ent classes. Moreover, we addressed issues associ-
ated with the size and the unbalanced corpus using the
technique of data generation. Thus, problems linked
to the lexical approach through the sentence embed-
ding technique ”BERT” that deals with the meaning
of words in the sentence. We got the most satisfac-
tory results (F-measure equal to 81%) for the follow-
ing symptoms classification degrees: believing they
are always right and isolation. We obtained the poor-
est results (F-measure equal to 57%) for the classifica-
tion of poor relationships with others symptom. This
difference is due to the language specificity linked to
each symptom class and the way of reacting of the al-
gorithm to each situation. In the same context, get the
best result for the classification of symptom believing
they are always right despite the lack of data com-
pared to the symptom poor relationships with others
for which we have enough data. This is justified by
the fact that the second symptom is measured by four
degrees with a high error rate, whereas in the first case
the symptom is measured using only two degrees.
Moreover, we can conclude that the task of data gen-
eration has helped us to overcome the problem of data
reduction. For the classification of the disease, we ob-
tained the best results using SVM algorithm since the
instances of our corpus are linearly separable. Re-
garding the average results obtained for the detection
of some symptoms compared to the results of disease
detection, this is due to the: (i) linguistic phenomena
such as irony, negations, etc, (ii) high number degree
of symptoms (not like binary classification of the dis-
ease), (iii) issues with using features in lexicon format
(general lexicon, an idea can be written in more than
one way, etc). Finally, this work gives Twitter the op-
portunity to track the state of its users (if there are new
symptoms that appear or disappear).
In this paper, we proposed a method to detect peo-
ple with paranoid personality disorder. This method
has an advantage compared to other works since it
provides explanatory results by detecting symptoms
of paranoid disease. Besides, it takes advantage of
a deep learning approach that combines at the same
time the extraction of features and the classification
tasks. Furthermore, it addresses the problems of un-
balanced data and reduced size of the corpus through
the task of generating data. The proposed method was
implemented and evaluated and the evaluation results
obtained are encouraging, indeed, the F-measure is
equal to 79%. As perspectives, we envisage testing
our method on other types of personality disorders
with a particular application field.
Ahmad, N. and Siddique, J. (2017). Personality assess-
ment using twitter tweets. Procedia computer science,
An, G., Levitan, S. I., Hirschberg, J., and Levitan, R.
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder
(2018). Deep personality recognition for deception
detection. In INTERSPEECH, pages 421–425.
Baik, J., Lee, K., Lee, S., Kim, Y., and Choi, J. (2016). Pre-
dicting personality traits related to consumer behavior
using sns analysis. New Review of Hypermedia and
Multimedia, 22(3):189–206.
Bakarov, A. (2018). A survey of word embeddings evalua-
tion methods. arXiv preprint arXiv:1801.09536.
Baumgartl, H., Dikici, F., Sauter, D., and Buettner, R.
(2020). Detecting antisocial personality disorder us-
ing a novel machine learning algorithm based on elec-
troencephalographic data. In PACIS, page 48.
Bernstein, D. P. and Useda, J. D. (2007). Paranoid person-
ality disorder.
Bleidorn, W. and Hopwood, C. J. (2019). Using ma-
chine learning to advance personality assessment and
theory. Personality and Social Psychology Review,
Celli, F. and Lepri, B. (2018). Is big ve better than mbti?
a personality computing challenge using twitter data.
In CLiC-it.
Cer, D., Yang, Y., Kong, S.-y., Hua, N., Limtiaco, N., John,
R. S., Constant, N., Guajardo-Cespedes, M., Yuan,
S., Tar, C., et al. (2018). Universal sentence encoder.
arXiv preprint arXiv:1803.11175.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,
W. P. (2002). Smote: synthetic minority over-
sampling technique. Journal of artificial intelligence
research, 16:321–357.
Eke, C. I., Norman, A. A., and Shuib, L. (2021). Context-
based feature technique for sarcasm identification
in benchmark datasets using deep learning and bert
model. IEEE Access, 9:48501–48518.
Ellouze, M., Mechti, S., and Belguith, L. H. (2020). Au-
tomatic profile recognition of authors on social media
based on hybrid approach. Procedia Computer Sci-
ence, 176:1111–1120.
Ellouze, M., Mechti, S., and Belguith, L. H. (2021a).
Approach based on ontology and machine learning
for identifying causes affecting personality disorder
disease on twitter. In International Conference on
Knowledge Science, Engineering and Management,
pages 659–669. Springer.
Ellouze, M., Mechti, S., Krichen, M., Belguith, L. H., et al.
(2021b). A deep learning approach for detecting the
behavior of people having personality disorders to-
wards covid-19 from twitter.
Feng, F., Yang, Y., Cer, D., Arivazhagan, N., and Wang, W.
(2020). Language-agnostic bert sentence embedding.
arXiv preprint arXiv:2007.01852.
Fernandes, E. R., de Carvalho, A. C., and Yao, X.
(2019). Ensemble of classifiers based on multiob-
jective genetic sampling for imbalanced data. IEEE
Transactions on Knowledge and Data Engineering,
alez-Gallardo, C. E., Montes, A., Sierra, G., N
arez, J. A., Salinas-L
opez, A. J., and Ek, J. (2015).
Tweets classification using corpus dependent tags,
character and pos n-grams. In CLEF (working notes).
Hall, M. and Caton, S. (2017). Am i who i say i am? unob-
trusive self-representation and personality recognition
on facebook. PloS one, 12(9):e0184417.
Haz, L., Rodr
ıa, M.
A., and Fern
andez, A.
(2022). Detecting narcissist dark triad psychological
traits from twitter.
Ishaq, A., Sadiq, S., Umer, M., Ullah, S., Mirjalili, S., Ru-
papara, V., and Nappi, M. (2021). Improving the pre-
diction of heart failure patients’ survival using smote
and effective data mining techniques. IEEE access,
olves, K., V
arnik, A., Schneider, B., Fritze, J., and Al-
lik, J. (2006). Recent life events and suicide: a case-
control study in tallinn and frankfurt. Social science
& medicine, 62(11):2887–2896.
Lin, H., Jia, J., Qiu, J., Zhang, Y., Shen, G., Xie, L., Tang,
J., Feng, L., and Chua, T.-S. (2017). Detecting stress
based on social interactions in social networks. IEEE
Transactions on Knowledge and Data Engineering,
Mbarek, A., Jamoussi, S., Charfi, A., and Hamadou, A. B.
(2019). Suicidal profiles detection in twitter. In WE-
BIST, pages 289–296.
Muhammad, A., Hendrik, B., and Iswara, R. (2019). Ex-
pert system application for diagnosing of bipolar dis-
order with certainty factor method based on web and
android. In Journal of Physics: Conference Series,
volume 1339, page 012020. IOP Publishing.
Ombabi, A. H., Ouarda, W., and Alimi, A. M. (2020).
Deep learning cnn–lstm framework for arabic senti-
ment analysis using textual information shared in so-
cial networks. Social Network Analysis and Mining,
Organization, W. H. (2001). The world health report 2001:
Mental health: new understanding, new hope.
Petz, G., Karpowicz, M., F
urschuß, H., Auinger, A.,
y, V., and Holzinger, A. (2015). Reprint of:
Computational approaches for mining user’s opinions
on the web 2.0. Information Processing & Manage-
ment, 51(4):510–519.
Pramodh, K. C. and Vijayalata, Y. (2016). Automatic per-
sonality recognition of authors using big five factor
model. In 2016 IEEE International Conference on
Advances in Computer Applications (ICACA), pages
32–37. IEEE.
Quan, Y., Zhong, X., Feng, W., Chan, J. C.-W., Li, Q., and
Xing, M. (2021). Smote-based weighted deep rotation
forest for the imbalanced hyperspectral data classifica-
tion. Remote Sensing, 13(3):464.
Reimers, N. and Gurevych, I. (2019). Sentence-bert: Sen-
tence embeddings using siamese bert-networks. arXiv
preprint arXiv:1908.10084.
Rekik, A., Jamoussi, S., and Hamadou, A. B. (2019). Vio-
lent vocabulary extraction methodology: Application
to the radicalism detection on social media. In In-
ternational Conference on Computational Collective
Intelligence, pages 97–109. Springer.
Salem, M. S., Ismail, S. S., and Aref, M. (2019). Personality
traits for egyptian twitter users dataset. In Proceedings
ICSOFT 2022 - 17th International Conference on Software Technologies
of the 2019 8th International Conference on Software
and Information Engineering, pages 206–211.
Schwartz, H. A., Eichstaedt, J. C., Kern, M. L., Dziurzyn-
ski, L., Ramones, S. M., Agrawal, M., Shah, A.,
Kosinski, M., Stillwell, D., Seligman, M. E., et al.
(2013). Personality, gender, and age in the language
of social media: The open-vocabulary approach. PloS
one, 8(9):e73791.
Stankevich, M., Latyshev, A., Kuminskaya, E., Smirnov, I.,
and Grigoriev, O. (2019). Depression detection from
social media texts. In Elizarov, A., Novikov, B., Stup-
nikov., S (eds.) Data Analytics and Management in
Data Intensive Domains: XXI International Confer-
ence DAMDID/RCDL, page 352.
Umar, A. and Qamar, U. (2019). Detection and diagnosis of
psychological disorders through decision rule set for-
mation. In 2019 IEEE 17th International Conference
on Software Engineering Research, Management and
Applications (SERA), pages 33–37. IEEE.
Varshney, V., Varshney, A., Ahmad, T., and Khan, A. M.
(2017). Recognising personality traits using social
media. In 2017 IEEE International Conference on
Power, Control, Signals and Instrumentation Engi-
neering (ICPCSI), pages 2876–2881. IEEE.
Wang, B., Wu, Y., Vaci, N., Liakata, M., Lyons, T.,
and Saunders, K. E. (2021). Modelling paralin-
guistic properties in conversational speech to detect
bipolar disorder and borderline personality disorder.
In ICASSP 2021-2021 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing
(ICASSP), pages 7243–7247. IEEE.
Wang, C., Wang, B., and Xu, M. (2019a). Tree-structured
neural networks with topic attention for social emo-
tion classification. IEEE Access, 7:95505–95515.
Wang, L., You, Z.-H., Chen, X., Li, Y.-M., Dong, Y.-
N., Li, L.-P., and Zheng, K. (2019b). Lmtrda: Us-
ing logistic model tree to predict mirna-disease as-
sociations by fusing multi-source information of se-
quences and similarities. PLoS computational biol-
ogy, 15(3):e1006865.
Deep Learning CNN-LSTM Approach for Identifying Twitter Users Suffering from Paranoid Personality Disorder