SIAS: Suicidal Intentions Alerting System
Georgios Domalis
1
, Christos Makris
1
, Pantelis Vikatos
1
, Anastasios Papathanasiou
2,3
,
Efterpi Paraskevoulakou
2
and Manos Sfakianakis
2,3
1
Computer Engineering & Informatics Department, University of Patras, Patras, Greece
2
Department of Informatics, University of Piraeus, Piraeus, Greece
3
Cyber Crime Division, Hellenic Police, Athens, Greece
Keywords:
Alerting System, Classification Model, Suicidal Intention.
Abstract:
In this paper, we present an alerting system based on an efficient classification model for detecting suicidal
people using natural language processing and data mining techniques. The model uses linguistic features
which are derived from an analysis of handwritten and electronic messages/notes. The model was trained and
validated with fully anonymised real data provided by the Cyber Crime Division of Greek Police as well as
available suicidal notes from social media. The alerting system is intended as a prevention, management tool
for automatic detection of suicidal intentions.
1 INTRODUCTION
Suicide is a serious problem of social and public
health perspective. According to recent World Health
Organization
1
more than 30,000 suicide deaths in the
United States and nearly 1 million suicide deaths
worldwide occur every year constituting suicide one
of the 20 leading causes of death. The Internet as well
as social media can play an important role in preven-
tion of suicide cases. In addition, current scientific
research (Burnap et al., 2015; Colombo et al., 2016;
Sueki, 2015) is focused on discovering patterns and
association between users’ behavior in social media
with suicidal intention. Regarding official statistics
from Greek Law Enforcement Agency
2
the rates of
interventions Cyber Crime Division, related to suici-
dal intentions have been extremely increased during
the current decade. The efficient investigation and
management of cases that involve a report of suicide
on the internet requires collaboration between Law
Enforcement Agency (Cyber Crime Division) and re-
sponsible ISPs. Law Enforcement Agency encoun-
ters difficulties in the investigation of such cases and
cybercrimes due to the strict legislative framework.
First of all the cross-border nature and the involve-
ment of many jurisdictions impede the investigation
and detection of such cases. Therefore, it is neces-
1
htt p : //www.who.int/mental health/prevention/en/
2
htt p : //www.hellenicpolice.gr
sary the international police cooperation in immedi-
ate time given the serious and urgent nature of these
cases (online suicides). However, there are a lot of
constraints and problems encountered by law enforce-
ment authorities in conducting investigations when-
ever the digital data and suicidal posts are stored on
computing cloud environments. More specifically,
many challenges that exist in these environments have
to be handled involving multiple levels, such as recog-
nition, collection, preservation, examination, inter-
pretation and reporting of digital data and evidence.
A major problem arises related to the fact that, even
if such data are identified, the characteristics of the
cloud environment complicate the collection of ev-
idence through legal procedure, since it is possible
by involving many jurisdictions. Another problem
is the large amount of data and information stored
in cloud computing environments, which adds diffi-
culty to identify those data related to the specific re-
search carried out every time. The policy and guide-
lines for addressing these challenges is imperative to
effectively investigate announced suicide cases on the
Internet and law enforcement generally on the World
Wide Web.
The goal of this paper is the implementation of
an alerting system called SIAS (Suicidal Intentions
Alerting System) for predicting suicide intention us-
ing an efficient classification model. The classifica-
tion model introduces a set of characteristics which
Domalis, G., Makris, C., Vikatos, P., Papathanasiou, A., Paraskevoulakou, E. and Sfakianakis, M.
SIAS: Suicidal Intentions Alerting System.
DOI: 10.5220/0006297402910297
In Proceedings of the 13th International Conference on Web Information Systems and Technologies (WEBIST 2017), pages 291-297
ISBN: 978-989-758-246-2
Copyright © 2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
291
are derived from analysis of texts and are related to
cohesion, emotional instability, part of speech, cer-
tain emotional lexicons and personality traits. Natu-
ral language techniques have been used for the fea-
ture extraction directly from the textual sources. We
train and evaluate machine learning algorithms in or-
der to conclude with, one which performs efficiently
in terms of F-measure for our datasets. We use pre-
annotated texts divided into two categories. The first
one constitute a set of citizens’ texts with the proven
suicidal intention and other with normal behavior pro-
vided by diverse sources.
The rest of the paper is structured as follows. Sec-
tion 2 overviews related work, we motivate our re-
search from current challenges regarding and related
studies. In Section 3, we provide an overview of the
system architecture that we propose with the detailed
description of each module. In Section 4 the classifi-
cation is mentioned with the data preparation and the
training/evaluation of the model. Section 5 overviews
details of the implementation of the system for mod-
ules and sub-modules respectively and presents a ref-
erence to our experimental results. Finally, in Sec-
tion 6, we discuss the strengths and limitations of our
approach and we conclude with an outlook to future
work.
2 RELATED WORK
The concept of prediction of suicidal people through
electronic or hand written messages has gained the
interest of researchers in the field of natural language
processing, sentiment analysis and machine learning
(Burnap et al., 2015; Thompson et al., 2014). In
Pestian et al. (Pestian et al., 2010) they introduce
methods that distinguish genuinely and elicited sui-
cide notes using a content analysis based on natural
language processing.
Also, several studies are focused on predicting mil-
itary and veteran suicide risk (Soltaninejad et al.,
2014; Thompson et al., 2014). (Thompson et al.,
2014) describes language models in order to detect
suicidal ideation from texts derived from social media
and specifically, Facebook posting. The results of this
study show that word pairs are more useful than single
words for model construction. Study (Burnap et al.,
2015) introduces an approach with text mining tech-
nique using classification models in order to catego-
rize each text derived from Twitter to multiple classes,
related to suicidal ideation and topics such as report-
ing of a suicide, memorial, campaigning and support.
Another study (Sueki, 2015) discovers the association
between suicide-related Twitter posts with suicidal
behavior. Especially this research focuses on young
people using the Internet and social media to identify
and predict suicide intentions based on the texts and
behavior of Twitter users. The linking of personality
traits with the suicidal ideation constitutes the aim of
study (Soltaninejad et al., 2014). Regarding to this re-
search personality traits such as neuroticism, extraver-
sion and conscientiousness traits may predict suicidal
ideation on the other hand traits such as agreeable-
ness and openness weakly correlate with suicide. An
approach using a vocabulary of topics which suici-
dal persons are used to talking is presented in (Ab-
boute et al., 2014). Twitter messages are parsed and
checked using the topics by a process combining nat-
ural language processing and learning methods to in-
dicate suicidal risky behavior. In addition, in the study
(Lightman et al., 2007) well-known tools for linguis-
tic analysis such as Coh-Metrix (Graesser et al., 2004)
and LIWC (Pennebaker et al., 2001) have been used
to investigate the correlation of features with suici-
dal and non-suicidal songwriter’s lyrics. In (Mairesse
et al., 2007) the authors present firstly a detailed cor-
relation analysis between Big Five personality traits
and the features contained in software, e.g. LIWC;
then they described an automatic recognition of user
personality traits using regression and classification
models. In another study (Stirman and Pennebaker,
2001) the word use of suicidal and non suicidal po-
ets are examined and correlations between the suici-
dal ideation and features from LIWC are extracted.
In this work, we deviate from the above approaches
due to the combination of linguistic features with per-
sonality traits which are strongly correlated with sui-
cidal intentions improving the performance of predic-
tion.
3 SYSTEM ARCHITECTURE
In this section, we introduce our model a detailed de-
scription of the alerting system architecture. The pro-
posed alerting system is focused to immediate identi-
fication of suicidal intentions.
The system is composed of the following modules:
The targets identification. The alerting system
starts by targeting of certain citizens (subjects)
which their psychological disorder could lead to
suicidal attempt. The targets are reported by cit-
izens, psychological clinics or Law Enforcement
Agency.
The user profile creation. This module deals with
the registration of targets’ personal info and the
creation of users’ profile. Coordinating Opera-
tional Center of Cyber Crime Division finds or
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
292
receives information about the suicidal intention
over the Internet or other communication tech-
nologies, following a thorough online police in-
vestigation, which includes correspondence with
concerned websites (Facebook, Twitter, Linkedin
etc.) and companies providing telecommunica-
tions and Internet Services (ISPs) in order to iden-
tify certain citizens.
The feature extraction. The module of feature ex-
traction accepts raw texts for each individual tar-
get. In this step, feature vectors are produced by
analysis of text, creating characteristics related to
cohesion and emotional instability, linguistic and
personality traits. The feature vector is introduced
to the classification model for suicidal intention
declaration.
The classification model. This module constitutes
the predictor of the suicidal intention. The clas-
sification model has been trained and evaluated
with texts derived from, Cyber Crime Division,
open data and posts from well-known social me-
dia. According to our approach the prediction of
suicidal intention constitutes a binary classifica-
tion problem in which a feature vector is placed
in a category (Y/N).
The monitoring & alerting. The monitoring &
alerting module is actually the platform which
presents in real time the indicators of all targets
as well as the classification.
4 CLASSIFICATION MODEL
The efficiency of the proposed system depends on the
performance of classification model. The main chal-
lenge that we face is the lack of available data sources
due to the nature and anonymity of this type of data.
Cyber Crime Division of Greek police intervenes in
cases related to suicidal intentions in which citizens
inform of their attempt in social media, blogs or open
sources. Respecting personal information and accord-
ing to ethical principles and Greek legislation, a real
text corpus (fully anonymised) of suicidal cases, in
which Cyber Crime Division intervened in the past,
was used for the evaluation of our system. In addi-
tion the dataset was enriched with text with normal
behavior provided by diverse sources. Furthermore,
we added suicidal notes from two social media Tum-
blr
3
and Whisper
4
. Also, we introduced a detailed
graph model that makes easy the decision whether a
3
htt ps : //www.tumblr.com/
4
htt ps : //whisper.sh/
text implies or not a suicidal manifestation as in it is
shown in Figure 2; we explain this in the sequel.
4.1 Data Preparation
In this section, there is a brief description of the fea-
tures which are used for the training procedure of dif-
ferent classification algorithms. The choice of co-
hesion and linguistic metrics were based on previ-
ous research concerning the relationship between psy-
chological health and language use as it is described
in (Lightman et al., 2007; Stirman and Pennebaker,
2001). In addition, we introduce personality traits as
additional characteristics.
There is a fruitful discussion through academics about
the relation of personality traits with suicidal in-
tention (Carrillo et al., 2001; Kotrla Topi
´
c et al.,
2012; Rozanov and Mid’ko, 2011; Soltaninejad et al.,
2014). Figure 2 presents a graph with the association
of the ve basic dimensions of personality, that re-
main stable in individuals forming the Big Five Model
(McCrae and John, 1992), with two factors of Auto-
Destructive Behavior and Depression which imme-
diate influence the suicidal ideation. The Depres-
sion and Auto-Destructive Behavior are recognized
as significant predictors of various suicidal manifes-
tations, in particular, phenomena of hopelessness and
suicidal ideation (Kotrla Topi
´
c et al., 2012; Rozanov
and Mid’ko, 2011). According to Big Five, the hu-
man personality is described as a vector of five val-
ues of traits. The combination of Big Five personal-
ity dimensions explain the dynamics of a personality.
For example, a person may be very talkative (high
Extraversion), not very tolerant and sensitive (low
Agreeableness), systematic and punctual (high Con-
scientiousness), easily anxious (high Neuroticism)
and extremely curious (high Openness). Each per-
sonality trait correlates differently with Depression
and Auto-Destructive Behavior. More specifically
individuals that present open to fantasy, but not to
actions are prone to harmful Auto-Destructive in-
tentions and depression respectively (Carrillo et al.,
2001). Furthermore, an extrovert person seems to be
less risky to deal with depression or to harm himself
(Kotrla Topi
´
c et al., 2012). Also, people that lack
of sufficiency, self-discipline, self-control in decision
making increase the probability of belonging to the
depression and suicidal class (Rozanov and Mid’ko,
2011; Soltaninejad et al., 2014). In study (Soltanine-
jad et al., 2014) is mentioned that the high degree of
neuroticism which includes anger, hostility, impulsiv-
ity and vulnerability constitutes a significant of de-
veloping psychiatric disorders such as mood disor-
ders, depression which may lead to suicide (Solta-
SIAS: Suicidal Intentions Alerting System
293
Figure 1: Alerting system architecture
Figure 2: Association of Personality Traits with Suicidal
Ideation.
ninejad et al., 2014). Last but not least people who
react sensitively, tolerant and kind of situations could
lead confront with auto destructiveness and depres-
sion (Kotrla Topi
´
c et al., 2012).
The personality traits in Big Five model can be
extracted based on a questionnaire that determines
user personality as described in (John and Srivastava,
1999). However, we follow an approach for unsuper-
vised recognition of personality traits deriving from
text analysis (Celli, 2012).
Cohesion & Emotional Instability, Linguistic and
Personality Traits constitute the 3 sets of features as it
summarily described in Table 1. The prediction fea-
tures are produced by the linguistic analysis of texts
for each individual.
Cohesion and Emotional Instability Metrics:
Argument Overlap: This index measures aspects
of cohesion. The argument overlap index helps to
decide readability, complexity and grade level of
a particular corpus.
Latent Semantic Analysis (LSA): LSA is a mea-
sure of cohesion like argument overlap. The LSA
index compares the contextual similarity between
sentences recognizing text’s cohesiveness.
Word Concreteness: This feature measures the
terms’ concreteness of individuals in the written
texts.
Communication Words: This measure is the ra-
tio of communication words in the sentences by
using a certain dictionary of words, such as talk,
share and converse. We created these dictionar-
ies adding the synonyms from Wordnet (Miller,
1995).
Tense/Aspect Ratio: This index declares the de-
gree that sentences are linked by the time relation
in order to extract the overall temporal cohesion
on the sentences.
References to Time: The feature exposes the ra-
tio of verbs which are referred to future tense and
words related to time in general.
Death-Themed Words: This measure is the ratio
of death-oriented words in the sentences by us-
ing a certain dictionary. The dictionary includes
words related to death and dying and their syn-
onyms.
Swear Words: A list of vulgar expressions has
been created to extract the ratio of swear words
in the sentences.
Emotion Words: The ratio of words that address
affective or emotional processes. We divide these
words to positive and negative emotion.
Linguistic Metrics:
Punctuation metrics: The number of commas,
question marks, exclamation marks, parenthesis
and all punctuation divided by the total number
of tokens in the text.
Reference metrics: The number of @, first person
(singular and plural) pronouns divided by the total
number of tokens in the text.
Part of speech metrics: The number of first person
singular pronouns, prepositions, pronouns, first
person plural pronouns, second person singular
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
294
pronouns, numbers divided by the total number
of tokens.
Emoticons: The number of emoticons that express
negative and positive feelings divided by the total
number of tokens.
Word/token metrics: This feature set includes the
mean word frequency, the type/token ratio, the
number of words longer than 6 letters divided by
the total number of tokens. word count: the num-
ber of words of the text.
Special patterns: The number of external links
and the number negative particles divided by the
total number of tokens in the text.
Personality Traits:
Agreeableness (A): This personality dimension
includes attributes such as affability, tolerance,
sensitivity, trust and kindness.
Conscientiousness (C): Common features of
this dimension include organization, punctuality,
achievement-orientation and dependency.
Extraversion (E): This trait includes individuals
such as outgoing, talkative, sociable and enjoying
social situations.
Neuroticism (N): Individuals high in this trait
tend to be anxious, irritable, temperamental and
moody.
Openness (O): This trait includes curiosity, orig-
inality, intellectuality, creativity and openness to
new ideas.
4.2 Training & Evaluation
The feature vectors are enriched with the category for
each individual Y , N for suicidal and non-suicidal in-
tention respectively in order to form the datasets for
training and evaluation of the classifiers. We follow
the same approach as previous studies (Burnap et al.,
2015; Thompson et al., 2014; Abboute et al., 2014)
introducing the dataset to different classifiers to in-
vestigate the appropriate one which fits efficiently for
this type of data. The performance was evaluated by
F-measure. Also, the parameter selection is tested us-
ing a greedy approach in each classifier separately.
5 EXPERIMENTAL RESULTS
5.1 Implementation
We gathered 200 samples of suicidal and non-suicidal
intention related notes as it is shown in Table 2. The
Table 1: The User Profile Feature Vector.
Features # Description
Linguistic Metrics 21 Punctuation(5),
Reference(2),
Part-of-speech(6),
Emoticons(2),
Word/token(4),
Special patterns(2)
Cohesion Metrics 21 Tense/Aspect ratio,
References to
Time(3), Argu-
ment Overlap(9),
Communication
Words, Death
Themed Words,
Emotion Words(3),
Word Concrete-
ness, LSA, Swears
Words
Personality Traits 5 Openness,
Neuroticism,
Agreeableness,
Conscientiousness,
Extraversion
suicidal notes are derived from Cyber Crime Divi-
sion of Greek police in cases that information about
their suicide attempt has been exposed in social me-
dia, blogs and open sources. Furthermore, we use sui-
cidal notes from 2 well-known social networks (Tum-
blr, Whisper). The features were separated into 3 sets
which each one produces a different dataset. The first
set includes Cohesion & Emotional Instability Met-
rics (Cohesion), the second set includes the combi-
nation of the first set with Personality traits (Coh-
B5) and the third one the aggregation of all features.
We selected Naive Bayes, SVM, Rotation Forest, Ad-
aBoost and J48 as representative classifiers in order
to examine the performance for this type of data.
The classifiers are evaluated by the F-measure metric
which is the harmonic mean of precision and recall.
We separated each dataset to a train and a test set,
using two approaches:
K-Fold Cross-Validation (K=10 Fold).
Leave-One-Out Cross-Validation
The concept of using both techniques is that split-
ting with 10-Fold Cross-Validation, important infor-
mation can be removed from the train set. However,
the Leave-One-Out Cross-Validation technique evalu-
ates the classification performance based on one sam-
ple.
The text processing as well as the feature extraction
and training of classifiers was implemented in Python
SIAS: Suicidal Intentions Alerting System
295
Table 2: Number of Instances per Class.
Class Instances
Suicidal Intention 114
Non-Suicidal Intention 106
2.7 using NLTK
5
and scikit-learn
6
modules.
5.2 Results & Discussion
In this section, the results of our research are pre-
sented. We evaluate our dataset with 10-Cross-
Validation and Leave-One-Out methods. Figure 3 and
Figure 4 depict performance for all classifiers. Ac-
cording to the result, we observe that the performance
of the classifiers in terms of F-measure is better in the
dataset with the aggregation of all features than re-
moving the Linguistic and Personality Traits sets. We
consider that the add on features improves the per-
formance of all classifiers except SVM and Rotation
Forest, which deviates in the Coh - B5 set of features
in Leave-One-Out evaluation.
Rotation Forest classifier outperforms in comparison
to the other classifiers regarding F-measure. How-
ever Rotation Forest presents the highest deviation be-
tween the two methods of evaluation. Furthermore,
the use of linguistic metrics and personality traits in
the training of the classifier enhances the performance
10% approximately in the 10-Cross-Validation. The
combination of Cohesion metrics with Big Five traits
improves or remains the performance. The aggre-
gation of all features improves the F-measure which
is 0.805 in the worst case using J48 classifier and
reaches 0.895 in the best case using Rotation Forest.
In our understanding the use of Rotation Forest ben-
efits the efficient prediction of suicidal intention of
certain targets and enhance the alerting system. Also
the majority of the extracted features from text analy-
sis is language independent and our system can easily
be expanded to any language using a specific part of
speech and dictionaries for communicative, death and
swears words.
Table 3: Cohesion Metrics.
Classifier 10-CV Leave-One-Out
Naive Bayes 0.727 0.727
SVM 0.795 0.804
Rotation Forest 0.8 0.814
AdaBoost 0.764 0.736
J48 0.704 0.677
5
htt p : //www.nltk.org/
6
htt p : //scikit learn.org
Table 4: Cohesion Metrics and Big-Five.
Classifier 10-CV Leave-One-Out
Naive Bayes 0.737 0.727
SVM 0.777 0.786
Rotation Forest 0.804 0.786
AdaBoost 0.764 0.736
J48 0.741 0.713
Table 5: All Metrics.
Classifier 10-CV Leave-One-Out
Naive Bayes 0.836 0.764
SVM 0.841 0.85
Rotation Forest 0.895 0.868
AdaBoost 0.818 0.832
J48 0.805 0.814
Figure 3: 10-cross validation (F-measure).
Figure 4: Leave-One-Out (F-measure).
6 CONCLUSIONS
In this work, we looked into the design of the alert-
ing system (SIAS) for the immediate intervention of
authorities in suicidal intention. Our methodology in-
cludes 5 phases with the Targeting, Using Profile Cre-
ation, Feature Extraction, Classification and Alerting.
This paper has concentrated on building an efficient
classifier with natural language processing and data
mining techniques to predict the suicidal intention us-
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
296
ing features derived from text analysis. The results
depict with high accuracy the suicidal manifestation
in the text of users in danger to commit suicide.
As a future work, we are considering to explore al-
ternative factors which are not introduced in our ap-
proach and potentially influence the prediction of sui-
cidal intention such as the behavior in social net-
works.
REFERENCES
Abboute, A., Boudjeriou, Y., Entringer, G., Az
´
e, J.,
Bringay, S., and Poncelet, P. (2014). Mining Twitter
for Suicide Prevention, pages 250–253. Springer In-
ternational Publishing, Cham.
Burnap, P., Colombo, W., and Scourfield, J. (2015). Ma-
chine classification and analysis of suicide-related
communication on twitter. In Proceedings of the 26th
ACM Conference on Hypertext & Social Media,
HT ’15, pages 75–84, New York, NY, USA. ACM.
Carrillo, J., Rojo, N., Sanchez-Bernardos, M., and Avia, M.
(2001). Openness to experience and depression. Euro-
pean Journal of Psychological Assessment, 17(2):130.
Celli, F. (2012). Unsupervised personality recognition for
social network sites. In Proc. of Sixth International
Conference on Digital Society. Citeseer.
Colombo, G. B., Burnap, P., Hodorog, A., and Scourfield,
J. (2016). Analysing the connectivity and communi-
cation of suicidal users on twitter. Computer Commu-
nications, 73, Part B:291 300. Online Social Net-
works.
Graesser, A. C., McNamara, D. S., Louwerse, M. M., and
Cai, Z. (2004). Coh-metrix: Analysis of text on co-
hesion and language. Behavior research methods, in-
struments, & computers, 36(2):193–202.
John, O. P. and Srivastava, S. (1999). The big ve trait
taxonomy: History, measurement, and theoretical per-
spectives. Handbook of personality: Theory and re-
search, 2(1999):102–138.
Kotrla Topi
´
c, M., Perkovi
´
c Kova
ˇ
cevi
´
c, M., and Mla
ˇ
ci
´
c,
B. (2012). Relations of the big-five personality di-
mensions to autodestructive behavior in clinical and
non-clinical adolescent populations. Croatian medi-
cal journal, 53(5):450–460.
Lightman, E. J., McCarthy, P. M., Dufty, D. F., and McNa-
mara, D. S. (2007). Using computational text anal-
ysis tools to compare the lyrics of suicidal and non-
suicidal songwriters. In Proceedings of the 29th An-
nual Meeting of the Cognitive Science Society, pages
1217–1222. Citeseer.
Mairesse, F., Walker, M. A., Mehl, M. R., and Moore, R. K.
(2007). Using linguistic cues for the automatic recog-
nition of personality in conversation and text. J. Artif.
Int. Res., 30(1):457–500.
McCrae, R. R. and John, O. P. (1992). An introduction to
the five-factor model and its applications. Journal of
personality, 60(2):175–215.
Miller, G. A. (1995). Wordnet: a lexical database for en-
glish. Communications of the ACM, 38(11):39–41.
Pennebaker, J. W., Francis, M. E., and Booth, R. J. (2001).
Linguistic inquiry and word count: Liwc 2001. Mah-
way: Lawrence Erlbaum Associates, 71:2001.
Pestian, J., Nasrallah, H., Matykiewicz, P., Bennett, A., and
Leenaars, A. (2010). Suicide Note Classification Us-
ing Natural Language Processing: A Content Analy-
sis. Biomedical informatics insights, 2010(3):19–28.
Rozanov, V. A. and Mid’ko, A. A. (2011). Person-
ality patterns of suicide attempters: gender differ-
ences in ukraine. The Spanish journal of psychology,
14(02):693–700.
Soltaninejad, A., Fathi-Ashtiani, A., Ahmadi, K., Mir-
sharafoddini, H. S., Nikmorad, A., and Pilevarzadeh,
M. (2014). Personality factors underlying suicidal be-
havior among military youth. Iran Red Crescent Med
J, 16(4):e12686.
Stirman, S. W. and Pennebaker, J. W. (2001). Word Use in
the Poetry of Suicidal and Nonsuicidal Poets. Psycho-
somatic Medicine, 63(4):517–522.
Sueki, H. (2015). The association of suicide-related twitter
use with suicidal behaviour: A cross-sectional study
of young internet users in japan. Journal of Affective
Disorders, 170:155 – 160.
Thompson, P., Bryan, C., and Poulin, C. (2014). Predict-
ing military and veteran suicide risk: Cultural aspects.
In Proceedings of the Workshop on Computational
Linguistics and Clinical Psychology: From Linguis-
tic Signal to Clinical Reality, pages 1–6, Baltimore,
Maryland, USA. Association for Computational Lin-
guistics.
SIAS: Suicidal Intentions Alerting System
297