Predicting Personality from Twitter
Jennifer Golbeck
∗
, Cristina Robles
∗
, Michon Edmondson
∗
, and Karen Turner
∗
∗
University of Maryland;{jgolbeck,crobles,michonk8,kturner}@umd.edu
Abstract—Social media is a place where users present them-
selves to the world, revealing personal details and insights into
their lives. We are beginning to understand how some of this
information can be utilized to improve the users’ experiences with
interfaces and with one another. In this paper, we are interested
in the personality of users. Personality has been shown to be
relevant to many types of interactions; it has been shown to be
useful in predicting job satisfaction, professional and romantic
relationship success, and even preference for different interfaces.
Until now, to accurately gauge users’ personalities, they needed to
take a personality test. This made it impractical to use personality
analysis in many social media domains. In this paper, we present
a method by which a user’s personality can be accurately
predicted through the publicly available information on their
Twitter profile. We will describe the type of data collected, our
methods of analysis, and the machine learning techniques that
allow us to successfully predict personality. We then discuss the
implications this has for social media design, interface design,
and broader domains.
Index Terms—personality, social media
I. INTRODUCTION
Social networking on the web has grown dramatically over
the last decade. In January 2005, a survey of social networking
websites estimated that among all sites on the web there
were roughly 115 million members [14]. Just over five years
later, Twitter alone has exceeded 200 million members. In the
process of creating social networking profiles, users reveal a
lot about themselves both in what they share and how they
say it. Through self-description, status updates, photos, and
interests, much of a user’s personality comes out through their
profile.
For decades, psychology researchers have worked to un-
derstand personality in a systematic way. After extensive
work to develop and validate a widely accepted personality
model, researchers have shown connections between general
personality traits and many types of behavior. Relationships
have been discovered between personality and psychological
disorders [42], job performance [4] and satisfaction [24], and
even romantic success [46].
This paper attempts to bridge the gap between social media
and personality research by using the information people
reveal in their online profiles. Our core research question asks
whether social media profiles can predict personality traits. If
so, then there is an opportunity to integrate the many results
on the implications of personality factors and behavior into the
users’ online experiences and to use social media profiles as
a source of information to better understand individuals. For
example, the friend suggestion system could be tailored to a
user based on whether they are more introverted or extraverted.
Previous work has shown that the information in users’
Facebook profiles is reflective of their actual personalities, not
an “idealized” version of themselves [3]. We expect Twitter
to have similar characteristics, and that plus a broad user base
of 200 million people makes it an ideal platform for study.
We administered the Big Five Personality Inventory to 279
subjects through a Twitter application. In the process, we
gathered their 2000 most recent public Twitter posts (tweets).
This was aggregated, quantified, and passed through a text
analysis tool to obtain a feature set. Using these statistics, we
were able to develop a model that can predict personality on
each of the five personality factors to within between 11% and
18% of the actual values.
The ability to predict personality has implications in many
areas. Existing research has shown connections between per-
sonality traits and success in both professional and personal re-
lationships. Social media tools that seek to support these rela-
tionships could benefit from personality insights. Additionally,
previous work on personality and interfaces showed that users
are more receptive to and have greater trust in interfaces and
information that is presented from the perspective of their own
personality features (i.e. introverts prefer messages presented
from an introvert’s perspective). If a user’s personality can be
predicted from their social media profile, online marketing and
applications can use this to personalize their message and its
presentation.
We begin by presenting background on the Big Five Per-
sonality index and related work on personality and social
media. We then present our experimental setup and methods
for analyzing and quantifying Twitter profile information. To
understand the relationship between personality and social
media profiles, we present results on correlations between each
profile feature and personality factor. Based on this, we de-
scribe the machine learning techniques used for classification
and show how we achieve large and significant improvements
over baseline classification on each personality factor. We
conclude with a discussion of the implications that this work
has for social media websites and for organizations that may
utilize social media to better understand the people with whom
they interact.
II. B
ACKGROUND AND RELATED WORK
A. The Big Five Personality Inventory
The “Big Five” model of personality dimensions has
emerged as one of the most well-researched and well-regarded
measures of personality structure in recent years. The models
five domains of personality, Openness, Conscientiousness,
extroversion, Ageeableness, and Neuroticism, were conceived
2011 IEEE International Conference on Privacy, Security, Risk, and Trust, and IEEE International Conference on Social Computing
978-0-7695-4578-3/11 $26.00 © 2011 IEEE
DOI
149
Fig. 1: A person has scores for each of the five personality
factors. Together, the five factors represent an individual’s
personality.
by Tupes and Christal [47] as the fundamental traits that
emerged from analyses of previous personality tests [29].
McCrae & Costa [28] and John [21] continued five-factor
model research and consistently found generality across age,
gender, and cultural lines [29]. Additional research has proved
that different tests, languages, and methods of analysis do
not alter the models validity [29], [10], [21], [27]. Such
extensive research has led to many psychologists to accept the
Big Five as the current definitive model of personality [43],
[34]. It should be noted that the models dependence on trait
terms indicates that the Big Five traits are based on a lexical
approach to personality measurement [43], [9], [10], [16]. The
Big Five traits are characterized by the following:
• Openness to Experience: curious, intelligent, imaginative.
High scorers tend to be artistic and sophisticated in taste
and appreciate diverse views, ideas, and experiences.
• Conscientiousness: responsible, organized, persevering.
Conscientious individuals are extremely reliable and tend
to be high achievers, hard workers, and planners.
• extroversion: outgoing, amicable, assertive. Friendly and
energetic, extroverts draw inspiration from social situa-
tions.
• Agreeableness: cooperative, helpful, nurturing. People
who score high in agreeableness are peace-keepers who
are generally optimistic and trusting of others.
• Neuroticism: anxious, insecure, sensitive. Neurotics are
moody, tense, and easily tipped into experiencing negative
emotions.
B. Applications of the Big Five
Much work has been done with personality as it relates to
our lives and the choices we make. In terms of relationships
with others, many relationships have been identified. Personal-
ity type is linked to whom users choose to friend on Facebook.
[45] found that extraversion, agreeableness, and openness all
correlated with friendship selection. Personality features have
also been tied to many aspects of romantic relationships,
including partner choice, level of attachment and success
[8], [46]. In terms of interpersonal conflict, studies have
associated Big Five traits with coping responses, vengefulness,
and rumination [32],[5]. Social relationships aside, personality
also relates to preferences. Rentfrow and Gosling [39] is one
of many studies that found that personality is a factor that
relates to the music an individual prefers to listen to. Jost et
al. [23] also found that the personality type of an individual
was able to predict whether they would be more likely to
vote for McCain or Obama in 2008. Research has also found
personality differences between self-professed “dog people”
and “cat people” [37], [17]. Within the context of marketing
and advertising, Big Five personality traits have been shown to
accurately predict a consumers preference for national brands
or independent brands [48]. Studies like this show a promising
future for the integration of personality analysis and consumer
profiling.
Many studies have demonstrated the usefulness of person-
ality profiles within the professional context. Hodgkinson and
Ford [20] found that personality traits affect job performance
and satisfaction, and Barrick and Mount [4] correlated specific
traits with occupational choices and proficiency. Big Five
dimensions have proved valid predictors for team performance
[31], counterproductive behaviors [41], and entrepreneurial
status [49], among many other factors. [6] also revealed rela-
tionships between personality and behavior among managers,
and Barrick and Mount found recurring personality profiles
among both high-autonomy and low-autonomy positions in
the workforce [5].
In the space of Human-Computer Interaction, one of the
pioneering studies on the connection between personality and
interface preference was presented in [30]. Users listened to
audio readings of five book reviews which were written from
the perspective of introverts vs. extroverts. Subjects were able
to identify the personality differences between the reviews and
showed an attraction to those which were closest to their own
personality type. When the personality type matched, subjects
were even more likely to buy the book being reviewed.
This work was extended into ideas of Graphical User
Interface design in [25]. Different GUIs were developed to
represent introverted vs. extroverted personality types. As in
[30], subjects could identify the personality differences and
preferred the interface that matched their own personality type.
C. Personality Research and Social Media
To the best of our knowledge, our work is among the first to
look at the relationship between profile information provided
in social networks and personality traits. However, there have
150
Fig. 2: Average scores on each personality trait shown with
standard deviation bars.
been a few previous studies on how personality relates to social
networking more generally.
It has been shown in [40] that extroversion and consci-
entiousness positively correlate with the perceived ease of
use of social media websites. extroversion was also shown
to have a positive correlation with perceived usefulness of
such sites. Not surprisingly, extroversion was also shown to
correlate with the size of a user’s social network in several
studies [2], [44], [45]. There have also been mixed results for
other personality traits. Work in [45] showed that individuals
with high agreeableness scores were selected more often as
friends and that people tended to choose friends with similar
agreeableness, extroversion, and openness scores. This was
not repeated in [44], but a correlation between openness and
number of friends.
III. D
ATA COLLECTION
We created a Twitter application with two functions. First, it
administered a 45-question version of the Big Five Personality
Inventory [22] to users. Subjects would take the test and for
each, we collected the most recent 2,000 tweets from the user
(or all tweets if they had less than 2,000).
We had fifty subjects who were recruited through posts on
Twitter, Facebook, and relevant mailing lists. Twitter does not
collect or release demographic information about its users and,
since we would have no general baseline for comparison, we
did not collect it for our subjects.
Average scores on the personality test are shown in figure
2 and in table I.
For each user, we began by collecting a simple set of
statistics about their accounts and their tweets. These included
the following:
• Number of followers (people following the user)
• Number of following (people the user follows)
• Density of the social network
• Number of “@mentions” - An @mention is when a user
mentions the name of another user by adding an @ to
the front of the username, as is convention on Twitter
• Number of replies - Using the Twitter API, we could see
how many of the user’s tweets were direct replies to other
user’s tweets.
• Number of hashtags - Hashtags (e.g. #cscw2012) are a
way of tagging a tweet to be part of a given topic or
event. They are also used in “games” where users come
up with tweets to go with a tag (e.g. #firstdraftmovielines
is used with altered first movie lines created by users).
• Number of links
• Words per tweet
For the number of @mentions, replies, hashtags, and links,
we used the raw numbers and the average per tweet.
Our primary analysis was a basic processing of the text of
the tweets. This was done by merging the collected tweets for
a given user into a single “document” and analyzing that.
Previous research has shown that linguistic features can be
used to predict personality traits [26], [36]. . Data collected in
[36] was used in both studies. They had three separate sources
of text, ranging from an average of 1,770 words to over 5,000
words per person.
There is potential to apply these linguistic analysis methods
to help predict personality by analyzing a person’s tweets.
However, the text samples used in earlier studies are much
larger than are available to us through any twitter posting.
Aggregating many tweets from a user gives more information,
but as a series of disconnected statements rather than a
coherent document as was used in other studies. Thus, it is
unclear if Twitter text will be as connected to personality
as was the case in other work. Tweets are much different
sources of text. Each one is limited to 140 characters, and
a compilation of tweets from a given user is more a stream
of disjointed thoughts than a coherent narrative as is found in
the text used in previous personality studies. Thus, it was not
entirely clear whether tweets would be a useful source of data
for this type of analysis.
There were an average of 1914 words per user, and the
distribution is shown in figure 3. The number of words ranged
from 50 to 5724. These came from an average of 142.2 tweets,
with one using having a maximum of 350 tweets and another
with a minimum of 4.
Following the methods used in [26], [36] as well as other
studies of social media behavior, such as [13], we utilized
two main tools to analyze the content of users’ tweets. The
first is that Linguistic Inquiry and Word Count (LIWC) tool
[35]. LIWC produces statistics on 81 different features of
text in five categories. These include Standard Counts (word
count, words longer than six letters, number of prepositions,
etc.), Psychological Processes (emotional, cognitive, sensory,
and social processes), Relativity (words about time, the past,
the future), Personal Concerns (such as occupation, financial
issues, health), and Other dimensions (counts of various types
of punctuation, swear words). We excluded the Standard
Counts and Other Dimension features to eliminate what is
likely to be noise on the type of text we have. The exceptions
are that we included word count, words per sentence, and
swear word counts since these reflect verbosity and tone of
151
Fig. 3: Number of words per user.
the user. For the other three categories, the values are given
as the percentage of words in the input that match words in a
given category. For example, it counts the number of “social”
words such as “talk”, “us”, and “friend”, or “anxiety” words
like “nervous”, “afraid”, and “tense”. Correlations between
these features and personality traits (e.g. anxiety words and
neuroticism scores) would not be surprising. This produced
79 text features.
In addition, we ran the text again the MRC Psycholinguistic
Database, a list of over 150,000 words with linguistic and
psycholinguistic features of each word. These include: Kucera-
Francis written frequency, number of categories, and num-
ber of samples; Brown verbal frequency; Familiarity rating;
Meaningfulness via Colorado norms and via Paivio Norms;
Concreteness; age of acquisition; Thorndike-Lorge written
frequency; and the number of letters, phonemes, and syllables.
We computed the average non-zero score for each feature over
all the words from each user.
In addition, we performed a word by word sentiment anal-
ysis of each user’s tweets. Using the General Inquirer dataset
[1], which provides a hand annotated dictionary that assigns
words sentiment values on a -1 to +1 scale, we computed a
score for each user that was the average sentiment score for
all words used in their list of tweets.
IV. P
ERSONALITY AND TWITTER BEHAVIOR
CORRELATIONS
We began by running a Pearson correlation analysis between
subjects’ personality scores and each of the features obtained
from analyzing their tweets and public account data. These are
shown in table II.There are a number of significant correlations
here, however none of them are strong enough to directly
predict any personality trait. Correlations that were statistically
significant for p<0.05 are bolded.
Many of the correlations make intuitive sense. For example,
conscientiousness is negatively correlated with words about
death (e.g. “bury”, “coffin”, “kill”) and with negative emotions
and sadness, suggesting conscientious people tend to talk less
about unhappy subjects. At the same time, the trait is positively
Fig. 4: Features used for predicting personality.
TABLE I: Average scores on each personality factor on a
normalized 0-1 scale
Agree. Consc. Extra. Neuro. Open.
Average 0.697 0.617 0.586 0.428 0.755
Stdev 0.162 0.176 0.190 0.224 0.147
correlated with the use of “you”, indicating the same people
tend to talk about or to others. Agreeable people also tend to
use “you” a lot, but are less likely to talk about achievements
and money.
However, there are not such intuitive explanations for other
correlations. For example, the number of parentheses used is
negatively correlated with both extraversion and openness. It
is unclear why this is the case, or if these are perhaps falsely
significant data points. However, since our focus in this paper
is on predicting personality rather than on focusing on any
particular correlation, we do not assign much weight to any
of these connections. A space of future work would be to
probe more deeply into these correlations over a larger data
set.
V. P
REDICTING PERSONALITY
To predict the score of a given personality feature, we
performed a regression analysis in Weka [18]. We used two
regression algorithms: Gaussian Process and ZeroR, each with
a 10-fold cross-validation with 10 iterations. Two algorithms
had similar performance over the personality features. Results
are shown in table III.
We found that Openness was the easiest to compute and
neuroticism was the most difficult, consistent with the results
152
TABLE II: Pearson correlation values between feature scores and personality scores. Significant correlations are shown in bold
for p<0.05. Only features that correlate significantly with at least one personality trait are shown.
Language Feature Examples Extro. Agree. Consc. Neuro. Open.
“You” (you, your, thou) 0.068 0.364 0.252 -0.212 -0.020
Articles (a, an, the) -0.039 -0.139 -0.071 -0.154 0.396
Auxiliary Verbs (am, will, have) 0.033 0.042 -0.284 0.017 0.045
Future Tense (will, gonna) 0.227 -0.100 -0.286 0.118 0.142
Negations (no, not, never) -0.020 0.048 -0.374 0.081 0.040
Quantifiers (few, many, much) -0.002 -0.057 -0.089 -0.051 0.238
Social Processes (mate, talk, they, child) 0.262 0.156 0.168 -0.141 0.084
Family (daughter, husband, aunt) 0.338 0.020 -0.126 0.096 0.215
Humans (adult, baby, boy) 0.204 -0.011 0.055 -0.113 0.251
Negative Emotions (hurt, ugly, nasty) 0.054 -0.111 -0.268 0.120 0.010
Sadness (crying, grief, sad) 0.154 -0.203 -0.253 0.230 -0.111
Cognitive Mechanisms (cause, know, ought) -0.008 -0.089 -0.244 0.025 0.140
Causation (because, effect, hence) 0.224 -0.258 -0.155 -0.004 0.264
Discrepancy (should, would, could) 0.227 -0.055 -0.292 0.187 0.103
Certainty (always, never) 0.112 -0.117 -0.069 -0.074 0.347
Perceptual Processes
Hearing (listen, hearing) 0.042 -0.041 0.014 0.335 -0.084
Feeling (feels, touch) 0.097 -0.127 -0.236 0.244 0.005
Biological Processes (eat, blood, pain) -0.066 0.206 0.005 0.057 -0.239
Body (cheek, hands, spit) 0.031 0.083 -0.079 0.122 -0.299
Health (clinic, flu, pill) -0.277 0.164 0.059 -0.012 -0.004
Ingestion (dish, eat, pizza) -0.105 0.247 0.013 -0.058 -0.202
Work (job, majors, xerox) 0.231 -0.096 0.330 -0.125 0.426
Achievement (earn, hero, win) -0.005 -0.240 -0.198 -0.070 0.008
Money (audit, cash, owe) -0.063 -0.259 0.099 -0.074 0.222
Religion (altar, church, mosque) -0.152 -0.151 -0.025 0.383 -0.073
Death (bury, coffin, kill) -0.001 0.064 -0.332 -0.054 0.120
Fillers (blah, imean, youknow) 0.099 -0.186 -0.272 0.080 0.120
Punctuation
Commas 0.148 0.080 -0.24 0.155 0.170
Colons -0.216 -0.153 0.322 -0.015 -0.142
Question Marks 0.263 -0.050 0.024 0.153 -0.114
Exclamation Marks -0.021 -0.025 0.260 0.317 -0.295
Parentheses -0.254 -0.048 -0.084 0.133 -0.302
Non-LIWC Features
GI Sentiment 0.177 -0.130 -0.084 -0.197 0.268
Number of Hashtags 0.066 -0.044 -0.030 -0.217 -0.268
Words per tweet 0.285 -0.065 -0.144 0.031 0.200
Links per tweet -0.061 -0.081 0.256 -0.054 0.064
153