Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

Lingual markers for Automating personality profiling - Background

and road ahead

First Author:
Mohmad Azhar
Department of computer science,
University of Kashmir Hazratbal,
Srinagar 190006, India
Residence: Hazratbal, Srinagar, 190006
Tel.: +91-9149539973
E-mail: mohmadazhar.student@kashmiruniversity.net

Second Author:
Dr. Manzoor Ahmad Chachoo
Department of Computer Science
University of Kashmir Srinagar, India – 190006
Email: manzoor@kashmiruniversity.ac.in
Google scholar url: https://scholar.google.com/citations?user=iK9NpkcAAAAJ&hl=en&oi=ao

1
Lingual markers for Automating personality profiling - Background and road
ahead
Mohmad Azhar , Dr Manzoor Ahmad Chachoo
1 2

1 Department of computer science


University of Kashmir, Hazratbal Srinagar, Srinagar 190006, India

Abstract

Personality is a psychological concept which embodies the unique characteristics of an individual. An individual's
distinct traits are embodied by the psychological concept of personality. The Lexical Hypothesis states that language
use and the terms people use to describe one another can help us decide personality qualities. Huge improvements in
data collecting and processing have been brought about by technological breakthroughs. These could help to develop
autonomous personality assessment models by deriving linguistic markers from the data present in social media,
telecommunication signals, and even signals collected from human-machine Interaction. Numerous studies have
cantered on using machine learning to automate personality recognition from text. However, there are questions in
terms of their performance, reliability as well as ethical usage. To find solutions, we extensively review and analyse
the existing research in the field of personality computing using lingual markers in text. A content-oriented
classification of the techniques used is provided. We also examine the existing literature for gaps and limitations with
a detailed comparative analysis. The field of personality computing has the potential to impact every field of human
life but the progress as of now is limited. Our review will help researchers to build from what has been achieved so
far for faster progress in the field.
Keywords: Social-signal processing, Human-Computer Interaction, Personality computation, Behaviour modelling, Emotional
Intelligence, Text Processing, Natural language Processing, Personality Profiling

1. Introduction

The dynamics of human nature and its uniqueness has been a subject of study at least since the Greek era.
Theophrastus states, “Often before now have I applied my thoughts to the puzzling question — one, probably, which
will puzzle me forever — why it is that, while all Greece lies under the same sky and all the Greeks are educated alike,
it has befallen us to have characters so variously constituted.” [1]. Personality is the combination of behaviour,
cognition, and emotions of an individual arising because of various biological and sociological processes. It has a
profound impact on governing our life choices, preferences, and desires. Personality assessment applies to our every
decision of life, and thinking ahead may supply important insights into the future behaviours of a person. Psychologists
have always tried to find out what is the basis of personality, which makes it unique to every individual. Self-
personality tests are the only way however to uniquely identify the personality of a human being. The Woodworth
Psychoneurotic Inventory [2] is commonly cited as the first personality test. It was developed during World War I for
the U.S. military to screen recruits for shell shock risks (post-traumatic stress disorders). Various other inventories
like 16PF [3], NEO-PI-R [4] and MBTI Indicator [5] have been developed for evaluating personality traits.

Personality computing is the science of analysing the personality traits of a person. It aims to fit the various
emotional, cognitive, and behavioural responses to a computational model. Automating personality analysis will help
us to characterize the key traits of a person, and may determine latent mechanisms which make us different from one
another. It will help us to uniquely identify the personality of a person and provide platforms for human-computer
interaction by enabling synthetic personalities for the machines. There are diverse applications which would receive
help from such developments. In their survey, Mohammedi and Vinceralli [6] summarized the main problems that
need to be tackled in the domain; Automatic personality Recognition (APR), which aims to recognize true personality
innate to the individual, Automatic personality perception (APP), which aims to predict the personality of others
perceived by others, and Automatic personality synthesis (APS), that intends to develop a synthetic personality for
machines for interaction with other machines humans. Our focus is the sub-domain of Automatic Personality
Recognition, which involves Personality recognition from text sources (APRT). With the current availability of
enormous data, futuristic computation mechanisms might reveal patterns that may point us back to what originally
makes us unique. Over the past two decades a lot of researchers have been interested in personality recognition using

2
widely varied data sources and machine learning algorithms. The domain, even though nascent, is bound to have a
massive impact on areas like social signal processing, Human-computer Interaction, recommendation systems,
domotics, job recommendations, counselling, guidance and even suicide prevention. We have thoroughly reviewed
the APRT works over the last two decades including the psycho-linguistic works which form the background for the
domain. The main purpose of this study is to supply a dissection of the available literature to evaluate achievements,
shortcomings, and possibilities for betterment. In section 2 we provide an outline of APRT. A detailed summary of
existing literature works forms sections 3 and 4. Some Open challenges in the domain and possible solutions are
discussed in Section 5. Section 6 concludes the paper envisaging future directions for the field.

2. Personality computing: State of the art

In this section we supply a bird’s eye view of what has been achieved in the last two decades. For the sake of
convenience, we have divided the existing literature base in a five-fold manner, the existing datasets, low level features
used as input, the psychological personality base, computational techniques, and the Performance metrics.

a. Datasets

Digitalization and advent of social media has supplied us enormous amounts of data which can be used for
automatic personality assessment. However, curation and annotation of this data is a concern. Researchers have chosen
diverse sources ranging from formal/informal texts, transcriptions of spoken text, emails, biographies, personal blogs
but most have collected social media data along with either self or observer personality reports. We have listed four
distinct datasets which form a representative of a separate class of text data and have been used by a vast majority of
researchers.
Table 1: Major Text based Datasets used in personality computing literature.

Dataset Description Number of Language Words per Subject Personality Measure/


subjects Affective Dimensions
/ Identities
Essays 2479 extemporal 1203 English 766 Big 5
essays from
psychology
students.
MyPersonality1 Personality data and 250 English 575 Big 5
Facebook profile
information.
PAN-AP-2015 A multi lingual 726 English, Dutch 1258 Big 5
collection of tweets Spanish, Italian
with gold standard
personality labels.
YouTube [10] Transcriptions of 404 English --- Big 5
YouTube video
blogs.

Essays (Formal)

In an attempt of Author profiling from formal texts, Pennebaker and King [7] (Penn99) collected ESSAYS
consisting of 2400 stream of consciousness texts written by 1203 psychology students in University of Pennsylvania.
The essays were labelled with self-reported personality traits using BFI which stays a standard dataset for text-based
personality computation till date.

MyPersonality (social media-Facebook)

As a part of the MyPersonality project (2007-12), David Stivell and Michal Kosinki developed an application
intended for collection of Facebook data for personality computation. It allowed users to fill out a personality
questionnaire and opt to donate their Facebook data for research. A small subset of this data was released by Celli et

3
al [8] having Facebook posts of 250 users annotated with self-reported personality assessments and several Network
based features. The dataset has been used by a lot of researchers for Automatic Personality Recognition from text.

PAN-AM-2015 (social media-Twitter-Multi)

Rangel et al [9] released a multi-lingual twitter dataset, containing tweets of 726 Twitter users with self-
reported personality labels. The dataset has been actively used by researchers focused on building generalized models
that can be applied to other languages.

YouTube (Apparent Personality)

The YouTube Vlogs dataset (originally provided by Idiap research institute, Spain). The dataset was first used
by Biel et al in [10]. It consists of audio-visual recordings of 404 YouTube Vloggers who talk about a variety of topics
ranging from personal, social to political. The data is gender-neutral (52 percent female). The videos and the
transliterations were produced in the English language. Automatic labelling the database was done according to five
traits of the OCEAN/Big 5 model by Amazon Mechanical Turk [10] using Gosling Ten-Item Personality Inventory
[11]

b. Psycho-lingual personality markers

The personality of a person is an inside phenomenon but it affects our behaviour in every way. It is possible to
take personality and affective cues from the overall looks of a person, their linguistic style, handwriting, how they
talk, or even walk to deduce personality information about them. Psychologists and lingual experts have set up the
existence of such personality markers across a wide range of linguistic variables using the Lexical Hypothesis. The
lexical Hypothesis assumes that the most relevant individual differences are encoded into the language, and more
important the difference, the more likely it is to be expressed in a single word. Table 2 illustrates the low-level lingual
features, some of which have been used in the existing personality computing works. The feature extraction process
follows either or both of two basic mechanisms, Top down (Closed Vocabulary) or Bottom Up (Open Vocabulary).

Closed Vocabulary

The Closed vocabulary approach uses external resources with established word categories like LIWC, MRC,
WordNet, and NRC Lexicons to extract low level features for personality trait estimation.

Various Psycho-linguistic style bearing tools and categories have been developed. Pennebaker et al [7]
(Penn99) identified 77 independent word categories associated with personality and behaviour of an individual.
Colhart [35] identified 26 Psycholinguistic categories for 150, 837 English words. Moffitt et al developed Structured
programming for linguistic cue extraction (SPLICE) for extraction of linguistic cues from language. Related categories
were identified by researchers for associating Affect [37], Emotion [39-42], Valence and arousal [41]. Categories
based on measuring the semantic sense between were identified in WordNet [44], concept modelling was achieved by
associating semantic features between words in [45] and EmoSenticNet [47] evaluates the sentiments in language.
Speech and dialogue acts [47-49] were identified in language to provide a categorical base for Affect modelling for
language.

Such word categories provide a transparent linkage between word categories which can be used for automatic
prediction of Personality and behaviour of a person using text.

Open Vocabulary

Open vocabulary is data driven and seeks linguistic cues associated with personality traits. This approach
accesses the choice of words from a user to assign personality traits to them. It can be done on a word, sentence, or
document level. Word level representations like Part of Speech Tags [12], Bag of Words [13] take all the words from
the text without any consideration between how they are linked. Temporal features [11] are used to understand the
time when the person has written them. Named Entities [14] and function words [15] are the established word
categories from grammar rules.

4
Word count and Frequency features like N-Grams show how frequently a person uses a particular group of
words. Term Frequency - Inverse document Frequency (TF-IDF) [16] and Term Frequency - Inverse document
Frequency (TF-IGM) [17] evaluate how frequently we use a word.

Analysing the contextual structure of how we shape our sentences can be done with the help of tools like
Doc2Vec18], Topics [21] which provide representation at the paragraph/document level. Readability [19] of the text
can be valuable resource in identifying a person. Semantic modelling of the text by identifying the Polarity [21] is
helpful in identifying whether the person is positive or negative.
Table 2: Low level features extracted for personality estimation using text

Approach Mode Low level Features extracted


Temporal Features [11]
Word Level Part of Speech (POS) Tags [12]
(Context Independent) Bag of Words( BoW) [13]
Named Entity [14]
Function Words [15]
Word Count/ Uni/Bi/N-Grams [16]
Frequency Features TF-IDF [16]
TF-IGM [17]
Doc2Vec [18]
Readability [19]
Clause Analysis
Phrases [20]
Topics [21]
Open Vocabulary
(Bottom Up) Semantic Features Polarity [22]
Skipgram [24]
Word2Vec
Common Bag of Words (CBoW) [24]
Word Embedding
Glove [26]
FastText [27]
Topic2Vec [25]
Paragraph Vector-Distributed Memory (PV-DM) [28]
LM/ElMo [29]
ULMFiT [23]
Gated Recurrent Unit (GRU) [30]
Language Model
AWD-LSTM [31]
Embedding
BERT [32]
BERT LimitBERT [33]
Roberta [34]
Linguistic Inquiry and Word Count (LIWC) [7]
Mairesse features [35]
MRC Psycholinguistic Database [36]
Dictionary of Affect in Language (DAL). [37]
SPLICE [38]
Psycho Linguistic Style H4Lvd features [39]
SenticNet [40]
Closed Vocabulary
AFINN [41]
(Top Down)
NRC Emotion Lexicon [42]
NRC VAD Lexicon [43]
WordNet [44]
Word Sense ConceptNet [45]
EmoSenticNet [46]
Speech /
Speech Acts [47, 48,49]
Dialogue Acts

Word Embedding models provide word level representations which could be directly to the computational
model. The Word2Vec [24] model was introduced by Google research team for creation of word representations. It
also uses a neural network with Softmax activation and follows either Skipgram or CBOW architecture for
construction of word representation. Global Vectors for Word Representation (Glove) [26] is a Stanford built model
for distributed word representation. It is an unsupervised model for acquiring semantic vector representations of words.
To build word vectors, the Glove model looks at the co-occurrences of words in a corpus. FastText [27], a tool created
by Facebook, creates vector representations of words using a neural network. It supports continuous bag of words
(CBOW) and Skipgram models using Softmax or hierarchical loss functions, and negative sampling.

5
Pre-trained Language models like ElMo [29], ULMFit [23] and BERT [32] are trained on large corpora to get
language embeddings from text. These are used to catch the associations in the structure of how words form sentences
and sentences shape a paragraph which can be extremely useful in understanding the personality of an individual.

The Open Vocabulary approach is very promising, but is prone to overfitting and as such needs exceptionally
large corpora for training. There is also a lack of transparency due to the closed-box nature of these representations,
so even if the results are exceptionally good, we will not know how they were achieved.

c. Psychological models

Numerous theories have been given for personality analysis but most fall into four themes, Psychoanalytic,
Humanistic, Social Cognitive and Trait theory perspective. However, trait-based theories most successfully predict
quantifiable perspectives within the lifestyle of individuals. The most prominent trait models are listed in table 3.
Table 3: Major psychological models for personality modelling

Personality model Description Traits


Groups personality traits into three categories, cardinal traits,
which shape the person, his/her attitudes, and his/her Cardinal
Allport’s Trait model: Gordon
behaviours; central traits, which are the factors that decide Central
Allport [51]
most of the individual behaviours; and secondary traits, which Secondary
may only be revealed in certain situations.
Considers the human basic values that are recognized in each Security, Conformity, Tradition,
Basic Human Values: Shalom society. In the proposed theory, 10 important different values Benevolence, Universalism, Self-
Schwartz [52] are detected and the disagreement dynamics, as well as the Direction, Stimulation, Hedonism,
agreement between these values, is also mentioned Achievement, Power.
Abstractedness, Apprehension
Dominance, Emotional Stability
Liveliness, Openness to change
Cattell’s 16 personality factor 16 essential personality factors, which are listed under five Perfectionism, Privateness
model [53] major categories. Reasoning, Rule-Consciousness
Self-Reliance, Sensitivity
Social boldness, Tension
Vigilance, Warmth
Psychoticism
Eysenck’s Giant Three model [54] Originally known as the PEN model Extraversion
Neuroticism
Covers four personality dimensions:
1. Extrovert vs Introvert (E/I)
The Myers-Briggs Type Indicator 16 Personality types e.g.
2. Sensing vs Intuition (S/N)
(MBTI) model: [55] INFP, ESTJ, ISTP
3. Thinking vs Feeling (T/F)
4. Judging vs Perceiving (J/P)
A three-dimensional model which distinguishes people in Narcissism, Machiavellianism, and
The Dark Triad [56]
terms of dark, non-pathological categories. psychopathy
Honesty-Humility, Agreeableness,
A six-dimensional framework which predicts several Emotionality, Extraversion,
HEXACO Model [58]
personality phenomena that Big 5 model does not. Conscientiousness, Openness to
Experience
This is so far the most generalised model which classifies the Openness, Contentiousness,
Big 5 / Five Factor Model
personality in five main traits, extending the Big three model Extroversion, Agreeableness, and
(OCEAN) [50]
of Eysenck. Neuroticism.

A lot of psychologists have tried to show the number of traits that uniquely distinguish an individual. This
has resulted in several theories with the number of traits ranging from 2 to 16. Such traits were obtained by a factor
analysis over the adjectives used in personality description questionnaires. Big five [57] or “Five Factor Model” [50]
is the most widely accepted trait theory, so we have limited our review to include only works using the Big Five
personality traits.

The Big 5 (OCEAN) traits can be broadly counted as:

6
Openness to experience (O): Imaginative versus Realistic; this describes the ability of a person to accept new
things, adapt to new environments.

Conscientiousness (C): Careless versus Organized; People with this trait like to work ahead with a plan to
make the maximum of their time.

Extroversion (E): Shy versus Sociable; People who rate high on extroversion are comfortable making
connections with new people and tend to be enthusiastic and talkative.

Agreeableness (A): Friendly versus Uncooperative; this trait decides if a person would suit better in a
collaborative or a social work atmosphere.

Neuroticism (N): Emotionally stable versus insecure; highly neurotic individuals are very volatile and not able
to control their emotions.

d. Computational models

The computational model is the algorithmic black-box that takes input (text) and provides output (personality).
The result depends on how the task has been modelled for computation; a personality score for a person or whether
they have a particular trait. The problem modelling as such would require distinct levels of granularity for the system
to discriminate between individual users. The is done in either of the following ways:
Classification:
When the output generation is limited to extremes of personality traits, such as extrovert vs. introvert,
Contentious or Organized, Binary Classification models can be developed. Due to multiple traits the problem is
inherently a multi-label Classification yet only a few researchers chose to follow that. Multi-Class and Structured
Classification has also been used by some of the existing works. Even though the results would aid a lot of practical
applications but clearly such modelling of the problem would not be accurate as personality varies continually among
members of the population.
Regression:
If personality score is modelled as a set of scalar values, we can treat the computation as a regression problem.
This would provide us with the exact level of personality score that can be attributed to an individual. This method is
precisely synonymous to the classical score-based personality assessments. Further, if we apply a certain threshold
score obtained from regression modelling can also yield a personality label class High or Low. We have found a lot
of researchers use Univariate in which separate models are trained for each trait and some use multi-variate regression
where a single model is trained for training the traits jointly.

Ranking:
With scalar values, we can also rank the individual personality scores by modelling it as a Ranking problem.
Mai07 used RankBoost for pairwise ranking as per the personality traits. Wald12 uses machine learning for ranking
individuals as per their personality traits from Facebook data. There are some works that use ordinal regression though
we could find no such work for text-based data. Ranking the individuals as per personality traits has some practical
applications as well like in recruitment and dating industries but it is not an exact descriptor of Personality prediction
because we can rank people using the scores from regression modelling but not the opposite.
Clustering:
There is an enormous potential of using clustering with unseen data using unsupervised models for clustering
people with similar personality traits. This approach has been used very less in the personality computation domain,
but we did find some works that have treated personality recognition as a clustering problem. However, they have
used Jungian based trait theories as a means for the cluster so they have not been included in our review. Clustering
can also be used for finding the exact number of personality traits present in human language and otherwise.

7
Figure 1: Computational Models used for APRT

A hierarchical classification of computational models that have been used has been given in Figure 1. After
establishing what the model does, using above approaches, we come to how a computational model operates to achieve
it. We have the various works in personality computation literature may be divided into Supervised, Unsupervised,
Semi-Supervised and Logic based:

Supervised

Supervised learning models use data annotated with personality scores which are either self-reported
(calculated using questionnaires provided to the individual) or observer reports (calculated by external
judges/psychology experts). The computational models use this data to establish patterns to predict personality traits
from unseen examples.

Unsupervised

Models using these techniques use un-labelled data for learning to find underlying patterns in the data to form
clusters of individuals with similar traits.

Semi-Supervised

Semi-supervised learning uses a combination of Supervised and Unsupervised techniques to learn individual
personality differences from the underlying text data.

8
Ontology based modelling

The Ontology based modelling systems use formal representation of the latent patterns in text language used
by an individual. Such modelling requires unambiguous definitions of personality differences which can be put down
into well formed representations.

e. Performance metrics

The trained computational models are to be evaluated for effective performance. Various performance metrics
have been used by researchers for calculating the efficiency of personality computation models. There is no uniform
metric that has been used by the research community for evaluation of personality computation, however, some of the
metrics that have been used are listed in Table 3. It is important to have a standard metric for comparison of existing
works to form an exact opinion on how much has been achieved in the domain. Further, psychometric standards for
validating the automatic personality assessments are also required.
Table 4: Metrics used for performance evaluation

Task Metrics Description


Accuracy (Acc) It is defined as the ration of correctly classified examples to the number of total
examples.
Precision (P) It is defined as the ratio of number of correctly predicted examples classified to a label
to total number of examples predicted to that label.
Classification
Recall (R) It is defined as the ratio of number of correctly predicted examples classified to a label
to actual number of examples of that label.
F1 – Score (F1) This is calculated as the weighted mean of Precision and Recall.
Area under curve (AUC) It is calculated
Mean Absolute Error (MAE) It is calculated using mean of the absolute values of the difference between actual and
predicted values
Mean Square Error (MSE) It is calculated as the mean or average of the squared differences between predicted and
Regression
expected target values in a dataset.
Root Mean Square Error (RMSE) Square root of MSE is calculated to obtain this as to get the exactness of the error
R – Score Square root of coefficient of determination
Subset Accuracy (SA) It reports whether the subset of labels predicted matches the ground truth.
It reports on an average how many times a label was incorrectly predicted. For each
Multi-Label Hamming Loss (HL)
incorrect prediction, the value increases by 1 and smaller HL means better performance.
Classification
Micro F1 (MinF) It is the F1 Score for an individual label
Macro F1 (MacF) It is the joint F1 Score of all the label

Personality prediction is a challenging task for humans and the same applies to computational models. However,
the recent advances in computer science have led to the belief that given enough data, computers might be able to
identify the latent personality constructs better than a human observer. To understand and evaluate the progress of
domain (APRT), we base our review on the following questions:

 How good are text data sources for automatic modelling of Personality?
 Are the features that have been used, adequate for understanding personality differences in text?
 Which Psychological theory is the best for automatic recognition of personality?
 What constitutes a good performing computational model?
 How good are the existing performance metrics for performance evaluation of automatic personality
recognition systems?

9
3. Literature review: Lingual Makers for APRT:

The first generation of personality computation researchers were focused to show correlation between the
lingual markers used in their writings and the personality traits, the cue establishment phase. What follows is these
markers yielding automatic personality assessment models, the cue utilization phase. It is also important to evaluate
intrinsic and extrinsic performance of the cues and the assessment models, the cue validation phase. Though the
phases run in parallel, to understand the progress of text personality computation we shall explore them in isolation.

a. Cue Establishment

This section supplies an overview of the existing works focused on setting up lingual markers that distinguish
between individuals.

Pennebaker and King [7] (Penn99), proved linguistic style as an independent and meaningful way of
automatic personality recognition. They developed the LIWC categories correlating personality traits and language
use. Gill and Oberlander [59] (Gill02) studied the role of lexical features in predicting Extroversion. They report
LIWC and MRC represents Psychotism and Neuroticism better than Extroversion. They also conclude Extrovert
language is less formal, positive, and longer than the Introvert. However, both are closed vocabulary features and do
not take account of the context. Gosling et al [60] (Gos02) proposed two mechanisms for personality expression,
Identity claims and Behavioural residues. Identity claims are the symbolic statements made by individuals about how
they want to be regarded. Behavioural residues are the unintentional traces of a person's behaviour. Vazire and Gosling
[61] (Gos04) established personality identification by looking at personal websites, which consist of identity claims.
The results suggest Openness to be the easiest trait to be judged. Further it was found out that the scores of Extroversion
and Agreeableness were slightly enhanced, yet accurate. The correlations supply compelling evidence that Identity
claims alone let us know about the personality traits of a person. Mehl, Gosling and Pennebaker [62] (Mehl06)
evaluated the association between Big 5 traits and informal language used in daily context. They collected language
usage and several other features using an Electronic Activated Recorder, correlating it with big five personality traits.
The results suggested the time spent talking was indicative of students Extroversion, class attendance was positively
correlated with Conscientiousness and swear word use was negatively correlated to Agreeableness. It was also seen
that the correlation of language and traits has a remarkable difference across genders. Gosling et al in [63] (Gos07)
proved Facebook profile as a valid means of personality identification. Their findings suggest moderate to strong
consensus for identification of all personality traits from Facebook profiles. However, it also suggested self-
enhancement for traits like Emotional Stability and Openness to Experience.

Tauscik and Pennebaker [64] (Tau10) address the validity concerns of LIWC categories and its potential
usage for identifying Personality and behaviour differences. They also highlight the possibility of advancements for
developing a similar categorical model incorporating the structural and semantic concerns in language. Yarkoni [65]
(Yar10) employed a large corpus for finding associations between word categories and personality traits as well as
facets of the traits. Their results suggest significant correlation that can be used in personality assessments based on
language use. Further, they also investigated associations between personality traits and their facets with individual
words. Their effort highlights the importance of studying the influence of personality on language at multiple levels
of analysis to form new categorical with more insights into Personality assessment compared with LIWC. Holtgraves
[66] (Hol10) used 224 University students for examining associations between text messages and personality traits.
The results suggest significant correlations between LIWC categories from the text messages and personality traits of
the students. Based on their earlier work, Iacobelli, Nowsan, Gill and Oberlander [67] (Ia11) explored use of bi-grams
for personality trait estimation of bloggers. The results suggest that using Bi-grams was better for modelling
personality traits. This suggests the importance of language structure in identifying personality traits. Qiu et al [68]
(Qiu12) examined the associations between personality traits and linguistic cues in microblogs. Due to insignificant
correlation, self-reports and single observer reports are not sufficient. They also calculated Inter-Observer agreement,
showing moderate to strong agreement on all the big 5 traits. Shwartz et al [69] (Shwa13) used Facebook based data
for Differential language Analysis (DLA), an open vocabulary analysis and got significant correlations for the big 5
traits with R score of up to 42 percent compared to the LIWC baseline score of 29 percent. This supplies evidence of
elements in language analysis that have not been captured by the lexical categories set up in LIWC.

10
Park, Kosinki et al [70] (Park14) proposed a Language based Assessment (LBA) for modelling personality
traits using Facebook statuses from extended MyPersonality dataset. They trained a Ridge regressor on late fusion of
word/phrase relative frequency with Topic usage, reduced using univariate feature selection and randomised PCA.
The reported results suggest a significant correlation between the LBA results, self, and observer personality reports.
Hassanein et al [71] (Has21) used twitter data to interpolate the relationship between Big 5 traits, Human Needs and
Personal values. Pearson correlation was used to measure the correlation between the attributes and the traits. Stajner
and Yenikent [72] (Staj21) studied why it is difficult to model personality based on MBTI scores as compared to Big
5. Giorgi et al [73] (Gio22) showed regional differences in personality traits with social media language. A language-
based assessment was performed on over 6 million twitter users from different counties of the United States. The
results suggest use of language can help us find the personality trait differences between people across regions.

b. Cue Utilization

After the establishment of links between language and personality traits, we require efficient computational methods
to model personality from language. In this section we analyse the approaches that predict personality traits from
several types of textual data. As stated before, we have only considered works based on textual data and Big 5
personality traits. The task of Automatic Personality Recognition form text (APRT) has received a lot of attention
from researchers using diverse techniques. Several Shared workshops WCPR 13 [8], WCPR 14 [74], PAN-AM 15 [9]
for APRT have been conducted to establish benchmarks for the domain. For the sake of convenience, we have listed
the existing works as per core learning mechanisms used viz; Classical Machine Learning, Logic Based Learning,
Deep Learning, Graph Based Learning in Section 4.

c. Cue Validation

In this section we review the existing works that have worked on confirming the personality assessment based on
lingual markers.

Buchanan and Smith in [75] (Buc99) investigated the use of the Internet for personality recognition. They
propose a psychometric approach for personality assessment based on the World Wide Web (WWW). This could be
adapted for Automatic Recognition of Personality from User Generated Content (UGC) in formal as well as informal
contexts. Youyou, Kosinki and Stiwell in [76] (Kos15a) found the predictions of regression-based models better than
human observers. The Automatic assessments correlated strongly with self-ratings of the participants exceeding the
correlation scores with human based judgements. Novikov et al [77] (Nov21) investigate the accuracy of a subset of
automatic personality models. They screen the works having distinct training, validation, and test phases. Pearson
correlation was used to evaluate the performance of the existing works as diverse metrics have been used. They also
suggest validation of automatic prediction models with psychometric measures for robustness and generalizability.

4. Literature review: Computational modelling of APRT

This section supplies a brief overview of the various works trying Automatic Personality recognition from Text
(APRT). A diverse range of machine learning models and feature sets has been employed with different datasets and
performance metrics. A lot of researchers have reported moderate to good accuracy for the task, however given the
different mechanism of how they approach the problem and confirm their results, it is exceedingly difficult to have a
clear idea of how good one model fares compared to others.

a. Supervised Learning

Most of the researchers have employed labelled datasets to train their models for APRT, however there is a lot of
divergence in the training mechanism. We have divided them into the following sub-categories:

Classical Machine Learning Models

One of the first works that used text for personality profiling using machine learning was Argamon et al [78]
(Arg05). Support vector machine with linear kernel (SMO) is used to classify high and low classes of neuroticism and
extraversion in authors of informal text using computational stylo-metrics and systematic functional grammar (SFG).
The corpus used was based on essays written by psychology students at University of Texas from 1997 to 2003.

11
The reported accuracy for Binary classification was around 58 percent which was an absolute 8 percent
improvement over their baseline. (Division by thirds, upper third is high, lower third is low, middle third is ignored).

Mairesse and Walker [79] (Mai06) tried modelling personality traits using a combination of lexical features
and Speech acts. They performed Regression as well as Binary Classification of Essays and EAR Corpus datasets and
their results suggest improvements over the baseline. The improvements however seem better for the informal/spoken
language from EAR Corpus than the formal texts from the Essays. Oberlander and Nowsan [80] (Ob06), in an attempt
of personality profiling with Weblogs, used Naïve Bayes (NB) model for Binary as well as multi-class classification
with n-grams as features. Due to skewed distribution over the yes and the no classes, openness trait was excluded for
this work. For their hardest multiclass classification task, they improved the classification accuracy of Extraversion
by 10.9 percent (32.2% relative) and Agreeableness by 30.4% (77.2% relative) from the Arg05 baseline. In a continued
work [81] (Now07), they used the classifiers trained with the first data to classify a larger, noisier corpora. The results
are not as good as the original yet promising. This suggests transfer learning can be used in which models trained on
clean, balanced datasets can be generalized for deployment on noisy unseen data.

A seminal work in text-based personality computation by Mairesse et al [35] (Mai07), extracted 88 LIWC word
categories and 14 MRC Psycholinguistic categories from two datasets, ESSAYS, and conversation extracts from EAR
Corpus. Taking the Big 5 traits as reference, they tackle personality recognition from text as classification, regression,
and ranking problems. For the Binary classification, six different algorithms: C4.5 decision tree learning (J48), Nearest
Neighbour (NN), NB, Ripper (JRip), AdaBoost, and SMO were used. SMO gave the best results for ESSAYS with
62.11 percent accuracy for Openness to experience, while as for EAR Corpus NB and AdaBoost performed the best
with an accuracy of 73 percent for Extraversion. For regression a linear regression model, an M5’ regression model
tree, an M5’ regression model tree returning a linear model, a REPTree decision tree, and a model based on Support
vector machines with linear kernels (SMOReg) were used. The M5’ model tree performs the best with an error
decrease of 6.7 percent from the baseline for ESSAYS dataset while models perform badly over the small sized EAR
Corpus. For the ranking problem, RankBoost was used. For the ESSAYS data, Openness to experience produced the
best ranking model with loss of 0.39 and agreeableness produced the worst with loss of 0.46, while for EAR Corpus
Agreeableness and Conscientiousness produced loss of 0.31 and 0.33 outperforming the rest.

In [82] Estival et al (Est07) used email data of 1033 respondents with a joint total of 9836 email messages.
They used Part of Speech (POS) tags and a custom named entity organizer (NE). Five demographic traits were
distinguished along with the Big5 traits. The best reported accuracies for the psychometric traits are 56.7 percent for
LibSVM. Golbeck et al in [83] (Gol11a) worked on prediction on personality from social media. They extracted at
least 2000 tweets of 279 users along with their self-reported personality assessments using BFI. LIWC and MRC
features were obtained using a text analysis tool along with a word-by-word sentiment analysis of the tweets. Gaussian
process (GP) and ZeroR (ZR) algorithms were used for regression analysis of the features to predict personality
information. Openness was the best computed trait and neuroticism was the worst with mean absolute errors (MAE),
12 and 18 percent respectively on a normalized 0-1 scale. In [84] (Gol11b) they used 161 statistical features from
Facebook profile pages of the 279 people for personality analysis. LIWC and certain hand-crafted features were used
for regression using M5’Rules (M5R) and GP. The MAE for each personality trait was around 11 percent for both
the algorithms used. In [85] Iacobelli et al (Ia11) performed a personality estimation of bloggers based on a variation
of features involving bi-grams, n-grams, stop-words and LIWC categories. They used SVM for Binary classification
with the best accuracy for Openness 84.4 percent, and the worst accuracy for Neuroticism at 70.5 percent. Quercia et
al in [86] (Qu11) also have worked on personality prediction from social media. They tried to map the influence scores
of 335 twitter users to personality traits. Regression analysis was done by using the M5R algorithm with 10-fold cross
validation that achieved results with around 0.88 root mean square error (RMSE). They analysed the correlations
between personality traits and user behaviour, classifying the users as listeners, popular, hi-read, and influential.

Adali and Golbeck [87] (Gol12) analysed behaviour on social media for predicting personality traits of people.
They used GP and ZeroR for regression on twitter-based data and the reported MAE for Big 5 traits was in the range
of 11 to 19 percent. Bai et al [88] (Bai12) have also predicted personality from user behaviour on social networks.
They used Chinese social network platform Renren. A custom set of 41 features including basic information, emotion
and time distribution were used on 209 individuals. C4.5 classification algorithm with gain ration was used and

12
achieved accuracy around 0.8 for at least four of the five traits. Kermanidis, L in [89] (Ker12) used SVM with Bagging
for recognition of personality traits from Greek essays using lingual features combined from LIWC and MRC. The
reported accuracies range from 64 to 86 percent. Wald et al [90] (Wal12) used several machine learning models,
RepTree, DTable and Linear Regression for ranking individuals according to their personality traits using their own
data collected from Facebook. Shen, Brdiczka and Liu [91] (Liu13) used Emails for personality prediction using three
generative models, Joint, sequential and Survival models. NB, DT and SVM techniques were used for training and
the reported results suggest Survival mode, in which the features are selected independently for each label performs
the best. The joint model has the worst performance as it assumes features being selected jointly by all labels. For the
selection model, labels are sequentially selected and they in turn select the features. They also tried to measure trait
specific predictive power of their extracted email features using Information Gain.

Firoj et al in their effort to recognise personality traits in [92] (Fir13) followed, bag of words approach using
uni-grams as features. They trained SMO, Bayesian Logistic Regression (BLR), and Multinomial Naïve Bayes (mNB)
classifiers with MyPersonality dataset and the reported accuracies range from 50.1 – 69.6 percent, with mNB achieving
the best performance for binary classification. Verhoeven et al [93] (Vdd13) used SVM and ensemble learning to train
classifiers for each of the five traits. SMO based meta-learning experiments using most frequent character trigrams
(MFT) were used for Facebook as well as Essays datasets independently. The reported results have an average f-score
of 0.72 over a 50 percent baseline. In [94] Farnadi et al (Fa13) predicted personality traits for both the datasets using
four sets of features including LIWC categories, temporal features like frequency and number of status updates, social
network features and other features like number of statuses per user, number of capitalised words/letters/urls/proper
names and words used more than once. They used SVM, kNN and NB algorithms with a weighted average of
precision, recall and f-measure based on class average as evaluation measures. The reported results suggest that time
features perform poorly with NB, while other features/algorithm combinations work quite well. Cross domain
generalisations from MyPersonality to Essays supplied poorer results than the other way around, still outperforming
the baselines. This concludes that the training set size matters in performing personality prediction using cross domain
datasets.

Tomlison et al in their effort to predict conscientiousness [95] (To13) from Facebook statuses. They relied on
using the verbal specificity and objectivity of verbs used for predictive accuracy of personality traits. It was found out
that conscientiousness is normally distributed within the dataset and as such better results may be reached by an
approach that predicts only the outliers. A linear regression model was used to analyse the features and the best
reported RMSE was 0.63. Facebook data was also used by Markovikj et al in [96] (Mar13) for personality trait
prediction. They used a variety of lexical and affective features as well as social and demographic information from
the MyPersonality dataset. SVM with Boosting was used for predicting personality traits and the reported accuracy
was almost 90 percent. Iacobelli et al in [97] (Iac13) tried personality trait recognition on both the datasets. The posts
in MyPersonality corpus were combined based on their authors, and unigrams, bi grams and n grams were combined
as features for each document, and the same was done for Essays. Conditional Random Fields were used for structured
classification in addition to NB, SMO, and Logistic Regression (LR). The reported accuracies fall in between 44.8 –
66 percent for MyPersonality and 51.62 – 61.81 percent for Essays.

Speech act annotation was used for MyPersonality corpus by Appling et al [98] (App13). They reported the
correlations to personality traits, which suggested some of the traits positively being predicted with speech acts and
others negatively related. As per their work Openness does not predict any speech act though. Saif et al in [99] (Sai13)
predicted personality traits using different lexical features like the ones used by Mai07, Token Unigrams, Average
Information Content (AIC), NRC and Turney Lexicons. They trained a SVM classifier with the Essays dataset and
the reported accuracy ranges from 55-60 percent. Poria et al [100], (Por13) combined common sense-based features
with psycholinguistic features and frequency-based features to be employed for classification using SMO based
supervised classifiers. Essays dataset was used and the reported results in the range of 63 to 66 percent.

In [101], Zuo et al used a weighted machine learning k-Nearest neighbour ML-kNN to predict personality traits
from Essays dataset. They used custom linguistic and emotion base features and the reported accuracy is about 65
percent. In [102], Gou et al (Gou14) performed automatic recognition of personality from twitter data. They used
LIWC word categories for the Big 5 trait profiling of 256 twitter using employees of IBM. Lima and Castro in [81]

13
(Lim14) used social media data for personality prediction. Supervised learning for binary classification with NB, SVM
and MLP was applied on three available datasets based on twitter data. The reported prediction accuracy is about 83
percent. In Pratama et al [103] (Prat15), a modified version of MyPersonality dataset was used for personality analysis
using MNB, KNN, SVM and a combination of the three. The reported results are around 60 percent. They also
performed respondent testing for the automatic personality prediction system using twitter data and the accuracy for
that was around 65 percent (multi-Label) Arroju, Hassan and Farnadi [104] proposed an Ensemble Regression Chains
Corrected (ERCC), a multivariate regression model for predicting personality traits using N-grams and extended
LIWC categories. They used multi-lingual twitter data from PAN 2015.

Poddar et al in [105] (Pod15) introduced adjectival marker technique for personality recognition from ONSW.
They used biographical data of 574 personalities, extracting information from websites using a python web crawler.
Regression using LASSO resulted in personality recognition accuracy ranging from 81 to 92 percent. Lukito [106]
(Luk16) used feature reduction on the Essays dataset. Information gain and PCA resulted in a 30 to 95 percent
reduction in the original feature size of LIWC categories. The algorithms used were SMO, LibSVM and Linear
Logistic Regression (LLR) against a baseline ZeroR and the results reported are in the range of 51 to 62 percent.
Pramodh et al [107] (Pra16) used Essays and MyPersonality datasets for personality recognition of authors. They
employed polarity-based matching, POS tagging and parsing for personality score computation and the reported
accuracies ranged from 62 to 66 percent. Ong et al [108] (Ong17) proposed a personality prediction system using N-
grams of twitter information. They use XGBoost and SVM on custom Twitter data and the best reported results are
76.23 and 97.99 per cent. Tandera et al [109] (Tan17) used both shallow and deep learning algorithms for personality
classification on MyPersonality and a custom Facebook based dataset. The traditional algorithms used are NB, SVM,
LR, GB and LDA with LIWC, SPLICE and SNA feature sets with an average accuracy from 61 to 67 percent. They
also used deep learning algorithms MLP, LSTM, GRU and CNN with word embedding features obtained with GloVe
and reported average accuracies from 59 to 83 percent.

Zaaba et al [110] (Zab17) used Facebook data for personality prediction of Malaysian users. Linguistic features
from the LIWC categories along with activities and structural features were analysed for correlation with the
personality to find the optimal sets per trait. Yata et al [111] (Yat18) used MyPersonality dataset for personality
prediction. They used MNB, SVM and kNN models with several multi-label classification techniques, Binary
Relevance (BR), Classifier Chains (CC) and Random k-Labelsets (RAkEL). The reported results suggest kNN with
CC and RAkEL are the optimal performers. Arjaria et al in [112] used Multi Label Naïve Bayes for predicting
personality traits from ESSAYS and the best reported results have an average accuracy around 87 percent. However,
they have not mentioned anything about the features that they used for training the model.

Artissa, Asror and Faraby [113] (Art19) tried variations of pre-processing techniques like Stemming and
Lemmatization using a MNB Classifier on MyPersonality dataset. Ergu et al. [114] (Erg19), used Turkish tweets for
personality computation. A variety of machine learning models (KNN, Decision tree (DT), RF, AdaBoost, Stochastic
Gradient Descent (SGD), Gradient Boosting (GB), and SVM) were tested to achieve this goal. The best accuracy was
reported with models trained on users' most recent 50 tweets, ranging from 0.76 to 0.97. Rohit et al. [115] (Roh20),
conducted research to predict user personality from social media profile status information. Then they further
categorized the users’ personalities into one of the OCEAN model's categories based on the analysis results and their
model achieved good accuracy. Ong et al [116] (Ong21) tried personality prediction for Indonesian users with twitter
information. They use XGBoost and SVM on custom Twitter data and the reported results show superior performance
for Agreeableness and Openness traits. Safitri and Setiawan [117] (Saf22) experimented with different various Kernels
using LIWC and TF-IDF features to predict personality traits from custom Twitter data. The results suggested RBF
kernel with joint LIWC and Tf-IDF as the best performer.

Vu eta al [118] (Vu18) explored various lexical-semantic features based to improve Personality prediction on
four diverse text datasets. The utilised combinations of WordNet based semantic knowledge with Word sense
disambiguation to train SVM binary classifiers for personality trait estimation. The reported results show improvement
over the Mairesse baseline for Essays dataset.

14
Ensemble Learning

Fernandes et al [119] (Fer20) uses adjective selection from Sausier's test to predict personality. Gradient
Boosted regression and Classification Trees were used on a custom dataset obtained by crowdsourcing. The reported
results suggest this to be a significant method for personality trait estimation using the adjectives best suited for a
particular trait.

Artificial Neural Networks

One of the first works using neural networks in text-based personality computation is Kalghati et al [120]
(Kal15). They used a twitter API to fetch tweets which after pre-processing are fed to a multi-layer perceptron (MLP),
one for each trait among the Big 5.

Recurrent Neural Network

In [121] Su et al, (Su16) used Chinese Linguistic Inquiry and word count (CLIWC) and Chinese Knowledge
Information Processing (CKIP) POS tags to extract linguistic features from MCDC dataset. 10 RNN’s were
independently constructed to model the items of the BFI-Inventory and then a Hidden Markov Model (C-HMM) was
used to predict the personality traits in a dyadic conversation with reported results up to about 88 percent. Liu et al
[122] (Liu16) used RNN based character to word to sentence feature method for prediction of personality traits from
a multi lingual dataset from PAN-AM-2015. This model is feature engineering free and language independent and
personality traits can be predicted purely based on text without any added features.

Xianyu et al in [123] used Heterogeneity-Entropy Neural Network (HENN) based on Deep Boltzmann Network
(DBN) and Auto-encoders for prediction of personality traits from data collected from Renren and Sina websites. Bag-
of-Textual-Words was used to extract a 1000-dimensional feature vector and Support vector regression achieved Mean
Absolute Error (MAE) in the range of 0.12 to 0.72 for text modality. Sun et al [124] (Sun18) proposed a bidirectional
LSTM model concatenated with a CNN called 2CLSTM for personality prediction in Essays and YouTube datasets.
Glove was used for obtaining word embedding. Latent sentence group (LSG) was introduced for modelling structural
features at sentence level and CNN used to learn the features. The average reported results for Essays and YouTube
are 55 and 60 percent, respectively.

An and Levitan in [125] (AnLe18) proposed a model that learns a linear combination of predictions made by
MLP and LSTM to produce a final fused prediction. They used MyPersonality dataset with a combination of LIWC
categories to the word embeddings generated by DAL and word vectors from Skip-gram and Glove. The reported
results have an average prediction accuracy of 67 percent. In [126] Yilmaz et al (Yil19) used LSTM based model for
personality detection from Turkish texts. They translated an existing dataset into Turkish and then an RNN based
model was used to predict the big five traits from the Turkish texts. The test accuracies reported are in the range of 55
to 68 percent.

Kazameini et al. [127] (Kaz20), created an efficient and robust text-based deep learning model for predicting
personality. They used context independent embedding word2vec and BERT, along with psycholinguistic features, to
a Bagged SVM (BB-SVM) classifier, reporting accuracy of about 59 percent.

Leanardi et al [128] (Lea20) proposed sentence embeddings for Personality prediction from text. They use an
optimised BERT architecture for extracting embeddings from text and propose a Stacked Neural Network. In three
different experiments, they used sentence Embeddings to train SVM repressor, FastText word embeddings to train
their stacked NN and finally Sentence Embeddings to train their stacked NN. The results of the Sentence embeddings
with Stacked NN outperformed the other two as well as other existing works on MyPersonality dataset.

In Xue et al. [129] (Xue21), created a unique semantic-enhanced personality recognition neural network
(SEPRNN) that can identify many personality features. They employed context-learning-based word-level semantic
representation followed by a fully connected layer to acquire text's higher-level semantics and when compared to
different baselines their suggested strategy improves accuracy significantly. Demerdash et al. [130], (Dem-21)
presented a deep learning approach for personality evaluation using fusion approaches pre-trained Language models

15
for transfer learning, and their proposed model outperforms the baseline results. Christian et al. [131] (Chr21),
proposed a new feature extraction strategy for various social networking data sources based on a multi-model deep
learning architecture paired with numerous pre-trained language models such as B ERT, RoBERTa, and XLNet, to
develop personality prediction systems, and their model achieved good accuracy. Jeremy et al [132] (Jer21) used
LSTM, Bi-LSTM and GRU models for personality prediction of Indonesian twitter users. The highest reported F
Score is 83 percent.

Mavis, Torslu and Karagoz [133] (Mav21) used Turkish twitter data for personality trait prediction. They
experimented with several classical models like SVC, DT and kNN using embeddings based on TF-IDF, word2Vec.
Comparative analysis of the results with MyPersonality dataset on the same models suggests that the approach can be
used on other languages as well. They also trained a LSTM network with the Turkish data but no significant
improvement was achieved except for Agreeableness trait. In Kosan [134] (Kos22) used LSTM and Bi-LSTM models
for personality prediction from a twitter-based dataset. They experimented with FastText embedding and varying pre-
processing techniques to arrive at their final model, which uses specific pre-processing methods like stemming,
lemmatization with fasText and Bi-LSTM. The average reported results for all the Big 5 traits are RMSE 0.1681.

Convolutional Neural Network

In [135] Majumder et al (Maj17) used CNN models for deep semantic feature extraction from the Essays
dataset, they also used word2vec with Mairesse baseline features and SVM, MLP and sMLP/FC algorithms were used
for classification with reported results ranging from 51 to 63 percent. Yu and Markov in [136] (Yu17) used CNN and
RNN for the shared task of WPCR-13 and the results are significantly better, with an F1 score in range of 60 to 65
percent. Giménez et al in [137] (Gim17) used CNN with word embeddings from Glove for personality analysis on the
PAN-AP-2015 corpus. They trained five models varying the number and height of kernels and the reported accuracy
is in the range 0.144 to 0.227 RMSE. Xue et al. [138] (Xue18), presented a textual context personality recognition
method based on deep learning. They suggested a hierarchical structure based on AttRCNN for this purpose, which
can learn deep semantic properties using user posts. The results are promising that prove that the suggested deep
semantic attributes surpass the baseline characteristics.

Rahman et al [139] (Rah19) used Deep CNN for personality classification on the Essays dataset. They used
three different activation functions in which leaky ReLU outperformed by four, and tanh outperformed CON over
sigmoid function, with average F1 scores of 33.11 %, 47.25 % and 49.07 % for sigmoid, tanh and ReLU, respectively.
Darlinsyah et al [140] (Dar19) proposed SENTIPEDE, a sentiment-based model combining CNN and LSTM for
personality prediction from short texts. They extracted Sentiment features from Twitter based data using VADER,
and then utilised pre-trained GloVe embeddings for training the joint model. The reported results suggest improved
accuracy over the chosen baselines. Mehta et al. [141] (Meh20), developed a unique deep learning-based approach to
assess personality from the Essays dataset for OCEAN traits and the Kaggle dataset for MBTI using language
modelling features in conjunction with conventional psycholinguistic features. For both datasets, the results showed
that language model embeddings often outperformed traditional psycholinguistic features, and suggested models
outperformed the selected baseline models. Deliami, Sadr and Nazari [142] (Del22a) used AdaBoost on top of CNN
for personality prediction from Essays. They proposed the use of separate pooling and classification layers to extract
relevant features, and AdaBoost for increasing the learning rate. Deliami, Sadr and Tarkhan [143] (Del22b) used a
combination of CNN and AdaBoost with varying filters for personality prediction form Essays and YouTube datasets.
The reported results outperform the selected baselines. It was concluded that using AdaBoost efficiently increases the
classification accuracy of classification.

Graph and Network Representation Learning Based

Guan et al [144] (Gua20) experimented with Network Representation Learning for prediction of Personality
traits from text data. The proposed Personality2Vec model uses a biased walk, modified Huffman Tree, and Skip-
Gram model to attribute a personality vector for each user in the data. Support vector Regression model was trained
using these vectors to predict personality traits and the reported results outperform the baseline methods chosen by
the authors. Wang et al [145] (Wang20) propose a Graph Convolutional Network for personality prediction from text.
They constructed a heterogeneous graph of user documents using TF-IDF to capture the correlations between the text

16
and the personality traits. The reported results show improved results over the chosen baselines for Essays and
MyPersonality datasets.

Wang et al [146] (Wang21) used a XLNet-Capsule for time series-based prediction of personality from text
data. The proposed model is separately used for extracting deep level features from the text data and then classification
of personality traits. The reported results suggest improved performance over a wide range of baselines selected by
the authors. Ramezani et al. [147] (Ram22) propose a Graph enabled text-based automatic personality prediction.
They built a knowledge graph of low-level text features to be used for training CNN, LSTM and Bi- LSTM Neural
networks achieving accuracies of up to 77 per cent.

Attention Based

Jiang et al in [148] (Jian20) used several deep learning algorithms for personality prediction on ESSAYS
dataset. The algorithms used were Attention based CNN (ABCNN), Attention based LSTM (ABLSTM), Hierarchical
Attention Network (HAN), BERT and RoBERTa. They also adapted the models for a fresh corpus developed called
FriendsPersona. The reported accuracies range from 55 to 67 percent for ESSAYS and 53 to 65 percent for
FriendsPersona dataset

Multi-Task Based

Li et al [149] (Li22) propose a Multi task classification for Personality and emotion detection using deep learning
models. They experimented with different sharing gates for information exchange between the LSTM based models.
It was found that the model using SoG with Model Agnostic Meta Learning outperformed all the other variants and
selected baselines.

b. Unsupervised

In an attempt of unsupervised personality recognition from social networks, Celli [150] (Cel12a) used
correlations from Mairesse et al to create a personality model for a social network-based dataset. The accuracy reported
is in line with that of Mairesse et al. This suggests it is possible to exploit existing correlations between lingual markers
and personality traits to develop comprehensive personality models which do not require subject details. Celli and
Rossi [151] (Cel12b) used a hybrid approach combining the features used by earlier works Mai07 and Qu11. They
developed an unsupervised personality recognition system that exploits correlations to build which does not require
previously annotated data. Liu and Zhu [152] (Liu16) proposed use of stacked Autoencoders for unsupervised learning
of Linguistic Representation Feature Vector (LRFV) based on SLIWC and FFT from Sina microblog. The features
obtained were used to train a Linear Regression model and results outperform the selected baselines. Alsadhan and
Skillcorn [153] (Als17) devised a word count based technique for personality prediction from small texts. They use
Singular value decomposition of document-word matrices for classification by a Frobenius norm. Evaluations on Big
5 as well as MBTI on a varied range of datasets suggest that the proposed technique can be used for low resource
languages as well due to its language independent nature. Celli and Lepri [154] (Cel18) performed a comparative
personality prediction using separate twitter based multi-lingual datasets having Big 5 and MBTI labels. The results
suggest algorithms trained on MBTI could have better performance.

c. Semi-Supervised

Lima and Castro in (Lim13) [155] used a group of tweets rather than the user profile for personality analysis.
They propose a Bayesian Personality Predictor in which each personality trait is separated into one binary classifier.
The benchmark was performed with tweets about Brazilian TV shows. The reported results show an accuracy of
around 84 percent. The same in (Lim14) [156] used a semi-supervised, multi-label approach for personality
recognition. They propose PERSOMA which uses 3 modules in cascade. The first module pre-processes the corpus
extracting meta features in two categories, grammar (lexical) and social (network). The second module transforms the
multi-label problems into single-label ones. The final module performs multi-label classification using five binary
classifiers. Three machine learning algorithms (NB Classifier, SVM and MLP) were used for training of three twitter-
based datasets (Obama-McCain Debate, Sanders, SemEval2013) and the reported results are 83 percent correct.

17
Tighe et al [157] (Tig16) used feature reduction techniques for personality trait classification in Essays. Features
extracted by LIWC categories were reduced with Information gain and PCA. Linear Logistic Regression, libSVM and
SMO were used to train and reported results were comparable to the predecessors with a far lower number of features.
Tighe and Cheng [158] (Tig18) experimented with about 600 models using various combinations of features and
outlier focus for personality recognition from twitter-based data. They used Linear, Ridge, Logistic Regressors and
SVM for regression as well as classification tasks. The reported results suggest that focusing on outliers improves the
results in the case of classification, and feature reduction reduced the accuracy.

Mao et al [159] (Mao18) proposed a particle swarm optimization technique for finding the best features for
Personality trait estimation. They extract a combination of Stylometric and TF-IDF features, which are subjected to
PSO to find the best set of features. Three classification algorithms kNN, NB and DT were trained with these features
and the reported results suggest significant improvements with the use of style features and PSO. Adi et al in [160]
(Adi18) used Stochastic Gradient Descent (SGD), XGBoost (XGB) and Super Learner (SL) algorithms with
Hyperparameter tuning and Feature selection for personality detection from tweets in Bahasa language. The results
achieved show significant improvement over their predecessors.

Carducci et al [161] (Car18) used pre-trained embeddings with transfer learning for personality prediction from
Twitter based text. They trained a SVM classifier with pre-trained FastText embeddings on MyPersonality model and
custom twitter data was used for testing. The reported result outperformed the baseline. They used MyPersonality and
a Twitter based dataset for personality analysis using SVR, LASSO and LR. Distributed word embeddings were used
with Transfer learning and the reported prediction had a MSE of 33 to 70 percent for MyPersonality dataset and 13 to
38 percent for the custom Twitter data.

Santos et al [162] (San18) compared the use of Psycholinguistic, data driven and facet-based approaches for
personality prediction in Brazillian-Portoguese, a low resource language. The suggested results suggest that
translations of the low resource language texts can be used with English based Psycho-linguistic categories. However,
the accuracy of using TF-IDF features and facet-based approach exceeds the one that uses either translated version of
LIWC or using LIWC on a translation of the original text.

Akrami et al [163] (Akr19) used Support Vector Regression (SVR) and Language embedding (ULMFiT)
algorithms for personality classification from text based on four different discussion forums and news sites in Swedish.
Two datasets were assembled, one large with lower reliability (DLR) and another smaller with high reliability (DHR).
Zheng et al [164] (Zhe19) used Pseudo Multi-view Co-training (PMC) as a semi-supervised technique for personality
classification using MyPersonality as the labelled dataset and a custom Facebook based unlabeled dataset. The features
vectors were obtained using LIWC, TF-IDF of n-grams and a word embedding using PV-DM. The reported prediction
was in the range of 62 to 71 percent F1 score

Tighe et al [165] (Tig20) explored use of Neural networks for personality prediction of Filipino twitter users.
They used various combinations of MLP with pre-trained as well as self-trained embeddings. They reported neural
networks using crafted features like TF-IDF performed better than word embeddings. Further, the performance was
poor for Openness models.

Pabón and Arroyave [166] (Pab21) experimented with different word level embeddings for predicting
personality traits. Word2Ved, GloVe and BERT embeddings from YouTube dataset were used for training SVM for
three different tasks Regression, Binary Classification, and a Tri-Class manifestation of personality traits (High,
Medium, and Low). The reported results suggest that Word2Vec and GloVe may be combined to provide better results.

Ontology based

Alamsyah et al [167] (Alm18) proposed an Ontology based approach for personality prediction from Twitter
data. Alamsyah et al [168] (Alm19) built an Ontology model to analyse personality traits from Twitter based text data.
In another effort, Alamsyah et al [169] (Alm21) improve their previously developed Ontology model by using N-
Grams in place of Radix tree.

18
5. Discussion and Open Challenges

Now that we have ascertained the state of the art, we will attempt to answer the questions that were asked towards
the start of this review.

 How good are text data sources for automatic modelling of Personality?

The data that has been used for personality computation is varied across various sources like social networks,
essays, biographies, and smartphone data. This results in lack of coherence and inconsistencies regarding effective
computational analysis of the techniques used. As of now there are no specific datasets that exclusively reveal
personality information. The existing results are filled with Information, subjective and objective biases. Subjective
Biases can arise because of various data variables such age, gender, time, culture/ethnicity, and Nationality in the
dataset. The importance of bias in the performance of behaviour/ personality analysis has been neglected in the existing
literature. Observer reports are better suited for high visibility traits like Extroversion, whereas for low visibility traits
such as Emotional stability, self-reports are a better fit. A combination of the two can be utilized to reduce the number
of biases. Further, stringent standards of rigorous psychometric testing must be put in place to minimize the amount
of noise and biases in the data to be used for personality computation. To address this challenge, we propose a gold
standard dataset incorporating multiple trait labels as well person specific information such as gender, nationality,
culture, education received.

 Are the features that have been used, adequate for understanding personality differences in text?

Effective recognition of personality traits from text requires selection of an optimal feature set. This would involve
identifying psychologically meaningful features that independently discriminate well between categories. Using
multiple features can lead to better estimation of personality, but lead to feature accumulation, raising storage and
computational costs. This has been tackled by estimation of the best features using feature selection algorithms.
However, deep learning models are end to end with inbuilt feature selection processes. Even if this yields optimum
results it is good to know what features are best for identification of personality traits from text. This will aid our
understanding of human behaviour. We propose using feature attribution to attain optimal features that will serve as
input for automatic personality recognition and help us provide a better understanding.

 Which Psychological theory is the best for automatic recognition of personality?

Conceptual clarity is required for causal explanations in personality which in turn depends on the theoretical base
from psychology [50-58]. The trait models are not well defined and lack an empirical basis. More emphasis can be
put into using machine learning models particularly Clustering, for analysing the lingual markers to provide an
empirical trait base that can be used for automatic personality computation.

 What constitutes a good performing computational model?

Robustness and Validity:

The existing literature is majorly based on supervised categorical computational models using labelled data, but
just few attempts have been made to check how well these generalize to unseen data. We suggest training the model
on a large clean corpus and then testing it on different data for development of robust personality recognition systems.
Further, more focus must be given on Semi-Supervised/ unsupervised approaches that would help us to tap the big
data available that can be used for automatic personality assessment of people. Also, in some of the existing works
we found no clear distinction between the Training, Validation and Test data, which makes them susceptible to
over/under-fitting hampering the performance.

Model Accuracy Versus Explainability :

The algorithmic black box, especially for the deep learning models, leads to a lack of trust on efficiency as well
as posing reproducibility challenges for personality computation. Better comprehensibility needs to be achieved
without compromising the precision and accuracy to help understand the internalization of personality computation.

19
This will not only help us to understand precisely what features are best suited for individual trait analysis but also
how we can pre-process the data to achieve better results.

Trust and Ethical Concerns:

Psychometric testing of the automatic personality assessments using parallel testing

Though seamless personality recognition from the vastly available data would be beneficial in a multitude of
manners but the use of such technologies does raise concerns of privacy violation and clear boundaries need to be
established for ethical usage.

 How good are the existing performance metrics for performance evaluation of automatic personality
recognition systems?

There is no existing standard metric for evaluating the accuracy of existing computational models. Throughout
the literature review we can see that different performance metrics like Acc, RMSE, and F-Score have been used. This
can be resolved by devising a uniform task specific standard metric which will facilitate comparative analysis. We
propose to build a correlation-based performance metric which can be used for comparative analysis of the so far work
done as well as for future reference.

6. Conclusion and Future Scope

In affective computation, detecting user behaviour and personality is a difficult and time-consuming task.
Numerous researchers have taken on the task of personality prediction from input text. In this paper, we have examined
recognizing personality from user-generated contexts. We presented an up-to-date overview of existing text-based
personality recognition methods using shallow as well as deep machine learning models. We also discussed important
concerns with current datasets, personality models, features, and approaches, as well as potential solutions. Although
the results have significantly improved with the use of deep learning models, but there are concerns of generalizability,
reliability, and validity of independent personality assessments by automatic models is still far-fetched. However, we
hope with coordinated efforts from psychologists, tests can be developed which are better suited for automatic
personality modelling. To conclude, much research attention is required for betterment in terms of establishing gold
standard dataset benchmarks, model architecture, feature attribution and performance metrics. Active collaboration of
psychologists, linguists and computer scientists needs to address the gaps in psychological theories for providing
empirical paradigms which would not only aid modelling of automatic personality prediction systems but also lead to
important insights into personality psychology.

Statements and Declarations

Limitations of the survey

We tried our best to include all the works in the domain of Automatic Personality Recognition from text, which
makes comparing the computational techniques extremely difficult. Further, non-inclusion of trait theories other than
the Big 5 model might have led to missing computational techniques that have better performance.

Conflicts of Interest

We declare there are no known competing monetary interests or personal relationships that could influence the
work reported in this paper.

References

1. Theophrastus. (4th Century BC). The characters.


2. Papurt, M. J. (1930). A study of the Woodworth Psychoneurotic Inventory with suggested revision. The
Journal of Abnormal and Social Psychology, 25(3), 335.
3. Cattell, H. E., & Mead, A. D. (2008). The Sixteen Personality Factor Questionnaire (16PF).

20
4. Costa Jr, P. T., & McCrae, R. R. (2008). The Revised Neo Personality Inventory (neo-pi-r). Sage
Publications, Inc.
5. Briggs, K. C. (1976). Myers-Briggs type indicator. Palo Alto, CA: Consulting Psychologists Press.
6. Vinciarelli, A., & Mohammadi, G. (2014). A survey of personality computing. IEEE Transactions on
Affective Computing, 5(3), 273-291.
7. Pennebaker, J. W., & King, L. A. (1999). Linguistic styles: Language use as an individual difference.
Journal of Personality and Social Psychology, 77(6), 1296–1312. https://doi.org/10.1037/0022-
3514.77.6.1296
8. Celli, F., Pianesi, F., Stillwell, D., & Kosinski, M. (2013, June). Workshop on computational personality
recognition: Shared task. In Proceedings of the International AAAI Conference on Web and Social Media
(Vol. 7, No. 1).
9. Rangel Pardo, F. M., Celli, F., Rosso, P., Potthast, M., Stein, B., & Daelemans, W. (2015). Overview of
the 3rd Author Profiling Task at PAN 2015. In CLEF 2015 Evaluation Labs and Workshop Working Notes
Papers (pp. 1-8).
10. Biel, J. I., & Gatica-Perez, D. (2012). The youtube lens: Crowdsourced personality impressions and
audiovisual analysis of vlogs. IEEE Transactions on Multimedia, 15(1), 41-55.
11. Hurst, M. F. (2006). Temporal Text Mining. In AAAI Spring Symposium: Computational Approaches to
Analyzing Weblogs (pp. 73-77).
12. Cutting, D., Kupiec, J., Pedersen, J., & Sibun, P. (1992, March). A practical part-of-speech tagger. In Third
conference on applied natural language processing (pp. 133-140).
13. Zhang, Y., Jin, R., & Zhou, Z. H. (2010). Understanding bag-of-words model: a statistical
framework. International journal of machine learning and cybernetics, 1(1), 43-52.
14. Grishman, R., & Sundheim, B. M. (1996). Message understanding conference-6: A brief history.
In COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics.
15. Chung, C., & Pennebaker, J. W. (2007). The psychological functions of function words. Social
communication, 1, 343-359.
16. Brown, P. F., Della Pietra, V. J., Desouza, P. V., Lai, J. C., & Mercer, R. L. (1992). Class-based n-gram
models of natural language. Computational linguistics, 18(4), 467-480.
17. Chen, K., Zhang, Z., Long, J., & Zhang, H. (2016). Turning from TF-IDF to TF-IGM for term weighting
in text classification. Expert Systems with Applications, 66, 245-260.
18. Le, Q., & Mikolov, T. (2014, June). Distributed representations of sentences and documents.
In International conference on machine learning (pp. 1188-1196). PMLR.
19. Davison, A. (1984). Readability—Appraising text difficulty. Learning to read in American schools: Basal
readers and content texts, 121-139.
20. Kelledy, F., & Smeaton, A. F. (1997, April). Automatic phrase recognition and extraction from text.
In Proceedings of the 19th Annual BCS-IRSG Colloquium on IR Research 19 (pp. 1-9).
21. Wallach, H. M. (2006, June). Topic modeling: beyond bag-of-words. In Proceedings of the 23rd
international conference on Machine learning (pp. 977-984).
22. Lapponi, E., Read, J., & Øvrelid, L. (2012, December). Representing and resolving negation for sentiment
analysis. In 2012 ieee 12th international conference on data mining workshops (pp. 687-692). IEEE.
23. Howard, J., & Ruder, S. (2018). Universal language model fine-tuning for text classification. arXiv
preprint arXiv:1801.06146.
24. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of
words and phrases and their compositionality. Advances in neural information processing systems, 26.
25. Niu, Liqiang, Xinyu Dai, Jianbing Zhang, and Jiajun Chen. "Topic2Vec: learning distributed
representations of topics." In 2015 International conference on asian language processing (IALP), pp. 193-
196. IEEE, 2015.
26. Pennington, J., Socher, R., & Manning, C. D. (2014, October). Glove: Global vectors for word
representation. In Proceedings of the 2014 conference on empirical methods in natural language
processing (EMNLP) (pp. 1532-1543).

21
27. Young, J. C., & Rusli, A. (2019, August). Review and Visualization of Facebook's FastText Pretrained
Word Vector Model. In 2019 International Conference on Engineering, Science, and Industrial
Applications (ICESI) (pp. 1-6). IEEE.
28. Yao, D., Bi, J., Huang, J., & Zhu, J. (2015, July). A word distributed representation based framework for
large-scale short text classification. In 2015 international joint conference on neural networks (IJCNN)
(pp. 1-7). IEEE.
29. Matthew, E. (2018). Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee,
Luke Zettlemoyer. Deep contextualized word representations. In Proc. of NAACL.
30. Dey, R., & Salem, F. M. (2017, August). Gate-variants of gated recurrent unit (GRU) neural networks. In
2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS) (pp. 1597-1600).
IEEE.
31. Merity, S., Keskar, N. S., & Socher, R. (2017). Regularizing and optimizing LSTM language models.
arXiv preprint arXiv:1708.02182.
32. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:1810.04805.
33. Zhou, J., Zhang, Z., Zhao, H., & Zhang, S. (2019). Limit-bert: Linguistic informed multi-task bert. arXiv
preprint arXiv:1910.14296.
34. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., ... & Stoyanov, V. (2019). Roberta: A robustly
optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
35. Mairesse, F., Walker, M. A., Mehl, M. R., & Moore, R. K. (2007). Using linguistic cues for the automatic
recognition of personality in conversation and text. Journal of artificial intelligence research, 30, 457-
500.
36. Coltheart, M. (1981). The MRC psycholinguistic database. The Quarterly Journal of Experimental
Psychology Section A, 33(4), 497-505.
37. Whissell, C., Fournier, M., Pelland, R., Weir, D., & Makarec, K. (1986). A dictionary of affect in language:
IV. Reliability, validity, and applications. Perceptual and Motor Skills, 62(3), 875-888.
38. Moffitt, K., Giboney, J., Ehrhardt, E., Burgoon, J. K., & Nunamaker, J. F. (2010). Structured programming
for linguistic cue extraction. The Center for the Management of Information.
39. Stone, P. J., Dunphy, D. C., & Smith, M. S. (1966). The general inquirer: A computer approach to content
analysis.
40. Cambria, E., & Hussain, A. (2015). SenticNet. In Sentic Computing (pp. 23-71). Springer, Cham.
41. Nielsen, F. Å. (2011). A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. arXiv
preprint arXiv:1103.2903.
42. Mohammad, S. M., & Turney, P. D. (2013). Nrc emotion lexicon. National Research Council, Canada, 2.
43. Mohammad, S. (2018, July). Obtaining reliable human ratings of valence, arousal, and dominance for
20,000 English words. In Proceedings of the 56th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers) (pp. 174-184).
44. Miller, G. A. (1995). WordNet: a lexical database for English. Communications of the ACM, 38(11), 39-
41.
45. Havasi, C., Speer, R., & Alonso, J. (2007, September). ConceptNet 3: a flexible, multilingual semantic
network for common sense knowledge. In Recent advances in natural language processing (pp. 27-29).
Philadelphia, PA: John Benjamins.
46. Poria, S., Gelbukh, A., Hussain, A., Howard, N., Das, D., & Bandyopadhyay, S. (2013). Enhanced
SenticNet with affective labels for concept-based opinion mining. IEEE Intelligent Systems, 28(2), 31-38.
47. Searle, J. R. (1975). Indirect speech acts. In Speech acts (pp. 59-82). Brill.
48. Searle, J. R. (1976). A classification of illocutionary acts1. Language in society, 5(1), 1-23.
49. Walker, M., & Whittaker, S. (1995). Mixed initiative in dialogue: An investigation into discourse
segmentation. arXiv preprint cmp-lg/9504007.
50. McCrae, R. R., & John, O. P. (1992). An introduction to the five‐factor model and its applications. Journal
of personality, 60(2), 175-215.

22
51. Allport, G. W., & Odbert, H. S. (1936). Trait-names: A psycho-lexical study. Psychological monographs,
47(1), i.
52. Schwartz, S. H. (2007). Basic human values: Theory, measurement, and applications. Revue française de
sociologie, 47(4), 929.
53. Cattell, H. E., & Mead, A. D. (2008). The Sixteen Personality Factor Questionnaire (16PF).
54. Eysenck, H. J. (1982). Personality, genetics, and behavior: Selected papers.
55. Newman, J. (1981). Myers, Isabel Briggs. The Myers-Briggs Type Indicator. Palo Alto, Ca., Consulting
Psychologists Press, 1976. Myers, Isabel Briggs (with Peter B. Myers). Gifts Differing. Palo Alto, Ca.,
Consulting Psychologists Press, 1980.
56. Paulhus, D. L., & Williams, K. M. (2002). The dark triad of personality: Narcissism, Machiavellianism,
and psychopathy. Journal of research in personality, 36(6), 556-563.
57. Goldberg, L. R. (1992). The development of markers for the Big-Five factor structure. Psychological
assessment, 4(1), 26.
58. Ashton, M. C., & Lee, K. (2007). Empirical, theoretical, and practical advantages of the HEXACO model
of personality structure. Personality and social psychology review, 11(2), 150-166.
59. Gill, A. J., & Oberlander, J. (2002). Taking Care of the Linguistic Features of Extraversion. In Proceedings
of the Annual Meeting of the Cognitive Science Society (Vol. 24, No. 24).
60. Gosling, S. D., Ko, S. J., Mannarelli, T., & Morris, M. E. (2002). A room with a cue: personality judgments
based on offices and bedrooms. Journal of personality and social psychology, 82(3), 379.
61. Vazire, S., & Gosling, S. D. (2004). e-Perceptions: personality impressions based on personal websites.
Journal of personality and social psychology, 87(1), 123.
62. Mehl, M. R., Gosling, S. D., & Pennebaker, J. W. (2006). Personality in its natural habitat: manifestations
and implicit folk theories of personality in daily life. Journal of personality and social psychology, 90(5),
862.
63. Gosling, S. D., Gaddis, S., & Vazire, S. (2007). Personality impressions based on facebook profiles. Icwsm,
7, 1-4.
64. Tausczik, Y. R., & Pennebaker, J. W. (2010). The psychological meaning of words: LIWC and
computerized text analysis methods. Journal of language and social psychology, 29(1), 24-54.
65. Yarkoni, T. (2010). Personality in 100,000 words: A large-scale analysis of personality and word use
among bloggers. Journal of research in personality, 44(3), 363-373.
66. Holtgraves, T. (2011). Text messaging, personality, and the social context. Journal of research in
personality, 45(1), 92-99.
67. Iacobelli, F., Gill, A. J., Nowson, S., & Oberlander, J. (2011, October). Large scale personality
classification of bloggers. In international conference on affective computing and intelligent interaction
(pp. 568-577). Springer, Berlin, Heidelberg.
68. Qiu, L., Lin, H., Ramsay, J., & Yang, F. (2012). You are what you tweet: Personality expression and
perception on Twitter. Journal of Research in Personality, 46, 710–718. doi:10.1016/j.jrp.2012.08.008.
69. Schwartz, H. A., Eichstaedt, J. C., Kern, M. L., Dziurzynski, L., Ramones, S. M., Agrawal, M., ... & Ungar,
L. H. (2013). Personality, gender, and age in the language of social media: The open-vocabulary approach.
PloS one, 8(9), e73791.
70. Park, G., Schwartz, H. A., Eichstaedt, J. C., Kern, M. L., Kosinski, M., Stillwell, D. J., ... & Seligman, M.
E. (2015). Automatic personality assessment through social media language. Journal of personality and
social psychology, 108(6), 934.
71. Hassanein, M. M., Rady, S., Hussein, W., & Gharib, T. (2021). Extracting Relationships between Big Five
Model and Personality characteristics in Social Networks. International Journal of Intelligent Computing
and Information Sciences, 21(2), 41-49.
72. Štajner, S., & Yenikent, S. (2021, April). Why Is MBTI Personality Detection from Texts a Difficult Task?.
In Proceedings of the 16th Conference of the European Chapter of the Association for Computational
Linguistics: Main Volume (pp. 3580-3589).
73. Giorgi, S., Nguyen, K. L., Eichstaedt, J. C., Kern, M. L., Yaden, D. B., Kosinski, M., ... & Park, G. (2022).
Regional personality assessment through social media language. Journal of personality, 90(3), 405-425.

23
74. Celli, F., Lepri, B., Biel, J. I., Gatica-Perez, D., Riccardi, G., & Pianesi, F. (2014, November). The
workshop on computational personality recognition 2014. In Proceedings of the 22nd ACM international
conference on Multimedia (pp. 1245-1246).
75. Buchanan, T., & Smith, J. L. (1999). Using the Internet for psychological research: Personality testing on
the World Wide Web. British journal of Psychology, 90(1), 125-144.
76. Youyou, W., Kosinski, M., & Stillwell, D. (2015). Computer-based personality judgments are more
accurate than those made by humans. Proceedings of the National Academy of Sciences, 112(4), 1036-
1040.
77. Novikov, P., Mararitsa, L., & Nozdrachev, V. (2021). Inferred vs traditional personality assessment: are
we predicting the same thing?. arXiv e-prints, arXiv-2103.
78. Argamon, S., Dhawle, S., Koppel, M., & Pennebaker, J. W. (2005, June). Lexical predictors of personality
type. In Proceedings of the 2005 Joint Annual Meeting of the Interface and the Classification Society of
North America (pp. 1-16).
79. Mairesse, F., & Walker, M. (2006). Words mark the nerds: Computational models of personality
recognition through language. In Proceedings of the Annual Meeting of the Cognitive Science Society
(Vol. 28, No. 28).
80. Oberlander, J., & Nowson, S. (2006, July). Whose thumb is it anyway? Classifying author personality
from weblog text. In Proceedings of the COLING/ACL 2006 Main Conference Poster Sessions (pp. 627-
634).
81. Nowson, S., & Oberlander, J. (2007, March). Identifying more bloggers: Towards large scale personality
classification of personal weblogs. In Proceedings of the International Conference on Weblogs and Social.
82. Estival, D., Gaustad, T., Pham, S. B., Radford, W., & Hutchinson, B. (2007, September). Author profiling
for English emails. In Proceedings of the 10th Conference of the Pacific Association for Computational
Linguistics (Vol. 263, p. 272).
83. Golbeck, J., Robles, C., Edmondson, M., & Turner, K. (2011, October). Predicting personality from
twitter. In 2011 IEEE third international conference on privacy, security, risk and trust and 2011 IEEE
third international conference on social computing (pp. 149-156). IEEE.
84. Golbeck, J., Robles, C., & Turner, K. (2011). Predicting personality with social media. In CHI'11 extended
abstracts on human factors in computing systems (pp. 253-262).
85. Iacobelli, F., Gill, A. J., Nowson, S., & Oberlander, J. (2011, October). Large scale personality
classification of bloggers. In international conference on affective computing and intelligent interaction
(pp. 568-577). Springer, Berlin, Heidelberg.
86. Quercia, D., Kosinski, M., Stillwell, D., & Crowcroft, J. (2011, October). Our twitter profiles, our selves:
Predicting personality with twitter. In 2011 IEEE third international conference on privacy, security, risk
and trust and 2011 IEEE third international conference on social computing (pp. 180-185). IEEE.
87. Adali, S., & Golbeck, J. (2012, August). Predicting personality with social behavior. In 2012 IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining (pp. 302-309). IEEE.
88. Bai, S., Zhu, T., & Cheng, L. (2012). Big-five personality prediction based on user behaviours at social
network sites. arXiv preprint arXiv:1204.4809.
89. Kermanidis, K. L. (2012, May). Mining authors’ personality traits from Modern Greek spontaneous text.
In Proc. of Workshop on Corpora for Research on Emotion Sentiment & Social Signals, in conjunction
with LREC (pp. 90-93).
90. Wald, R., Khoshgoftaar, T., & Sumner, C. (2012, August). Machine prediction of personality from
Facebook profiles. In 2012 IEEE 13th International Conference on Information Reuse & Integration (IRI)
(pp. 109-115). IEEE.
91. Shen, J., Brdiczka, O., & Liu, J. (2013, June). Understanding email writers: Personality prediction from
email messages. In International conference on user modeling, adaptation, and personalization (pp. 318-
330). Springer, Berlin, Heidelberg.
92. Alam, F., Stepanov, E. A., & Riccardi, G. (2013). Personality traits recognition on social network-
facebook. In Proceedings of the International AAAI Conference on Web and Social Media (Vol. 7, No. 2,
pp. 6-9).

24
93. Verhoeven, B., Daelemans, W., & De Smedt, T. (2013, June). Ensemble methods for personality
recognition. In Seventh International AAAI Conference on Weblogs and Social Media.
94. Farnadi, G., Zoghbi, S., Moens, M. F., & De Cock, M. (2013, June). Recognising personality traits using
facebook status updates. In Proceedings of the International AAAI Conference on Web and Social Media
(Vol. 7, No. 1).
95. Tomlinson, M. T., Hinote, D., & Bracewell, D. B. (2013, June). Predicting conscientiousness through
semantic analysis of facebook posts. In Seventh International AAAI Conference on Weblogs and Social
Media.
96. Markovikj, D., Gievska, S., Kosinski, M., & Stillwell, D. J. (2013, June). Mining facebook data for
predictive personality modeling. In Seventh International AAAI Conference on weblogs and social media.
97. Iacobelli, F., & Culotta, A. (2013, June). Too neurotic, not too friendly: structured personality
classification on textual data. In Seventh International AAAI Conference on Weblogs and Social Media.
98. Appling, D., Briscoe, E., Hayes, H., & Mappus, R. (2013, June). Towards automated personality
identification using speech acts. In Proceedings of the International AAAI Conference on Web and Social
Media (Vol. 7, No. 1).
99. Mohammad, S., & Kiritchenko, S. (2013, June). Using nuances of emotion to identify personality. In
Seventh International AAAI Conference on Weblogs and Social Media.
100.Poria, S., Gelbukh, A., Agarwal, B., Cambria, E., & Howard, N. (2013, November). Common sense
knowledge based personality recognition from text. In Mexican International Conference on Artificial
Intelligence (pp. 484-496). Springer, Berlin, Heidelberg.
101.Zuo, X., Feng, B., Yao, Y., Zhang, T., Zhang, Q., Wang, M., & Zuo, W. (2013, September). A weighted
ML-KNN model for predicting users’ personality traits. In Proc. Int. Conf. Inf. Sci. Comput. Appl.(ISCA)
(pp. 345-350).
102.Gou, L., Zhou, M. X., & Yang, H. (2014, April). KnowMe and Sha reMe: understanding automatically
discovered personality traits from social media and user sharing preferences. In Proceedings of the SIGCHI
conference on human factors in computing systems (pp. 955-964).
103.Pratama, B. Y., & Sarno, R. (2015, November). Personality classification based on Twitter text using
Naive Bayes, KNN and SVM. In 2015 International Conference on Data and Software Engineering
(ICoDSE) (pp. 170-174). IEEE.
104.Arroju, M., Hassan, A., & Farnadi, G. (2015). Age, gender and personality recognition using tweets in a
multilingual setting. In 6th Conference and Labs of the Evaluation Forum (CLEF 2015): Experimental IR
meets multilinguality, multimodality, and interaction (Vol. 23, p. 31).
105.Poddar, S., Kattagoni, V., & Singh, N. (2015). Personality Mining from Biographical Data with the"
Adjectival Marker" Technique. In BD (pp. 39-47).
106.Lukito, L. C., Erwin, A., Purnama, J., & Danoekoesoemo, W. (2016, October). Social media user
personality classification using computational linguistic. In 2016 8th International Conference on
Information Technology and Electrical Engineering (ICITEE) (pp. 1-6). IEEE.
107.Pramodh, K. C., & Vijayalata, Y. (2016, October). Automatic personality recognition of authors using big
five factor model. In 2016 IEEE International Conference on Advances in Computer Applications
(ICACA) (pp. 32-37). IEEE.
108.Ong, V., Rahmanto, A. D. S., Williem, W., Suhartono, D., Nugroho, A. E., Andangsari, E. W., &
Suprayogi, M. N. (11 2017). Personality prediction based on Twitter information in Bahasa Indonesia.
367–372. doi:10.15439/2017F359.
109.Tandera, T., Suhartono, D., Wongso, R., & Prasetio, Y. L. (2017). Personality prediction system from
facebook users. Procedia computer science, 116, 604-611.
110.Ahmad, Z., Lutfi, S. L., Kushan, A. L., & Yixing, R. T. (2017). Personality Prediction of Malaysian
Facebook Users: Cultural Preferences and Features Variation. Advanced Science Letters, 23(8), 7900-
7903.
111.Yata, A., Kante, P., Sravani, T., & Malathi, B. (2018). Personality Recognition using Multi-Label
Classification. International Research Journal of Engineering and Technology (IRJET), 5(03).
112.Arjaria, S., Shrivastav, A., Rathore, A. S., & Tiwari, V. (2019). Personality trait identification for written
texts using mlnb. In Data, Engineering and Applications (pp. 131-137). Springer, Singapore.

25
113.Artissa, Y. B. N. D., Asror, I., & Faraby, S. A. (5 2019). Personality Classification based on Facebook
status text using Multinomial Naïve Bayes method. 1192. doi:10.1088/1742-6596/1192/1/012003
114.Ergu, İ., Işık, Z., & Yankayış, İ. Predicting Personality with Twitter Data and Machine Learning Models.
In 2019 Innovations in Intelligent Systems and Applications Conference (ASYU) (pp. 1-5). IEEE.
115.Rohit, G. V., Bharadwaj, K. R., Hemanth, R., Pruthvi, B., & Kumar, M. (2020, August). Machine
intelligence based personality prediction using social profile data. In 2020 Third International Conference
on Smart Systems and Inventive Technology (ICSSIT) (pp. 1003-1008). IEEE.
116.Ong, V., Rahmanto, A. D. S., Williem, W., Jeremy, N. H., Suhartono, D., & Andangsari, E. W. (2021).
Personality Modelling of Indonesian Twitter Users with XGBoost Based on the Five Factor Model.
International Journal of Intelligent Engineering and Systems, 14, 248–261.
doi:10.22266/ijies2021.0430.22.
117.Safitri, G., & Setiawan, E. B. (2 2022). Optimization Prediction of Big Five Personality in Twitter Users.
Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), τ. 6, σσ. 85–91. doi:10.29207/resti.v6i1.3529
118.Vu, X. S., Flekova, L., Jiang, L., & Gurevych, I. (2018, January). Lexical-semantic resources: yet powerful
resources for automatic personality classification. In Proceedings of the 9th Global Wordnet Conference
(pp. 172-181).d
119.Fernandes, B., González-Briones, A., Novais, P., Calafate, M., Analide, C., & Neves, J. (2020). An
adjective selection personality assessment method using gradient boosting machine learning. Processes,
8(5), 618.
120. Kalghatgi, M. P., Ramannavar, M., & Sidnal, N. S. (2015). A neural network approach to personality
prediction based on the big-five model. International Journal of Innovative Research in Advanced
Engineering (IJIRAE), 2(8), 56-63.
121.Su, M. H., Wu, C. H., & Zheng, Y. T. (2016). Exploiting turn-taking temporal evolution for personality
trait perception in dyadic conversations. IEEE/ACM Transactions on Audio, Speech, and Language
Processing, 24(4), 733-744.
122.Liu, F., Perez, J., & Nowson, S. (2016). A language-independent and compositional model for personality
trait recognition from short texts. arXiv preprint arXiv:1610.04345.
123.Xianyu, H., Xu, M., Wu, Z., & Cai, L. (2016, July). Heterogeneity-entropy based unsupervised feature
learning for personality prediction with cross-media data. In 2016 IEEE international conference on
multimedia and Expo (ICME) (pp. 1-6). IEEE.
124. Sun, X., Liu, B., Cao, J., Luo, J., & Shen, X. (2018, May). Who am I? Personality detection based on
deep learning for texts. In 2018 IEEE International Conference on Communications (ICC) (pp. 1-6). IEEE.
125.An, G., & Levitan, R. (2018, February). Lexical and Acoustic Deep Learning Model for Personality
Recognition. In INTERSPEECH (pp. 1761-1765).
126.Yılmaz, T., Ergil, A., & İlgen, B. (2019, October). Deep learning-based document modeling for personality
detection from Turkish Texts. In Proceedings of the Future Technologies Conference (pp. 729-736).
Springer, Cham.
127. Kazameini, A., Fatehi, S., Mehta, Y., Eetemadi, S., & Cambria, E. (2020). Personality trait detection
using bagged svm over bert word embedding ensembles. arXiv preprint arXiv:2010.01309.
128.Leonardi, S., Monti, D., Rizzo, G., & Morisio, M. (2020). Multilingual transformer-based personality traits
estimation. Information, 11(4), 179.
129.Xue, X., Feng, J., & Sun, X. (2021). Semantic-enhanced sequential modeling for personality trait
recognition from texts. Applied Intelligence, 51(11), 7705-7717.
130.El-Demerdash, K., El-Khoribi, R. A., Shoman, M. A. I., & Abdou, S. (2021). Deep learning based fusion
strategies for personality prediction. Egyptian Informatics Journal.
131.Christian, H., Suhartono, D., Chowanda, A., & Zamli, K. Z. (2021). Text based personality prediction
from multiple social media data sources using pre-trained language model and model averaging. Journal
of Big Data, 8(1), 1-20.
132.Jeremy, N. H., & Suhartono, D. (2021). Automatic personality prediction from Indonesian user on twitter
using word embedding and neural networks. Procedia Computer Science, 179, 416-422.

26
133.Mavis, G., Toroslu, I. H., & Karagoz, P. (10 2021). Personality Analysis Using Classification on Turkish
Tweets. International Journal of Cognitive Informatics and Natural Intelligence, 15, 1–18.
doi:10.4018/ijcini.287596
134.Kosan, M. A., Karacan, H., & Urgen, B. A. (2022). Predicting personality traits with semantic structures
and LSTM-based neural networks. Alexandria Engineering Journal, 61(10), 8007-8025.
135.Majumder, N., Poria, S., Gelbukh, A., & Cambria, E. (2017). Deep learning-based document modeling for
personality detection from text. IEEE Intelligent Systems, 32(2), 74-79.
136.Yu, J., & Markov, K. (2017, November). Deep learning based personality recognition from facebook status
updates. In 2017 IEEE 8th International Conference on Awareness Science and Technology (iCAST) (pp.
383-387). IEEE.
137.Giménez, M., Paredes, R., & Rosso, P. (2017, April). Personality recognition using convolutional neural
networks. In International Conference on Computational Linguistics and Intelligent Text Processing (pp.
313-323). Springer, Cham.
138.Xue, D., Wu, L., Hong, Z., Guo, S., Gao, L., Wu, Z., & Sun, J. (2018). Deep learning-based personality
recognition from text posts of online social networks. Applied Intelligence, 48(11), 4232-4246.
139.Rahman, M. A., Al Faisal, A., Khanam, T., Amjad, M., & Siddik, M. S. (2019, May). Personality detection
from text using convolutional neural network. In 2019 1st International Conference on Advances in Science,
Engineering and Robotics Technology (ICASERT) (pp. 1-6). IEEE.
140.Darliansyah, A., Naeem, M. A., Mirza, F., & Pears, R. (2019). SENTIPEDE: A Smart System for Sentiment-
based Personality Detection from Short Texts. Article in JOURNAL OF UNIVERSAL COMPUTER
SCIENCE, 25, 1323–1352. doi:10.3217/jucs-025-10-1323
141.Mehta, Y., Fatehi, S., Kazameini, A., Stachl, C., Cambria, E., & Eetemadi, S. (2020, November). Bottom-up
and top-down: Predicting personality with psycholinguistic and language model features. In 2020 IEEE
International Conference on Data Mining (ICDM) (pp. 1184-1189). IEEE.
142.Deilami, F. M., Sadr, H., & Nazari, M. (2022). Using Machine Learning Based Models for Personality
Recognition. arXiv preprint arXiv:2201.06248.
143.Deilami, F. M., Sadr, H., & Tarkhan, M. (2022). Contextualized Multidimensional Personality Recognition
using Combination of Deep Neural Network and Ensemble Learning. Neural Processing Letters.
doi:10.1007/s11063-022-10787-9.
144.Guan, Z., Wu, B., Wang, B., & Liu, H. (2020, July). Personality2vec: Network representation learning for
personality. In 2020 IEEE Fifth International Conference on Data Science in Cyberspace (DSC) (pp. 30-37).
IEEE.
145.Wang, Z., Wu, C. H., Li, Q. B., Yan, B., & Zheng, K. F. (2020). Encoding text information with graph
convolutional networks for personality recognition. Applied sciences, 10(12), 4081.
146.Wang, Y., Zheng, J., Li, Q., Wang, C., Zhang, H., & Gong, J. (6 2021). Xlnet-caps: Personality classification
from textual posts. Electronics (Switzerland), 10. doi:10.3390/electronics10111360
147.Ramezani, M., Feizi-Derakhshi, M. R., & Balafar, M. A. (2022). Knowledge Graph-Enabled Text-Based
Automatic Personality Prediction. arXiv preprint arXiv:2203.09103.
148.Jiang, H., Zhang, X., & Choi, J. D. (2020, April). Automatic text-based personality recognition on
monologues and multiparty dialogues using attentive networks and contextual embeddings (student abstract).
In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 10, pp. 13821-13822).
149.Li, Y., Kazemeini, A., Mehta, Y., & Cambria, E. (7 2022). Multitask learning for emotion and personality
traits detection. Neurocomputing, 493, 340–350. doi:10.1016/j.neucom.2022.04.049
150.Celli, F. (2012, March). Unsupervised personality recognition for social network sites. In Proc. of sixth
international conference on digital society (pp. 59-62).
151.Celli, F., & Rossi, L. (2012, April). The role of emotional stability in Twitter conversations. In Proceedings
of the workshop on semantic analysis in social media (pp. 10-17).
152.Liu and Zhu proposed use of stacked AutoEncoders for unsupervised learning of Linguistic Representation
Feature Vector (LRFV) based on SLIWC and FFT from Sina microblog. The features obtained were used to
train a Linear Regression model and results outperform the selected baselines.

27
153.Alsadhan, N., & Skillicorn, D. (2017, November). Estimating personality from social media posts. In 2017
IEEE international conference on data mining workshops (ICDMW) (pp. 350-356). IEEE.
154.Celli, F., & Lepri, B. (2018). Is big five better than MBTI? A personality computing challenge using Twitter
data. Computational Linguistics CLiC-it 2018, 93.
155.Lima, A. C., & de Castro, L. N. (2013, September). Multi-label semi-supervised classification applied to
personality prediction in Tweets. In 2013 BRICS Congress on Computational Intelligence and 11th
Brazilian Congress on Computational Intelligence (pp. 195-203). IEEE.
156.Lima, A. C. E., & De Castro, L. N. (2014). A multi-label, semi-supervised classification approach applied
to personality prediction in social media. Neural Networks, 58, 122-130.
157.Tighe, E. P., Ureta, J. C., Pollo, B. A. L., Cheng, C. K., & de Dios Bulos, R. (2016, July). Personality Trait
Classification of Essays with the Application of Feature Reduction. In SAAIP@ IJCAI (pp. 22-28).
158.Tighe, E., & Cheng, C. (2018, June). Modeling personality traits of Filipino twitter users. In Proceedings of
the Second Workshop on Computational Modeling of People’s Opinions, Personality, and Emotions in Social
Media (pp. 112-122).
159.Mao, Y., Zhang, D., Wu, C., Zheng, K., & Wang, X. (2018, December). Feature analysis and optimisation
for computational personality recognition. In 2018 IEEE 4th International Conference on Computer and
Communications (ICCC) (pp. 2410-2414). IEEE.
160.Adi, G. Y. N., Tandio, M. H., Ong, V., & Suhartono, D. (2018). Optimization for automatic personality
recognition on Twitter in Bahasa Indonesia. Procedia Computer Science, 135, 473-480.
161.Carducci, G., Rizzo, G., Monti, D., Palumbo, E., & Morisio, M. (2018). Twitpersonality: Computing
personality traits from tweets using word embeddings and supervised learning. Information, 9(5), 127.
162.Dos Santos, W. R., Ramos, R. M., & Paraboni, I. (2019). Computational personality recognition from
facebook text: psycholinguistic features, words and facets. New Review of Hypermedia and Multimedia,
25(4), 268-287.
163.Akrami, N., Fernquist, J., Isbister, T., Kaati, L., & Pelzer, B. (2019, December). Automatic extraction of
personality from text: Challenges and opportunities. In 2019 IEEE International Conference on Big Data (Big
Data) (pp. 3156-3164). IEEE.
164.Zheng, H., & Wu, C. (2019, February). Predicting personality using facebook status based on semi-
supervised learning. In Proceedings of the 2019 11th International Conference on Machine Learning and
Computing (pp. 59-64).
165.Tighe, E., Aran, O., & Cheng, C. (2020). Exploring Neural Network Approaches in Automatic Personality
Recognition of Filipino Twitter Users.
166.Pabón, F. O. L., & Arroyave, J. R. O. (12 2021). Automatic Personality Evaluation from Transliterations of
YouTube Vlogs Using Classical and State of the art Word Embeddings. Ingeniería e Investigación, τ. 42, σ.
e93803. doi:10.15446/ing.investig.93803.
167.Alamsyah, A., Putra, M. R. D., Fadhilah, D. D., Nurwianti, F., & Ningsih, E. (2018, May). Ontology
modeling approach for personality measurement based on social media activity. In 2018 6th International
Conference on Information and Communication Technology (ICoICT) (pp. 507-513). IEEE.
168.Alamsyah, A., Nurwiant, F., Rachman, M. F., Hudaya, C. S., Putra, R. P., Rifkyano, A. I., & Nurwianti, F.
(2019a). A Progress on the Personality Measurement Model using Ontology based on Social Media Text Cite
this paper Personalit y Measurement Design for Ont ology-based Plat form using Social Media Text Andry
Alamsyah Ont ology Modelling Approach for Personality Measurement based on Social Media Act ivit y A
Progress on the Personality Measurement Model using Ontology based on Social Media Text.
169.Alamsyah, A., Dudija, N., & Widiyanesti, S. (10 2021). New approach of measuring human personality traits
using ontology-based model from social media data. Information (Switzerland), 12.
doi:10.3390/info12100413.

28
Table 5. Performance Comparison of APRT (Essays)

Author Alg. Problem Approach Features Dataset Metric Result


Binary Function words
Arg05 SMO Top Down ESSAYS Acc 0.58
Classification (SFG)
Regression
LIWC ESSAYS
Mai07 SMO Classification Top Down Acc 0.57
MRC
Ranking
Vdd13 SVM Classification Hybrid n-grams Essays F1 score 0.73
SVM LIWC
Essays
Fa13 kNN Classification Hybrid Social Network F1 score 0.58
NB Temporal
Structured
Iac13 NB Bottom up n-grams Essays Acc 0.56
Classification
Mairesse
baseline
Sai13 SVM Classification Bottom up NRC Essays F1 Score 0.57
WordNet
Unigrams
SVM
Fir13 BLR Classification Bottom up Unigrams Essays F1 Score 0.563
mNB
Common Sense
Por13 SMO Classification Top Down LIWC Essays F1 Score 0.65
MRC
Weighted ML- Linguistic
Zuo13 Classification Bottom Up Essays F1 Score 0.65
kNN Emotional
LibSVM
Luk16 SMO Classification Top Down LIWC Essays Acc 0.61
LLR
POS tagging
Pra16 --- Correlation Bottom Up Essays F1 Score 0.665
Stop words
N-grams
Maj17 CNN Classification Hybrid Mairesse Essays Acc 0.63
Word2vec
Mairesse

SenticNet

Binary NRC-VAD
SVM Classification
Meh20 MLP Hybrid Essays Acc 0.606
LR Readability
Regression
BERT

Albert
RoBERTa
BERT

Binary
Kaz20 BB-SVM Bottom Up Word2Vec Essays Acc 0.59
Classification

ABCNN
Binary
Xue21 SEP-RNN Bottom Up GloVe Essays Acc 0.676
Classification
BERT Elmo Binary Bottom Up Language
Dem21 Essays Acc 0.81
ULMFiT Classification Embeddings

29
Table 6. Performance Comparison of APRT (Facebook)

Author Alg. Problem Approach Features Dataset Metric Result


LIWC
NB, SVM, LR, Binary
Tan17 Top Down SPLICE MyPersonality Acc 0.64
GB LDA Classification
SNA
SVM
Word 0.5
Car18 LASSO Regression Bottom Up MyPersonality MSE
Embedding’s 0.25
LR
DAL

Binary
An18 MLP-LSTM Hybrid Skipgram MyPersonality Acc 0.67
Classification

GloVe
Classification
Roh20-52 RF Bottom Up TF-IDF MyPersonality Acc 64.25
Regression
CNN Binary Word
Yu17 Bottom Up MyPersonality F1` 0.63
RNN Classification Embedding
MLP
LSTM Binary
Tan17 Bottom Up GloVe MyPersonality Acc 0.71
GRU Classification
CNN
HL 0.001

SA 0.998

Tokenization EBF 0.857

KNN Stemming MinF 0.998


Multi-Label
Yata18 NB Bottom Up MyPersonality
Classification
SVM Stop words MacF 0.998

TF-IDF AP 0.999

RL 0.001

OE 0.143
LIWC
Xue18 CNN Regression Hybrid MyPersonality MAE 0.3577
Semantic
LIWC
Binary DAL
AnLe18 MLP-LSTM Hybrid MyPersonality Acc 0.67
Classification Skipgram
Glove
Binary
Art19 MNB Bottom Up TF-IDF MyPersonality Acc 60.2
Classification
BERT
Binary Language
Dem21 Elmo Bottom Up MyPersonality Acc 0.81
Classification Embedding
ULMFiT
BERT
Binary Language
Chr21 XL-Net Bottom Up MyPersonality Acc 0.885
Classification Embedding
RoBERTa

30
Table 7. Performance Comparison of APRT (YouTube)

Author Alg. Problem Approach Features Dataset Metric Result


Word
Binary 0.55
Sun18 2CLSTM Bottom Up Embeddings YouTube Acc
Classification 0.6
LSG
Binary 0.802
Xue21 SEP-RNN Bottom Up Glove YouTube F1
Classification 0.676

Table 8. Performance Comparison of APRT (Twitter))

Author Alg. Problem Approach Features Dataset Metric Result


GP LIWC
Gol11a Regression Top Down Custom MAE 0.15
ZR Handcrafted
Binary
Gim17 CNN Bottom Up GloVe PAN-AM-15 RMSE 0.144
Classification
N-gram
Arroju15 ERCC Regression Bottom Up PAN-AM-15 Acc 68.5
LIWC
MNB

Binary
Prat15 KNN Bottom Up TF-IDF PAN-AM-15 Acc 0.65
Classification

SVM
Sko16 RF Regression Hybrid LIWC Custom RMSE 0.73
Binary
Liu16 Bi-RNN Bottom Up Character-Word PAN-AM-15 RMSE 0.14
Classification
Gim17 CNN Regression Bottom Up GloVe PAN-AM-15 RMSE 0.144
SGD

XGB N-Grams
Adi18 Regression Bottom Up Custom AUC 0.992
SL TF-IDF

SVM
SVM

LASSO
Word
Car18 Regression Bottom Up Custom MSE 0.25
Embedding
LR

SVM
XL-Net Language
Chr21 Classification Bottom Up Custom Acc 0.885
RoBERT a Embedding

31

You might also like