Predicting Temperament using Keirsey’s Model for

Portuguese Twitter Data

Cristina Fátima Claro, Ana Carolina E. S. Lima and Leandro N. de Castro

Institute of Problem Solving, Presbiteriana Mackenzie University, Rua da Consolação-930, São Paulo, Brazil

Keywords: Machine Learning, Social Media, Temperament’s Classification, Keirsey Temperament Model.

Abstract: Temperament is a set of innate tendencies of the mind related with the processes of perceiving, analyzing and

decision making. The purpose of this paper is to predict the user's temperament based on Portuguese tweets

and following Keirsey's model, which classifies the temperament into artisan, guardian, idealist and rational.

The proposed methodology uses a Portuguese version of LIWC, which is a dictionary of words, to analyze

the context of words, and supervised learning using the KNN, SVM and Random Forest algorithms for train-

ing the classifiers. The resultant average accuracy obtained was 88.37% for the artisan temperament, 86.92%

for the guardian, 55.61% for the idealist, and 69.09% for the rational. By using binary classifiers the average

accuracy was 90.93% for the artisan temperament, 88.98% for the guardian, 51.98% for the idealist and

71.42% for the Rational.

1 INTRODUCTION

A set of characteristics is defined according with the

personality and these describe the individual behav-

ior, the temperament and the emotion (Nor Rahayu et

al., 2016). Personality represents the mixture of char-

acteristics and qualities that builds the character of an

individual. Thus, personality prediction is of interest

in the areas of health, psychology, human resources

and also has many commercial applications. Several

researches investigate the link between human behav-

ior in social media, personality types and psycholog-

ical illnesses, such as depression and post-traumatic

stress (Plank and Dirk, 2015; Lima and de Castro,

2016).

Social media are composed of different types of

social sites, including traditional media, such as

newspaper, radio and television, as well as non-

traditional media, such as Facebook, Twitter and

others (Gundecha and Liu, 2012). Social media

mining is the process that allows the analysis and

extraction of patterns from social media data (Nor

Rahayu et al., 2016). In this context, this paper

develops a system to predict the temperament of

Twitter users, using tweets in the Portuguese

language. The temperament model used was

introduced by David Keirsey, and divides the

temperament into four categories: artisan; guardian;

idealist; and rational. In order to do so, we will use

the TECLA framework adapted to work with

Portuguese texts (Lima and de Castro, 2016). In

addition, it will be shown an analysis of the context

of words by temperament using the dictionary of

words Linguistic Inquiry and Word Count (LIWC)

(Pennebaker et al., 2015).

This paper is structured as follows. Section II

presents the David Keirsey temperament model used

in TECLA, and Section III describes the TECLA

framework. Section IV presents the methodology and

the results achieved and, finally, Section V concludes

and discusses future perspectives.

2 KEIRSEY’S TEMPERAMENT

MODEL

Temperament is a set of innate tendencies of the mind

that relates to the processes of perceiving, analyzing,

and decision making (Calegari and Gemignani,

2006). People seek success, happiness, love, pleasure,

etc., in different ways and with distinct intensities

and, therefore, there are different types of tempera-

ment (Hall et al., 2000).

The temperament has its history marked in the

proposal of the four humors described by Hippocra-

tes, which gave origin to the theory of the four humors

to interpret the state of health and illness of a person

(Hall et al., 2000). From this theory, Galen (190 AD)

250

Claro, C., Lima, A. and Castro, L.

Predicting Temperament using Keirsey’s Model for Portuguese Twitter Data.

DOI: 10.5220/0006700102500256

In Proceedings of the 10th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2018) - Volume 2, pages 250-256

ISBN: 978-989-758-275-2

created the model of the first temperament typology

(Ito and Guzzo, 2002).

David Keirsey, an American psychologist, di-

rected his studies to temperament in action, paying at-

tention to choices, behavior patterns, congruencies,

and consistencies. For Keirsey, psychological types

are driven by aspirations and interests, which moti-

vate us to live, act, move, and play a role in society

(Lima and de Castro, 2016; Keirsey, 1998; Calegari

and Gemignani, 2006).

The artisans are usually impulsive, they speak

what comes to their minds and tend to do what works;

whereas the guardians speak mainly of their duties

and responsibilities, and how well they obey the laws.

Idealists normally act from a good conscience and the

rationals are pragmatic, act efficiently to reach their

objectives, sometimes ignoring the rules and conven-

tions if necessary (Keirsey, 1996; Lima, 2016).

Keirsey's temperament can be obtained by map-

ping the result of the MBTI test (Myers-Briggs Type

Indicator), which uses four dimensions to classify us-

ers, totaling 16 psychological types (Keirsey, 1998;

Calegari and Gemignani, 2006; Plank and Dirk,

2015). Psychological types are acronyms formed by

letters that begin with E and I, (extraversion and in-

troversion), which are attitudes; S and N indicate sen-

sation and intuition, which is the process of percep-

tion; the letters T and F indicate thinking and feeling,

and usually use logical reasoning, think first and feel

later; and letters J and P indicate judgment and per-

ception, which are attitudes and reflect the individu-

als’ style in the external world (Hall et al., 2000).

The mapping of the MBTI into the Keirsey’s

model occurs by means of the classification of the ac-

ronyms defined by Myers-Briggs, as shown in Table

1 (Keirsey, 1998).

Table 1: Keirsey temperament model classification from

the MBTI.

Keirsey

Myers-Briggs

Artisan

ESTP

ISTP

ESFP

ISFP

Guardian

ESTJ

ISTJ

ESFJ

ISFJ

Idealist

ENFJ

INFJ

ENFP

INFP

Rational

ENTJ

INTJ

ENTP

INTP

3 THE TECLA FRAMEWORK

The TECLA framework (Temperament Classifica-

tion Framework) was developed by Lima & de Castro

(Lima, 2016; Lima and de Castro, 2016) with the ob-

jective of offering a modular tool for the classification

of temperaments based on the Keirsey and Myers-

Briggs models (Lima, 2016). It is structured in a mod-

ular form, giving greater independence for each stage

of the process and making it possible to couple and

test different techniques in each module (Lima,

2016). Figure 1 shows the TECLA’s modules, which

are detailed in the following.

Figure 1: The TECLA framework structure.

• Data Acquisition Module: Receives information

from the user to be classified, including the num-

ber of tweets, the number of followers and fol-

lowed, and a set of messages (tweets) from the

user;

• Message Pre-Preprocessing Module: Processes

the data by creating an object matrix (meta-base)

represented by meta-attributes. The information

in the TECLA are divided into two categories:

grammatical and behavioral. The behavior cate-

gory uses information from Twitter, such as num-

ber of tweets, number of followed, followers, fa-

vorites, and number of times the user has been fa-

vorited. The grammar category uses information

from LIWC, MRC, Taggers, or oNLP (Lima and

de Castro, 2016);

• Temperament Classification Module: Responsi-

ble for identifying the temperament of social me-

dia users. It performs the classification in the

Keirsey model by using a set of classifiers;

• Evaluation Module: Used to quantify the frame-

work performance (Lima and de Castro, 2016).

In the version proposed in this paper, the TECLA will

be adapted to work with texts written in Portuguese

and will use the information provided by the LIWC

(Pennebaker et al., 2015).

4 METODOLOGY AND RESULTS

The description to be presented in this section will

follow the modular structure of the TECLA frame-

work. First we will explain how we implemented each

Predicting Temperament using Keirsey’s Model for Portuguese Twitter Data

251

module of the framework and then the computational

results.

A. DATA ACQUISITION

To validate this work we used a data from the litera-

ture called Twisty, which has tweets in Portuguese

and is provided by CLiPS (The Computational Lin-

guists & Psycholinguistics Research Center)

(Verhoeven et al., 2016). The dataset is composed of:

user id; tweet id; other tweets id; confirmed tweets id;

Myers-Briggs Type Indicators (MBTI) result; and

gender. The tweets were captured by using the Twit-

ter API (Xavier and Carvalho, 2011), and we captured

the tweets, number of followers, number of favorites,

total number of tweets and total number of friends

(Kwak et al., 2010) of each user.

The original database consists of 4,090 user ids.

From this universe it was not possible to collect 222

user ids due to denied access, leaving 3,868 valid user

ids. Table 2 shows the descriptive analysis of the da-

tabase according to David Keirsey's model, where

Tweets_Statuses_Count refers to the number of

tweets from the opening of the user account, and

Tweets_base refers to the number of tweets collected.

Figure 2 shows the temperament distribution of

the users. It is noted that the idealist temperament is

the predominant one, totaling 44% of the database.

Figure 2: Distribution of the users by temperament.

B. PRE-PROCESSING AND CATEGORY ANALYSIS

At this stage the texts are prepared for the application

of the classification algorithms, which consists of the

removal of special characters, blank spaces, numbers,

symbols, URLs, tokenization, and stopwords removal

(Haddi et al., 2013; Spencer and Uchyigit, 2012). Af-

ter that, a bag-of-words technique was applied to

specify the importance of each attribute (token) by as-

signing a weight to each token based on its TF-IDF

(Feldman and Sanger, 2007).

Another way to structure documents is through

the use of dictionaries, such as the Linguistic Inquiry

and Word Count (LIWC), which allows the grouping

of words into psychologically meaningful categories.

The LIWC was created by Dr. James Pennebaker to

examine relationships between language and person-

ality (Komisin and Guinn, 2012; Pennebaker and

King, 1999). It is a textual analysis tool that structures

documents into categories by assigning each word to

the corresponding category (Pennebaker et al., 2015).

By using the Portuguese LIWC dictionary it was

calculated the frequency of words by temperament.

The goal here is to present the most spoken word cat-

egories for each temperament, as shown in Table 3.

In a first analysis it is observed that the rational tem-

perament usually has a higher average frequency of

categories, followed by the guardian temperament.

In a second analysis it is noticed that all the tem-

peraments have a higher frequency in the following

categories: funct, which are functional words like for,

not, very and others; pronoun, which are the pronouns

like I, mine, me and others; verb, which are verbs

such as cover, occur and other; social, which are so-

cial processes such as talking, accompanying and

other; cogmech, which is the cognitive category; and

relativ, which is relativity as are (Pennebaker et al.,

2015) a, turn, exit and other. There are users who tend

to write by hiding their identity and tend to present a

writing that expresses action and can show a greater

perception and logical reasoning.

In another analysis, it is noted prominence for the

ppron category for rational and idealist tempera-

ments; and ipron, present, preps, incl for guardian and

rational temperaments. These categories are related to

the linguistic dimensions that tend to write more pa-

Table 2: Temperament and Twitter data of the users (A = artisan, G = guardian, I = idealist, R = rational).

Total

Users

450

506

1.717

1.195

3.868

Tweets_Statuses_Count

12.343.807

15.648.860

65.593.286

45.198.150

138.784.103

Tweets_Base

674.211

738.755

2.570.646

1.751.624

5.735.236

Followers

292.413

423.549

1.497.093

1.799.686

4.012.741

Friends

168.893

225.371

825.969

640.529

1.860.762

Favorites

1.768.903

2.371.924

10.006.749

6.683.984

20.831.560

Artisan

450

12%

Guardian

506

13%

Idealist

1717

44%

Rational

1195

31%

Distribution of Temperament Types

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

252

pers, pronouns, auxiliary verbs, etc. The humans cat-

egory has higher frequency for the idealist and ra-

tional temperaments, and tends to write about people.

The affect category occurs more for the guardian and

idealist temperaments, and users tend to write pre-

senting positive, negative, and other emotions. Last,

the ingest category has higher average frequency for

the artisan and guardian temperaments, who tend to

write about body, health and others.

With a lower frequency of words and with the

same value in all the temperaments the future, family,

anx, see, health, death, assent and filler categories

stand out.

Table 3: Average frequency of LIWC categories by temper-

ament.

Category

Artisan

Guardian

Idealist

Rational

funct

5,22

5,31

5,30

5,39

pronoun

1,85

1,86

1,91

1,93

ppron

1,21

1,22

1,25

1,26

0,39

0,36

0,42

0,41

0,04

0,03

you

0,66

0,70

0,66

0,68

shehe

0,64

0,67

0,65

0,67

they

0,12

0,13

ipron

1,22

1,25

1,24

1,27

article

0,77

0,80

0,78

0,81

verb

1,87

1,88

1,91

auxverb

0,67

0,69

0,68

0,70

past

0,42

0,43

present

1,12

1,13

1,12

1,15

future

0,08

adverb

0,48

0,50

preps

1,48

1,52

1,46

1,49

conj

0,91

0,90

0,92

0,93

negate

0,24

0,25

quant

0,59

0,60

0,61

0,62

number

0,14

0,15

swear

0,71

0,72

0,73

social

2,16

2,17

2,21

2,24

family

0,04

friend

0,11

0,10

0,09

humans

1,11

1,15

1,14

Affect

1,08

1,10

1,09

1,08

Posemo

0,70

0,72

0,71

0,69

negemo

0,35

0,34

0,36

anx

0,05

anger

0,14

0,13

0,14

0,15

sad

0,17

0,16

0,17

cogmech

4,06

4,14

4,12

4,18

insight

0,72

0,73

0,74

0,75

cause

0,48

0,50

0,49

0,50

discrep

0,68

0,69

0,70

tentat

0,98

0,99

1,00

1,02

certain

0,38

0,39

inhib

0,53

0,55

0,54

0,55

incl

1,40

1,42

1,40

1,41

excl

0,79

0,80

0,82

percept

0,78

0,79

0,80

0,79

see

0,26

hear

0,18

0,17

0,18

feel

0,31

0,32

0,31

bio

0,71

0,69

0,71

0,72

body

0,32

0,31

0,32

health

0,13

sexual

0,21

0,20

0,21

ingest

1,08

1,10

1,06

relativ

2,30

2,36

2,28

2,31

motion

0,76

0,74

0,76

space

0,98

1,01

0,97

0,99

time

0,98

1,02

0,96

work

0,23

0,25

0,23

0,24

achieve

0,46

0,49

0,46

0,47

leisure

0,31

0,32

0,31

home

0,06

0,05

money

0,29

0,30

0,29

0,31

relig

0,08

0,09

0,08

death

0,06

assent

0,13

nonfl

0,26

0,27

0,26

0,27

filler

0,03

C. CLASSIFICATION

In the experiments carried out, 4% of the total dataset

was randomly sampled to be used due to the size of

the matrix to be processed and the unavailability of

Predicting Temperament using Keirsey’s Model for Portuguese Twitter Data

253

computational resources. To perform the tempera-

ment classification, the following classifiers available

in the Scikit-learn (Pedregosa et al., 2011) were used:

KNN; SVM; and Random Forest (Lima and de

Castro, 2016; Nor Rahayu et al., 2016). Each temper-

ament was divided into a binary problem, as proposed

by Lima and de Castro (Lima and de Castro, 2016).

For the tests, a cross-validation with 6 and 10-folders

was used, and the accuracy, precision, recall, and F-

measure were calculated. For the KNN classifier,

which uses the object classification according to the

K-nearest neighbors, K = 1, K = 2 e K = 3 and the

cosine similarity was used for determining the neigh-

bors. The tests were separated into LIWC word dic-

tionary LIWC (Pennebaker et al., 2015) and TF-IDF

(Feldman and Sanger, 2007).

1) LIWC

Table 4 shows the results achieved by the TECLA for

a validation with 6 folders executed 10 times. The

values in bold are the best average accuracy and F-

measure results obtained by the classifiers for each

temperature.

For the artisan temperament, the KNN algorithm

with K = 1 obtained an average accuracy of 80.44%

and an F-measure of 88.91%. With a better perfor-

mance, that is, a greater number of correctly labeled

objects, it was the SVM algorithm with an average

accuracy of 88.37%, followed by the Random Forest

with an average accuracy of 87.95%. The SVM pre-

sented a higher average accuracy and a 100% recall,

whilst the Random Forest presented a better precision

than SVM.

In relation to the guardian temperament, the most

assertive prediction was by the SVM algorithm, with

an average accuracy of 86.92% and F-measure of

93%, followed by the Random Forest with an average

accuracy of 86.32%. The lowest average accuracy

(78.36%) was for the KNN with K = 1.

The SVM also performed better for the idealist

and rational temperaments. The idealist temperament

had an average accuracy of 55.61% and F-measure of

71.46% and the rational temperament had an average

accuracy of 69.09% and F-measure of 81.71%. It is

concluded that these two temperaments had a good

performance in the labeling of objects due to the av-

erage accuracy being greater than 50%.

In general, the SVM obtained better accuracy for

all temperaments, but for the artisan and guardian

temperaments the Random Forests presented very

close average accuracies to the SVM.

2) TF-IDF

Table 5 shows the results achieved by the TECLA

framework for a cross-validation with 10 folders, ex-

ecuted 10 times. The values in bold are the best aver-

age accuracy and F-measure results obtained by the

binary classifiers for each temperature.

The artisan temperament obtained a high average ac-

curacy (90.93%) with the KNN, K = 3 and F-measure

of 95,09% which makes effective the result obtained

by the KNN effective, K = 3, followed by the SVM

with an average accuracy of 88.35%. The SVM had

the highest average accuracy for the guardian temper-

ament, but in this case the KNN for K = 3 had a very

low performance. One hypothesis for this low value

is the imbalance of the database, so when increasing

the number of neighbors the algorithm can not label

the object. For the idealist temperament the SVM and

KNN (K = 3) were practi cally even. For the rational

temper-ament the best performance was for the KNN

algorithm with K = 2, with an average accuracy of

Table 4: Accuracy (Acc), Precision (Pre), Recall (Rec) and F-measure (M-F) for the four temperaments using 6 folders and

10 iterations.

LIWC

1NN

2NN

3NN

Random Forest

SVM

Artisan

Acc

80,44% ± 0,71%

87,62% ± 0,37%

87,95% ± 0,16%

88,37% ± 0,00%

Pre

88,79% ± 0,14%

88,47% ± 0,07%

88,41% ± 0,06%

88,37% ± 0,00%

Rec

89,10% ± 0,87%

98,86% ± 0,44%

99,39% ± 0,15%

100,00% ± 0,00%

M-F

88,91% ± 0,47%

93,37% ± 0,23%

93,58% ± 0,09%

93,82% ± 0,00%

Guardian

Acc

78,36 ± 0,62%

85,67% ± 0,10%

85,74% ± 0,10%

86,32% ± 0,11%

86,92% ± 0,01%

Pre

87,05% ± 0,07%

86,94% ± 0,04%

87,03% ± 0,06%

86,92% ± 0,00%

Rec

88,22% ± 0,76%

98,27% ± 0,09%

98,36% ± 0,09%

99,02% ± 0,11%

100,00% ± 0,01%

M-F

87,61% ± 0,43%

92,25% ± 0,06%

92,30% ± 0,06%

92,63% ± 0,06%

93,00% ± 0,01%

Idealist

Acc

54,97% ± 0,46%

52,57% ± 0,61%

54,27% ± 0,40%

55,61% ± 0,01%

Pre

56,80% ± 0,27%

57,88% ± 0,57%

56,67% ± 0,26%

55,61% ± 0,01%

Rec

79,44% ± 0,80%

54,18% ± 0,97%

75,65% ± 1,04%

100,00% ± 0,00%

M-F

66,19% ± 0,33%

55,86% ± 0,67%

64,76% ± 0,49%

71,46% ± 0,02%

Rational

Acc

59,12% ± 0,58%

65,82% ± 0,49%

87,62% ± 0,37%

66,62% ± 0,26%

69,09% ± 0,03%

Pre

69,72% ± 0,21%

69,27% ± 0,13%

88,47% ± 0,07%

69,74% ± 0,14%

69,10% ± 0,01%

Rec

72,17% ± 1,20%

90,84% ± 1,16%

98,86% ± 0,44%

91,38% ± 0,40%

99,97% ± 0,04%

M-F

70,85% ± 0,65%

78,57% ± 0,47%

74,78% ± 0,27%

79,09% ± 0,19%

81,71% ± 0,03%

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

254

Table 5: Accuracy (Acc), Precision (Pre), Recall (Rec) and F-measure (M-F) for the four temperaments using 10 folders and

10 iterations.

TF-IDF

1NN

2NN

3NN

Random Forest

SVM

Artisan

Acc

87,31% ±3,01%

90,25% ± 0,09%

90,93% ± 0,06%

87,69% ± 0,07%

88,35% ± 0,07%

Pre

87,31% ± 3,01%

90,25% ± 0,09%

90,93% ± 0,06%

87,69% ± 0,07%

88,35% ± 0,07%

Rec

99,00% ± 3,00%

100,00% ± 0,00%

100,00% ±0,00%

100,00% ± 0,00%

M-F

92,56% ± 2,99%

94,73% ± 0,07%

95,09% ±0,06%

93,25% ± 0,08%

93,60% ± 0,07%

Guardian

Acc

88,98% ± 0,06%

85,08% ± 0,09%

27,13% ± 3,73%

88,25% ± 0,07%

90,93% ± 0,08%

Pre

88,98% ± 0,06%

85,08% ± 0,09%

16,27% ± 3,25%

88,25% ± 0,07%

90,93% ± 0,08%

Rec

100,00% ± 0,00%

19,00% ± 3,00%

100,00% ± 0,00%

M-F

94,04% ± 0,05%

91,68% ± 0,11%

17,50% ± 3,16%

93,55% ± 0,13%

95,15% ± 0,06%

Idealist

Acc

52,40% ± 2,14%

51,98% ± 0,10%

61,05% ± 0,08%

54,42% ± 3,29%

61,04% ± 0,12%

Pre

48,14% ± 1,05%

51,98% ± 0,10%

61,05% ± 0,08%

50,12% ± 5,65%

61,04% ± 0,12%

Rec

90,00% ± 0,00%

100,00% ± 0,00%

88,00% ± 8,72%

100,00% ± 0,00%

M-F

62,12% ± 0,99%

67,74% ± 0,20%

75,13% ± 0,27%

63,32% ± 6,78%

75,08% ± 0,34%

Rational

Acc

65,52% ± 1,43%

71,42% ± 0,10%

70,80% ± 0,11%

65,60% ± 1,82%

68,21% ± 0,06%

Pre

62,18% ± 0,72%

71,42% ±0,10%

70,80% ± 0,11%

65,40% ± 2,41%

68,21% ± 0,06%

Rec

90,00% ± 0,00%

100,00% ±0,00%

100,00% ± 0,00%

99,00% ± 3,00%

100,00% ± 0,00%

M-F

73,16% ± 0,50%

82,85% ±0,27%

82,51% ± 0,27%

78,21% ± 2,63%

80,70% ± 0,013%

82.85%. Again the SVM was the algorithm that pre-

sented the best average performance among the eval-

uated ones.

5 CONCLUSIONS AND FUTURE

WORK

Temperament influences the way we perceive and re-

act to the world. Understanding temperament is of

crucial importance to our lives and to position our-

selves properly in the market. Normally, one’s tem-

perament can be known by filling in tests, such as the

MBTI (Myers-Briggs Type Indicator). The hypothe-

sis of this research is that it is possible to identify the

temperament of a person in a passive way, only by

using data obtained from the social media of the per-

son. For this, a database of tweets containing the

MBTI result of Twitter users was employed. These

data were used to generate predictive models of tem-

perament.

The documents (Tweets) were structured with the

Portuguese dictionary LIWC that groups words into

categories. The calculation of the frequency of words

was carried out to show which category is most talked

about by artisan, guardian, idealistic and rational tem-

peraments. In this analysis it is possible to identify the

writing tendency of the user associated with the sub-

ject that is most identified, perception among oth-

ers.The tweets were structured using LIWC and TF-

IDF. For classification via LIWC the best accuracy

results were achieved for the artisan and guardian

temperaments trained with SVM. For the TF-IDF the

highest average accuracy was for the artisan, guardian

and idealist temperaments, also with emphasis on the

SVM algorithm. For the representation using the TF-

IDF the best average accuracy was observed for the

artisan and guardian temperaments for the KNN (K =

3) and SVM algorithms.

As a future work, we intend to carry out a case

study using the TECLA framework with a database

composed of a set of volunteer users who will answer

the MBTI test form and share their social profiles so

that we can use their data to train the TECLA frame-

work and classify temperament.

Another improvement to be made is the study of

the content of the documents to investigate why the

classifiers have low accuracy and how much the un-

balanced basis interferes in this result to verifying

whether there is need for treatment for the unbalanced

classes.

ACKNOWLEDGEMENTS

The authors thank Mackenzie University, Mack-

pesquisa, CNPq, Capes and FAPESP for the financial

support.

Predicting Temperament using Keirsey’s Model for Portuguese Twitter Data

255

REFERENCES

Calegari, M. d., & Gemignani, O. H. (2006). Temperamento

e Carreira. São Paulo: Summus Editorial.

Feldman, R., & Sanger, J. (2007). The Text Mining

Handbook: Advanced approaches in Analyzing

Unstructured Data. Cambridge university press.

Gundecha, P., & Liu, H. (2012). Mining social media: a

brief introduction. New Directions in Informatics,

Optimization, Logistics, and Production. Informs, pp.

1-17.

Haddi, E., Liu, X., & Shi, Y. (2013). The role of text pre-

processing in sentiment analysis. Procedia Computer

Science, 17, 26-32.

Hall, C. S., Lindzey, G., & Campbell, J. B. (2000). Teorias

da Personalidade. Porto Alegre: Artmed.

Ito, P. d., & Guzzo, R. S. (2002). Diferenças individuais:

temperamento e personalidade; importância da teoria.

Estudos de Psicologia, pp. 91-100.

Keirsey, D. (1998). Please Undestand Me II: Temperament,

Character, Intelligence. Prometheus Nemesis Book

Company.

Keirsey, D. M. (1996). Keirsey.com. (Corporate Offices)

Acesso em 12/10/2017 de 10 de 2017, disponível em

://www.keirsey.com/4temps/overview_temperaments.

asp

Komisin, M. C., & Guinn, C. I. (2012). Identifying

personality types using document classification

methods. In: FLAIRS Conference.

Kwak, H., Lee, C., Park, H., & Moon, S. (2010). What is

Twitter, a social network or a news media? Proceedings

of the 19th international conference on World wide

web. ACM., 591-600.

Lima, A. C. (2016). Mineração de Mídias Sociais como

Ferramenta para a Análise da Tríade da Persona

Virtual. São Paulo.

Lima, A. C., & de Castro, L. N. (2016). Predicting

Temperament from Twitter Data. Advanced Applied

Informatics (IIAI-AAI), 2016 5th IIAI International

Congress on. IEEE.

Nor Rahayu, N., Zainol, Z., & Yoong, T. L. (2016). A

comparative study of different classifiers for automatic

personality prediction. Control System, Computing and

Engineering (ICCSCE), 2016 6th IEEE International

Conference on. IEE, 435-440.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., . . . Perro, M. (2011). Scikit-

learn: Machine learning in Python. Journal of Machine

Learning Research 12.Oct, 12, 2825-2830.

Pennebaker, J. W., & King, L. A. (1999). Linguistic styles:

language use as an individual difference. Journal of

personality and social psychology, 77(6), 1296.

Pennebaker, J. W., Boyd, R. L., Jordan, K., & Blackburn,

K. (2015). The development and psychometric

properties of LIWC2015.

Plank, B., & Dirk, H. (2015). Personality Traits on Twitter-

or-How to Get 1, 500 Personality Tests in a Week.

Proceedigs of the 6th Workshop on Computational

Approaches to Subjectivity, Sentiment and Social

Media Analysis(WASSA 2015), pp. 92-98.

Spencer, J., & Uchyigit, G. (2012). Sentimentor: Sentiment

analysis of twitter data. Proceedings of European

conference on machine learning and principles and

practice of knowledge discovery in databases.

Verhoeven, B., Daelemans, W., & Plank, B. (2016).

Twisty: A Multilingual Twitter Stylometry Corpus for

Gender and Personality Profiling. Proceedings of the

10th International Conference on Language Resources

and Evaluation.

Xavier, O. C., & Carvalho, C. L. (2011). Desenvolvimento

de Aplicações Sociais A Partir de APIs em Redes

Sociais Online. UFG. Goiânia.

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

256