Sentiment Analysis of Serious Suicide References in Twitter Social

Network

Wael Korani

and Malek Mouhoub

Department of Computer Science, University of Regina, Regina, Saskatchewan, Canada

Keywords:

Sentiment Analysis, Twitter, Suicide Thoughts, Artiﬁcial Intelligence.

Abstract:

Sentiment analysis analyzes people emotions, attitudes, and opinion towards organizations, services, issues,

and individuals. Opinions are the core of almost all human activities because they consider a signiﬁcant

inﬂuencers of our behaviors. With the growing popularity of social media applications (micro-blogs, twitter,

comments, etc), users of these platforms express their emotions through their posts and comments. Suicide is

one of these dangerous emotions that threaten the public health of Canadians, and mortality form suicide is the

third leading cause of death in teenage. In this paper, we propose a suicide classiﬁer system called Auto Twitter

Suicide Detector System (ATSDS) that provides support to authorities to take appropriate actions in order to

protect communities from such kind of thoughts. The proposed twitter suicide detector system is a classiﬁer

system using data gathered from twitter to detect those related to suicide. Our system is built using deep neural

network on multi-purpose cluster computing system called spark. In order to asses the system performance,

in terms of accuracy, we have conducted several experiments and tuned neural network parameters to achieve

higher performance. The results returned are very promising.

1 INTRODUCTION

The International Statistical Classiﬁcation of Diseases

and Related Health Problems (ISCDRHP) refers to

suicide related behavior as ”intentional self-harm”.

Suicide-related behavior includes thoughts, behav-

iors, and communications related to suicide. In (Yip

et al., 2003), Yip et al. deﬁned suicide related ideation

as thoughts of ending one’s life or a wish to be dead.

A suicide attempt is one form of self-injury whereby

the attempt is to end one’s life. There are two types of

suicides: active and passive suicide. Active suicide is

an effective way of suicide and gives a slim chance of

interruption, such as hanging, shooting, and jumping

(Glass Jr and Reed, 1993). However, passive suicide

is a less violent way of suicide that allows interven-

tion, such as overdose, poisoning, and international

malnutrition, which is called indirect self destructive

behaviors (ISDBs). In (Conwell et al., 1996), Con-

well deﬁned ISDBs as ”an act of omission or com-

mission that causes self-harm leading indirectly, over

time, to the patient’s death”. ISDBs are common

among older adults who have suicide signs, such as

https://orcid.org/0000-0002-1419-1149

https://orcid.org/0000-0001-7381-1064

refusing to eat or drink and failing to take medications

(Brown et al., 2004).

Canadian Vital Statistics Death (CVSD) is re-

sponsible for reporting the cause of death in Canada,

which is an effective mechanism for monitoring the

death by suicide in Canada. In 2005, Public Health

Agency of Canadian Suicide reported on their web-

site that suicide was the eighth leading cause of death

for adults between (55-64) years (13.0 per 100,000).

In (Buchanan et al., 2006), Canadian Coalition for

Seniors Mental Health reported that older adults have

the highest rate of death by suicide across all age

groups. In (Navaneelan, 2012), Navaneelan reported

that the suicide rate in Canada declined from 12.7 per

100,000 between 1989 and 1992, down to 11.5 per

100,000 in 2009.

In the last couple of decade, social media plays a

crucial role in our social life. Most people want to be

in groups, where they can share ideas, experiences,

emotions, etc. Social media applications help people

share their ideas, problems, get solutions from other

like-minded people. In (Kaplan and Haenlein, 2010),

Kaplan and Haenlein deﬁned social media as ”a group

of internet-based applications that are built on the ide-

ological and technological foundation of Web 2.0 and

that allow the creating and exchange of user gener-

Korani, W. and Mouhoub, M.

Sentiment Analysis of Serious Suicide References in Twitter Social Network.

DOI: 10.5220/0008894003390346

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 339-346

ISBN: 978-989-758-395-7; ISSN: 2184-433X

339

ated content”. Users interact and share negative and

positive experience and learn from each other though

social media. Social media is available for any user

at any time. There is no limit in time and space on

social media, and users can share information at any

time and spread it in a second.

In 2006, Twitter was developed in a different way

of Facebook (Carlson, 2011). Twitter is another pop-

ular, widespread, and limited social network. It is

a micro-blogging that gives only 140 characters for

each message. An instance twitter message is called

tweet, and twitter friends are called followers. Posted

tweets from users’ friends will be shown on user’s

proﬁle page. Users on Facebook and Twitter can post

text, photo, link, or video. Twitter gives a user the

ability to create an instance message that introduces

an idea without any barriers. As the third quarter of

2016, the number of active Twitter users was grow-

ing each month, which was estimated to be around

317 million active users each month. Twitter becomes

more popular and has around 500 million instance

message every day.

Sentiment analysis can be performed using differ-

ent machine learning approaches. Pang suggested that

the current research on sentiment analysis focuses on

two major things: to identify the given text whether

it is subjective or objective. In addition, it may iden-

tify the polarity of the subjective texts (Pang et al.,

2008). Sentiment analysis has been used for range

of topics, such as movie review, products or services

reviews, political opinion, and emotions. In this pa-

per, we focus on sentiment analysis related to sui-

cide thoughts. Sentiment analysis was conducted on

suicide thoughts that have been reported using writ-

ten communication of suicide on the Web via bulletin

boards (Ikunaga et al., 2013). In (Matykiewicz et al.,

2009), unsupervised machine learning was also im-

plemented to distinguish between actual suicide notes

and newsgroups. Suicide thoughts are also released in

chat rooms with no restrictions (Becker and Schmidt,

2005).

Social media, specially twitter, along with senti-

ment analysis play a signiﬁcant role in improving the

suicide research by analyzing individuals activities

through their posts. In this paper, twitter social me-

dia is used to build a classiﬁer system, which is called

Auto Twitter Suicide Detector System (ATSDS). The

proposed system, TSDS, is capable of detecting twit-

ter users who have suicide thoughts or interested in

the suicide topic. The proposed system is built on

multi-purpose cluster computing system called spark

along with deep neural network. The accuracy of dif-

ferent models are evaluated to choose the best param-

eter for the neural network.

2 RELATED WORK

Few studies were conducted to use classiﬁcation

approaches to automatically identify suicide-related

communications in twitter and other social media.

Studies showed a strong positive correlation between

suicide rates and the volume of social media posts and

comments that related to suicide thoughts (Won et al.,

2013). In (Won et al., 2013), Won et al. concluded

that the social media data may help in national sui-

cide forecasting and preventing. Jashinsky suggested

that there is a relationship between suicide risk and

twitter conversation (Jashinsky et al., 2014). John et

al. analyzed twitter posts the 24 hours prior to the

death by suicide (Gunn and Lester, 2015). The results

showed that persons who committed suicide have pos-

itive emotions over the last 24 hours and a change in

focus from the self to others. Although the study con-

ducted over one case study, the authors later on used

more cases. The authors used the Linguistic Inquiry

and Word Count (LIWC) software to identify emo-

tional words (Pennebaker et al., 2001).

In (Poulin et al., 2014), Poulin et al. conducted an

experiment on a group of US war veterans who shared

their Twitter and Facebook over time. The authors

proposed a suicide prediction system based on clini-

cal notes of US war veterans, and the system showed

high performance (60% accuracy). In addition, the

authors concluded that persons who recorded fear, ag-

itation, and delusion behaviors had committed sui-

cide. In (Sueki, 2015), Sueki conduced an experiment

using posts of Twitter users to ﬁnd the relationship

between suicide-related tweets and suicidal behavior.

The results showed that some particular phrases such

as “want to commit suicide” was strongly associated

with lifetime suicide attempts. However, phrases that

suggest suicide intent, such as “want to die” have less

strong association with suicide, because such phrases

could be used when a person had a bad day. In (Ab-

boute et al., 2014), Abboute et al. proposed a sys-

tem to classify “risky” and “non risky” tweets with

accuracy 60%. The authors concluded a number of

emotions related to suicide, such as hurt, bulling, and

insults in the “risky” category.

3 DATASET OVERVIEW

Although a few studies were conducted to predict sui-

cide thoughts using machine learning models, there

is no reliable dataset was publicly published. Reli-

able dataset is one of the challenges in creating a ma-

chine learning model. The main reason that there is

no existing reliable dataset is that there is no agree-

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

340

ment about speciﬁc features to characterize suicide

notes. Thus, our dataset is pulled from Twitter and

distributed in ﬁles. The dataset has 1719 ﬁles that

contain 815871 tweets from different regions spe-

cially Canada. Tweets in our dataset are raw data that

we should ﬁrstly clean.

Our dataset has attributes, such as username,

country, time, location, posted message, etc. Tweets,

country, and city are all attributes that we need to cre-

ate our system. Tweets are ﬁltered based on the con-

tent of “suicide” word, and the resulted tweets are di-

vided into two classes “suicide” and “non suicide”.

The results show that the dataset includes 368 tweets

in suicide class. We then extended the ﬁlter using

some extra words that might have relationship with

the suicide thoughts in literature, such as killing my-

self, hate myself, hate this life, want to die, and hate

people. The results show more tweets belongs to sui-

cide class, which has 469 tweets. Finally, we mix the

suicide tweets with a new class of non suicide class

that has 1407 tweets. The entire dataset includes 1876

tweets. Figure 1 shows the histogram of the suicide

and non suicide classes. Figure 2 and 3 are two sam-

ples tweets of each class of these classes: suicide and

non suicide.

Figure 1: Histogram of suicide and non suicide tweets.

Figure 2: Sample of non suicide tweets.

Figure 4 and 5 show top regions in our dataset.

Figure 4 shows the top seven regions in the sui-

cide class. It shows that Regina/Canada, Saska-

toon/Canada, and Saskatchewan/Canada are the top

three regions in our dataset. However, Fig-

Figure 3: Sample of suicide tweets.

ure 5 shows the top eight regions in the non sui-

cide class. Regina/Canada, Saskatoon/Canada, and

Saskatchewan/Canada represent big share in our

dataset. In addition, the dataset contains signiﬁcant

number of tweets from Malta and Philadelphia/USA

in both classes.

The entire dataset will be used in training and

testing processes of creating the model, which causes

a bias in the model. We expect to get high accuracy

using this method. Then, the dataset will be divided

into two parts: training and testing dataset. The

training dataset represents 75% of the total dataset.

The training dataset will be used to create the model.

The rest of the dataset that represents 25% will be

used in testing the accuracy of the model. The testing

dataset is used to evaluate the performance of the

model. The accuracy of the neural network model

is used to evaluate the performance of our proposed

model.

Figure 4: Top seven regions of suicide class.

Second stage, 10-fold cross validation approach

is used, which is a recommended technique to avoid

producing a bias model. In this stage, the dataset is

divided into 10 parts where nine parts is used in train-

ing a model and one part is used in testing the created

model. This process is repeated for all combinations

of train-test splits. Figure 6 shows the process of ﬁve

folds cross validation. The cross validation process is

better than dividing the dataset into speciﬁc training

and testing data as shown in ﬁrst stage.

Sentiment Analysis of Serious Suicide References in Twitter Social Network

341

Figure 5: Top ten regions for the non related suicide tweets.

Figure 6: 5-fold Cross Validation Example.

4 PROPOSED MODEL

Our model is created using deep forward neural net-

work on Spark platform using Scala language as

shown in Figure 11. Spark is presented in this paper

for sentimental analysis of Twitter suicide posts. It

is an open source engine multi-purpose cluster com-

puting system for data processing. Spark is used in

many applications and among them machine learn-

ing applications. It has MLlip library that provides

a machine learning functionality, such as classiﬁca-

tion, clustering, regression, and prediction. MLlib has

two packages mllip (built on the top fo RDD) and ml

(built on the top of the dataframes). Spark is used for

all the operations that were implemented in this pa-

per such as training, cross validation, pipelines, clas-

sifying, and computing classiﬁer performance. These

operations reveal better understanding of the created

model. Parameters of the created model should be

tuned to ﬁnd the best values that improve the accu-

racy of the model.

4.1 Spark Core

Spark is an open source cluster computing developed

by UC Berkely AMPLap. In 2010, Spark is adopted

by Apache Software Foundation. Apache Spark is

an open source engine multi-purpose cluster comput-

ing system for data processing on a large scale. It

provides fast memory computing, and it consists of

high level tools such as Spark streaming, data frames,

SQL, MLlip for machine learning and GraphX for

graph processing as shown in Figure 7. The core en-

gine of Spark provides monitoring, scheduling, and

distributing of application across the computing clus-

ter. Spark is implemented in Scala language, which

runs on (JVM) Java Virtual Machine.

Spark has some great features: Spark API is

available in different languages, such as Scala, Java,

Python, and R. It runs on a web user interface for

checking, monitoring, results, and Spark jobs (Karau

et al., 2015). In the last few years, Spark becomes

very popular among the companies, such as eBay, Ya-

hoo, Amazon, Databrickes, Baidu, TripAdvisor, and

others.

Spark

SQL

Spark

Streaming

Apatche Spark

MLlib

Machine

Learning

GraphX

(graph)

Figure 7: Spark Stack.

4.2 Artiﬁcial Neural Network

In 1943, McCulloch proposed the ﬁrst mathematical

model of a neuron (McCulloch and Pitts, 1943). In

1958, Rosenblatt proposed the ﬁrst neural network

known as perceptron (Rosenblatt, 1958). In 2002, Yu

Hen proposed a mathematical computing paradigm

called artiﬁcial neural network that models the oper-

ations of biological neural system (Hwang and Hu,

2001). The building block of any neural network

model is the neuron. The neuron model that proposed

by McCulloch is the most widely used neuron, and the

multilayer perceptron is the most widely neural net-

work, which consists of several sequential connected

layers of perceptrons.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

342

There are several types of neural network, such

as feed-forward and recurrent networks. In the feed-

forward networks, the output signal of a neuron has

no inﬂuence on its inputs. However, the recurrent

networks, the output signals of neurons are feedback

given as their input signals. The multilayer perceptron

that has been used in our project is the feed-forward

networks.

4.2.1 Neuron Model

A neuron consists of net function and activation func-

tion (transfer function). Figure 9 shows few activation

functions that have been used in literature. However,

the net function is used to determine how the input

signals are combined inside the neuron. The formula

for net function is:

u =

∑

i=0

(1)

where w is the weight, and w

is the threshold and its

corresponding input x

is always equal one. In addi-

tion, the input x

does not form a connection between

two neurons as others do. The output of neuron is de-

noted by Y , which is the output of the net function u

by one of the activation function list in Figure 9.

Y=(f,w)

Figure 8: Neuron Model.

Figure 9: Commonly used transfer functions a - hyperbolic

tangent , b - logistic sigmoid , c - threshold.

4.2.2 Multilayer Perceptron Model

A single layer perceptron is able to classify only lin-

early separable data. A multilayer perceptron (MLP)

is a network that includes two or three layers of neu-

rons as shown in Figure 10. MLP consists of one input

layer and one output layer, and one or more hidden

layers. The MLP network is considered a fully con-

nected if every node in a given layer is connected to

every node in the next layer. It is used in many ap-

plications, because it has the ability to solve problems

that do not have an algorithmic solution or their solu-

tions are too complex to be found. Currently, artiﬁcial

neural network is used to solve problems that are un-

solvable using logical systems. Our model has one

input layer, two hidden layers, and one output layer

as shown in Figure 10. The weights of the layers are

optimized using an optimization technique during the

training process of the weights.

Input layer

Hidden layers

Output layer

u11

u48

v11

v86

w11

w62

Inputs

Outputs

Figure 10: Deep forward neural network.

4.3 Auto Twitter Suicide Detector

System (ATSDS)

The Limited-Memory BFGS (L-BFGS) is an op-

timization algorithm that approximates the Broy-

den–Fletcher–Goldfarb–Shanno (BFGS) algorithms

using limited memory. The basic idea of L-BFGS

is that it approximates a given objective function lo-

cally as a quadratic without calculating the second

partial derivatives of the objective function. Thus, L-

BFGS achieves faster convergence compared to the

ﬁrst-order optimization. It is a built in optimization

algorithm in MLlib, and it has several parameters,

such as Gradient, updater, numCorrections, maxNu-

mIterations, regParam, and convergence tolerance .

ATSDS is built using L-BFGS and DFNN as

shown in Figure 11. We study the effect of conver-

gence tolerance on the performance in terms of accu-

racy of our proposed system. The convergence tol-

erance controls how much change is allowed when

L-BFGS considered to converge. In our experiments,

convergence tolerance is tuned to achieve the best ac-

curacy of our proposed model.

Sentiment Analysis of Serious Suicide References in Twitter Social Network

343

L-BFGS

Apatche Spark

MLlib Machine Learning

DFNN

ATSDS

Figure 11: Auto Twitter Suicide Detector System (ATSDS).

5 EXPERIMENTATION AND

RESULTS

In our model, the DFNN is set to have four layers:

input layer, two hidden layer, and output layer. It is

represented as DFNN (input layer, ﬁrst hidden layer,

second hidden layer, output layer). The input layer

of the DFNN represents the number of features of

the model, and it is set to 100 features. Each of the

two hidden layers has 15 neurons. The output layer

has two neurons, because the model has two output

classes: suicide and non-suicide. The number of neu-

ron in each layer is chosen after conducting prelim-

inary experiments. In our experiments, we consider

tuning the convergence tolerance as an important pa-

rameter in our optimization algorithm (L-BFGS).

The convergence tolerance is changed in the rec-

ommended range [0.001 : 0.15] when the number of

neurons is ﬁxed to 15 in each of the hidden layers.

Then, we choose the best three convergence tolerance

values (0.001, 0.01, 0.015) to study along with differ-

ent number of neurons in hidden layers. The number

of neurons in each hidden layer is changed in range

[1 : 20] to achieve the best performance of our model.

The maximum number of iterations in each case is set

to 100,000 to create our ATSDS. The seed generator

is set to 1234 in all experiments so that the results will

be reproducible.

In order to evaluate the performance of our pro-

posed system, we conducted several experiments that

we report in this section. The accuracy is used to eval-

uate our system performance. The accuracy is the ra-

tio of the number of correctly predicted instances to

the total number of instances in the dataset. In our

experiments, we study the effect of the convergence

tolerance for L-BFGS and the number of neurons in

the hidden layers over the accuracy of our proposed

model.

5.1 Results and Discussion

In our experiments, the class label is whether it is

suicide related tweet or non-related suicide tweet.

Firstly, the model is trained on the entire dataset us-

ing two hidden layers of 15 neurons each, and then

the model is tested on the same entire dataset. The

result shows that the proposed model has high ac-

curacy (96.7%). However, this system has bias be-

cause the dataset for training and testing are the same.

Secondly, the dataset is then divided into two parts

75% training dataset and 25% testing dataset. The

results show that the proposed model has relatively

high accuracy 93.25%, but it is lower than the ﬁrst

case. In the second case, the model does not know

any information about the testing dataset. However,

when the train dataset is decreased to 25% and testing

dataset is increased to 75%, the accuracy of the pro-

posed model is decreased. In the third case, the cross

validation technique is implemented for building our

ATSDS, which is the most recommended technique

to avoid bias in the model. The number of neurons in

each hidden layer is tuned to achieve higher accuracy.

The results show that the convergence tolerance of our

optimization algorithm (L-BFGS) and the number of

neurons in the hidden layers have signiﬁcant effect on

the performance of our proposed model.

The effect of convergence tolerance and the num-

ber of neurons in each hidden layer are tested to pro-

duce a high accuracy model. The result shows that

number of hidden layers do not have that much effect

on the accuracy of the model. The best number of

neuron in the hidden layers is 15 as shown in Figure

12 and Table 1. Figure 12 shows the DFNN (100, 15,

15, 2), which has four layers. The input layer has 100

features; output layer has two classes suicide and non

suicide; each of the two hidden layers has 15 neurons.

The model has been tested on different convergence

tolerance values. The results show that the best value

of the convergence tolerance is 0.015.

DFNN(100,15,15,2)

Convergence tolerance

Accuracy

Figure 12: Accuracy vs convergence tolerance.

Figure 13 shows three different convergence val-

ues along with different number of neurons in the hid-

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

344

Number of neurons

Accuracy

tol = 0.01

tol = 0.015

tol = 0.001

Figure 13: Accuracy vs number of neurons.

den layers in range [1 : 20]. The results show that the

best convergence value is 0.015 along with different

number of neurons. However, decreasing the number

of neurons in both hidden layers below four neurons

or increasing the number of neurons above 18 neurons

deteriorates the accuracy of the model as shown the

red line in Figure 13. Thus, when convergence toler-

ance is set to 0.015 and the number of neurons of both

hidden layers are selected (5, 10, or 15) neurons, the

ATSDS achieves better accuracy. The results show

that when the convergence tolerance is increased to

(0.1 or 0.15), the accuracy of the system is deterio-

rated as shown in Table 1.

Table 1: The accuracy and Convergence tolerance.

DFNN Conv. tolerance Accuracy

(100,15,15,2) 0.001 0.9356235

(100,15,15,2) 0.015 0.9442099

(100,15,15,2) 0.01 0.9429968

(100,15,15,2) 0.1 0.7482279

(100,15,15,2) 0.15 0.7482279

Table 1 shows that the best convergence tolerance

values are (0.001, 0.015, and 0.01). Table 2 shows the

accuracy of our proposed model along with chang-

ing those best convergence tolerance values and the

number of neurons in both hidden layers. The out-

put systems have high accuracy between 93.31% and

94.42% regardless of the number of neurons in the

hidden layers.

6 CONCLUSION

The paper introduces a high accuracy Auto Twitter

Suicide Detector (ATSD) system to auto detect users

who are in danger with such kind of destructive sui-

cide thoughts. ATSD is built on multi-purpose cluster

computing system (Spark) using deep feed forward

neural network and L-BFGS. The ATSD system is

analyzed to choose the optimal parameters in terms of

Table 2: The accuracy and Convergence tolerance.

DFNN Conv. tolerance Accuracy

(100,1,1,2) 0.001 0.940160

(100,1,1,2) 0.015 0.935562

(100,1,1,2) 0.01 0.940536

(100,5,5,2) 0.001 0.939008

(100,5,5,2) 0.015 0.941045

(100,5,5,2) 0.01 0.937366

(100,10,10,2) 0.001 0.939207

(100,10,10,2) 0.015 0.940205

(100,10,10,2) 0.01 0.938012

(100,15,15,2) 0.001 0.935623

(100,15,15,2) 0.015 0.944209

(100,15,15,2) 0.01 0.942996

(100,20,20,2) 0.001 0.936995

(100,20,20,2) 0.015 0.935451

(100,20,20,2) 0.01 0.933187

the number of neurons and convergence tolerance val-

ues that increase the accuracy of the system. The pro-

posed system is evaluated using Twitter dataset and

achieved high accuracy (94.42%). We anticipate that

our auto detector system can produce reliable results

for Twitter posts, allowing authorities to mitigate the

risk of suicide thoughts the threat our society. As

future work, we intend to run more experiments on

different datasets using different classiﬁer to compare

their accuracy.

ACKNOWLEDGEMENT

We thank Dr Nathaniel Osgood, University of

Saskatchewan, Canada for his support with the Twit-

ter dataset.

REFERENCES

Abboute, A., Boudjeriou, Y., Entringer, G., Az

e, J.,

Bringay, S., and Poncelet, P. (2014). Mining twit-

ter for suicide prevention. In International Confer-

ence on Applications of Natural Language to Data

Bases/Information Systems, pages 250–253. Springer.

Becker, K. and Schmidt, M. H. (2005). When kids seek help

on-line: Internet chat rooms and suicide. Reclaiming

Children and Youth, 13(4):229.

Sentiment Analysis of Serious Suicide References in Twitter Social Network

345

Brown, L. M., Bongar, B., and Cleary, K. M. (2004). A

proﬁle of psychologists’ views of critical risk factors

for completed suicide in older adults. Professional

Psychology: Research and Practice, 35(1):90.

Buchanan, D., Tourigny-Rivard, M., Cappeliez, P., Frank,

C., Janikowski, P., Spanjevic, L., Malach, F., Mokry,

J., Flint, A., and Herrmann, N. (2006). National

guidelines for seniors’ mental health: the assessment

and treatment of depression. Canadian Journal of

Geriatrics, 9(supplement 2):S52–S58.

Carlson, N. (2011). The real history of twitter. business

insider. Featured Articles From The Business Insider,

13.

Conwell, Y., Pearson, J., and DeRenzo, E. G. (1996). Indi-

rect self-destructive behavior among elderly patients

in nursing homes: a research agenda. The American

Journal of Geriatric Psychiatry, 4(2):152–163.

Glass Jr, J. C. and Reed, S. E. (1993). To live or die: A

look at elderly suicide. Educational Gerontology: An

International Quarterly, 19(8):767–778.

Gunn, J. F. and Lester, D. (2015). Twitter postings and sui-

cide: An analysis of the postings of a fatal suicide in

the 24 hours prior to death. Suicidologi, 17(3).

Hwang, J.-N. and Hu, Y. H. (2001). Handbook of neural

network signal processing. CRC press.

Ikunaga, A., Nath, S. R., and Skinner, K. A. (2013). In-

ternet suicide in japan: A qualitative content analysis

of a suicide bulletin board. Transcultural psychiatry,

50(2):280–302.

Jashinsky, J., Burton, S. H., Hanson, C. L., West, J., Giraud-

Carrier, C., Barnes, M. D., and Argyle, T. (2014).

Tracking suicide risk factors through twitter in the us.

Crisis: The Journal of Crisis Intervention and Suicide

Prevention, 35(1):51.

Kaplan, A. M. and Haenlein, M. (2010). Users of the world,

unite! the challenges and opportunities of social me-

dia. Business horizons, 53(1):59–68.

Karau, H., Konwinski, A., Wendell, P., and Zaharia, M.

(2015). Learning spark: lightning-fast big data anal-

ysis. ” O’Reilly Media, Inc.”.

Matykiewicz, P., Duch, W., and Pestian, J. (2009). Cluster-

ing semantic spaces of suicide notes and newsgroups

articles. In Proceedings of the Workshop on Current

Trends in Biomedical Natural Language Processing,

pages 179–184. Association for Computational Lin-

guistics.

McCulloch, W. S. and Pitts, W. (1943). A logical calculus

of the ideas immanent in nervous activity. The bulletin

of mathematical biophysics, 5(4):115–133.

Navaneelan, T. (2012). Suicide rates: An overview. Statis-

tics Canada Ottawa, Canada.

Pang, B., Lee, L., et al. (2008). Opinion mining and senti-

ment analysis. Foundations and Trends

 in Informa-

tion Retrieval, 2(1–2):1–135.

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):2001.

Poulin, C., Shiner, B., Thompson, P., Vepstas, L., Young-

Xu, Y., Goertzel, B., Watts, B., Flashman, L., and

McAllister, T. (2014). Predicting the risk of suicide

by analyzing the text of clinical notes. PloS one,

9(1):e85733.

Rosenblatt, F. (1958). The perceptron: a probabilistic model

for information storage and organization in the brain.

Psychological review, 65(6):386.

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.

Won, H.-H., Myung, W., Song, G.-Y., Lee, W.-H., Kim, J.-

W., Carroll, B. J., and Kim, D. K. (2013). Predicting

national suicide numbers with social media data. PloS

one, 8(4):e61809.

Yip, P. S., Chi, I., Chiu, H., Chi Wai, K., Conwell, Y.,

and Caine, E. (2003). A prevalence study of suicide

ideation among older adults in hong kong sar. Inter-

national journal of geriatric psychiatry, 18(11):1056–

1062.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

346