The last decade has witnessed significant work in personality mining from lexical cues in social media data. Not much work has yet been undertaken in extracting these lexical cues from biographical data populating social media. Most of this work involves a large crowd of researchers leveraging dictionary-based approaches such as LIWC (which primarily focus on function words). By means of this paper we intend to introduce a novel method of personality mining from social media data called “Adjectival-marker Technique”. This method involves extracting lexical features from descriptive texts (e.g. biographical data) to train a learning model, so as to predict the respective personality traits of the subject. Conceptually, it draws heavily from the last 78 years of work in lexical psychology and the Big Five personality test. However, it is not only a computational variant of the primordial theories of lexical psychology, but is also competent in conferring a substantial accuracy of personality prediction, matching that obtained by psychometric tests. In this study, we propose a variant of the Lexical Hypothesis from psychology. This modified hypothesis is validated by the computational results of personality prediction achieved by the Adjectival Marker Technique discussed below. The paper also discusses some insights illustrating the coherence of people's judgments about the subject's personality (virtual personality). The average accuracy (i.e. matching that achieved by psychometric tests for Big 5) for prediction approximated to Extraversion - 82.82% Agreeableness - 89.62%, Conscientiousness 92.48% and Imaginativeness/Intellect - 81.67%.
Social Media has become the most abundantly used means
of communicating and propagating information online.
Most information here is extensively descriptive of the
users who channel themselves through it. It is not only the
user who gives away information about himself
There have been several personality models (The Big
Three, The Big Five, The Alternative Five, etc.) that claim
to encapsulate the traits that need to be summoned so as
to effectively predict user personalities from social media
data. However, out of all these models, the most robust
and tested model, which has been consistent for the last
few decades, is the Big Five (Big5) model
This research builds on the confluence of two major
domains, the primordial theories of the lexical hypothesis and
the recent computational techniques of data modeling.
Allport's personality trait names
The theories of psychology were influenced by various revolutionary concepts, for instance, “trait” - a theoretical construct which describes a basic dimension of a person's personality (Allport, 1937). The idea of trait gave birth to the “Lexical Hypothesis of Psychology”. The initial direction of this paper was solely governed by this exact hypothesis (worked upon by Klages, 1926/1932; Cattell, 1943; Norman, 1963; Goldberg, 1982)
“Those individual differences that are most salient and socially relevant in people's lives will eventually become encoded into their language; the more important such a difference, the more likely is it to become expressed as a single
word.”
The Lexical Hypothesis has been used in its entirety in
author's personality prediction systems, like the one for Greek
Language described by Kermanidis et al, (2012). Motivated
by the same inspiration, we too expected to extract author's
personality traits from the text they wrote. This involved
mobilizing huge datasets of web blogs and essays and
extracting “names” from them to determine the author's
personalities. However, by the course of our study, we found
out that this was not as effective as the initial hypothesis
proposed (Goldberg et al, 1982). The average accuracy of
the initial experimentation was less than 50%, which was
as good as a randomly predicted personality set.
Thus, we propose a modification of the Lexical
Hypothesis in psychology which suggests that the personality of a
person is predicted based on cumulative judgements of
various authors about him/her. These judgements are
indicative of the respective traits of the person described along
the lines of the Big5 personality Model. The “Adjectival
Marker ” Technique helps us unravel these judgements, and
is derived from the adjectival markers of Big5 personality
traits as discussed by Goldberg &
“Those individual differences that are most salient and socially relevant in people's lives will eventually (over time) become encoded into their language as well as that of people who describe them (via the knowledge they have of them, these people could be peers, associates, friends, family members, followers etc.); the more important such a difference, the more likely is it to become expressed as a single word”.
The “Adjectival Marker Technique” introduced in this paper is most accurate when it is used to analyze the personality of the subject who the social media resource is descriptive of and not the author himself. We also inferred an interesting observation that suggested that the views of different people describing the subject are coherent amongst themselves and also with the results of the psychometric tests. The average accuracy of the traits, based on the proposed hypothesis, for a series of data spread temporally and spatially (as compared to the results obtained by psychometric tests) in social media came out as discussed below.
We begin by presenting a brief background on the Lexical Psychology theories of personality determination and related work on personality in conjuction with social media in Section 2. We then present our dataset in Section 3 & 4 and methodology for analyzing, quantifying and modelling biographical data content for 574 personalities in Section 5. The study proceeds on to describe the adjectival features used along with the machine learning techniques for classification and demonstrate significant improvements that the model was able to achieve over baseline classification on each personality factor. In subsequent sections, the paper presents the results in Section 6 and analysis of the study, and discusses the methods we incorporated which were instrumental in escalating the accuracy of the model for each of the traits discussed earlier in Section 7. We finally wrap up the paper with brief discussions about the future work, sparked by this study in Section 8.
2.
The last few years have witnessed a considerable escalation in studies which are directed at mining user personalities from social media data. Those which are related to this work can be mined in mainly 2 sections. (i) Studies which are based on lexical cues to mine author's personality, (ii) Studies which have used social media based features to study the personality of the user.
The former section includes work by Tausczik and
Pennebaker (2010) wherein they mined author personality via
LIWC (Linguistic Inquiry and word count) approaches.
Another such study used linguistic features such as
function words, deictics, appraisal expressions and modal verbs
to classify 2 of the Big Five traits namely neuroticism and
extraversion
Personality Analysis in Social Media Analysis is a recently
observed phenomenon. Herein, some substantial work was
done by Goldbeck et al, (2011) wherein the authors
predicted the personality of 279 users from Facebook, using
either linguistic or social network specific features.
The data collected as a part of this study was by means of a Python-based crawler. We first used a simple web crawler to get a list of web-pages with the name of the respective “person” as the argument keyword to the crawler. These web pages were then filtered based on their meta-tags. To boost true positives, we only considered the pages which specified their content as “biographical” in the meta-tag descriptors. This resulted in mobilization of few Wikipedia resources, blog mentions and majorly some very descriptive biographical websites. We then manually cleaned the noisy data to assure entity disambiguation and irrelevant mentions. The same has been illustrated by means of Figure 1.
The Jungian Personality functions of 574 personalities were
extracted from the resource for eventual evaluation.2 Since
this was one of the most authentic reserves we found
consisting of personality listings (so as to evaluate the ones our
model predicts) we found it the most effective to be used
for evaluating our own model. The “Adjectival Markers”
that the paper is based on (as described below) are a proven
indicator to reflect the Big5 traits of personality. Thus, to
evaluate our computed predictive model via personalities
for the respective subjects by an exclusively listed source,
we scaled the Jungian Typology type to the closest traits of
the Big5 using correlation factors as shown in Table 2
Correlation
Emotional Stability Conscientiousness Accommodation / Agreeableness
Intellect
Sensing/Intuition
As illustrated, 4 final personality traits were scaled (each of which had medium to high correlation with the MBTI 2http://www.celebritytypes.com, wherein extensive cognitive functions have been used to derive the psychology of the given personalities.
Semi-Correlating Descriptions Jung/MBTI/Kiersey Global 5
INTP
INFJ
INTJ
ISTJ
ISFJ
ISTP
ISFP
ENFP
ENTP
ENFJ
ENTJ
ESTJ
ESFJ
ESTP
ESFP
The adjectives mined from the biographical data were
refined to extract the adjectival markers i.e. specific
adjectives descriptive of the subject of the biographical data.
These adjectival markers were used as features in the final
LASSO logistic regression model. The adjectival markers
extracted are based on the work of
Biographical data was mined for 574 personalities from online resources as discussed in the former Section 3.1. This data was divided into 2 categories. Testing data and Training data. Users with no substantial data (>100 words were discarded for the analysis as of now). The data mining undertaken for acquiring these datasets is spread across various social media resources including Wikipedia articles, blog posts, social Q & A sites and community media sites (sharing biographical book excerpts, for building datasets of word count >10,000)
The training data set, used to mine adjectival markers, comprised of biographic data content of 283 personalities. The word count of the dataset ranged from 500 - 10,000 words. The ratio of the number of adjectives to the total number of words in the dataset ranged from 0 to 0.005.
This data content was mined by means of a Python-based web crawler, which parsed biographic websites, Wikipedia, and social media mentions.
The testing dataset comprised of biographic data content of a different set of 291 personalities than the ones used for training. These were mined from the social media reserves like Wikimedia articles, blog posts about the respective personalities, social Q& A sites etc. The word count and the number of adjectives to the total number of words ratio ranged from 100 10,000 words and 0.0001 to 0.003 respectively.
Adventurous
II
The training data (283 users) was mined for
adjectival markers according to Saucier's adjectival marker list
Since there was only single source where traits of major personalities are classified (i.e. celebritytypes.com) we used it to evaluate our model. We used the remaining 291 personalities for evaluation of the model. The testing biographical data was mined for adjectival trait markers and their respective traits were predicted. The results of this evaluation have been discussed elaborately in the next section. Figure 2, which can be found above, is also illustrative of the procedure define above.
The results by the above illustrated method are elaborated
in this section. The average accuracies compared to the
personalities obtained via psychometric tests (discussed in
more detail in the following section) for considered four of
the Big 5 traits were: Extraversion - 82.82% Agreeableness
- 89.62%, Conscientiousness - 92.48% and
Imaginativeness/Intellect - 81.67%. These readings do not necessarily
demonstrate the prediction accuracy of the innate
personality of a person but match that predicted by the psychometric
tests with the given accuracies. They are also in league with
few other techniques predicting the same for instance, the
work of
Thus, this paper proposes a technique which illustrates mainfold elevation in the overall accuracy of personality prediction (as indicated by psychometric tests) via social media.
The results obtained illustrate that this method is competent for predicting the personalities of a person in coherence with other people's judgments about him/her. It gives substantial accuracies in the prediction of a person's personality matching with those obtained via psychometric tests. As an essential part of this study, we have also attempted to capture the variation in accuracy with the change in various factors, namely, word count of the corpus, and the ratio of the number of adjectives to the total number of words.3 These are mainly intended to explore a threshold for word count and the adjective distribution (for the given 3Please note that the accuracies discussed here are the accuracy of the prediction as evaluated by the results via psychometric tests for Big 5 and should not be confused with accuracies used for predicting the baseline of the universal personality of a person. technique) in the document set so as to get substantial results from the Adjectival Marker Technique. The following deductions can be made respective to each trait: 7.1.
Few collective observations can be drawn from the gathered data. As indicated in Figure 4, the accuracy in predicting the traits increases with an increase in the data word count. We also compared the accuracy results in predicting the respective traits on the basis of varying distribution of adjectives in the training dataset (Figure 5). The accuracy in predicting the traits is relatively low when the ratio of the AC/TWC is low and increases with a subsequent increase in the AC/TWC ratio. 7.2.
The accuracy in predicting Agreeableness is relatively low (73.33%) for data with word count < 5000 words, and escalates up to 99.11% for big data reserves (>20,000 words). We also compared the accuracy results of predicting “Agreeableness” on the basis of varying distribution of adjectives in the training dataset.
The prediction of the “Agreeableness” trait is relatively low when the ratio of the adjectival count versus total word count is low. It illustrates an accuracy of 84.00% when the ratio is less than 0.001, improving to 94.18% when the ratio is between 0.001-0.002. Finally it escalates to 95.62% when increased to be greater than 0.003 (Figure 5). As expected there is a consistent increase in accuracy with increase in word count and the ratio AC/TWC. 7.3.
The accuracy in predicting Conscientiousness varies from 86.66% when the word count of the data reserves is less than 5000 words, and subsequently increases with the increase in the number of words as shown in Figure 4. We also varied the adjective distribution with the word count so as to obtain respective accuracies for the same model. It varies from an accuracy of 88.00% when the ratio is less than 0.001, improving to 93.60% when the ratio is between 0.001-0.002, and finally to 95.44% when increased to be greater than 0.003 (Figure 5). As expected there is a consistent increase in accuracy with increase in word count and the ratio AC/TWC.
The accuracy in predicting Imaginativeness varies from 93.33% at wordcount lower than 5000 words, and goes upto 99.88% for big data reserves (Figure 4).
The peaks observed in the variation of accuracy for “Imaginative” trait over the distribution of adjectives (AC/TWC) range from 85.71% accuracy for AC/TWC = 0.001, 90.69% accuracy for AC/TWC = 0.002 and finally 98.42% for AC/TWC >= 0.003 (Figure 5).
The accuracy varies from 97.70% for word count < 5000 words and subsequently increases to 99.88% as shown in Figure 4.
The accuracy of this trait varied from 85.71% for AC/TWC = 0.001 and went on to increase upto 99.68% for AC/TWC = 0.002 and then 99.70% for AC/TWC >= 0.003. The correlations for each of word count with accuracy and AC/TWC with accuracy for each of the above mentioned coefficient implies that for “Adjectival Markers” these are highly correlated to one another. This can also be validated by the graph in Figure 6.
By means of this study we propose a simpler yet
effective method to facilitate personality extraction of people
in social media. In order to achieve this we have also
reworked some perennial theories of Lexical Psychology and
modified them with the newer concepts of machine
learning models. This technique brings about a wave of novelty
in the wide spread lexical concepts and techniques used to
achieve user personality understanding in biographical data
reserves. It is a significant contribution in the field of
Computer Human interaction, since it is not just based on the
modern model training techniques of artificial intelligence,
but also finds solid ground in the foundational theories of
human psychology. One major drawback of this study is
that, it is (as of now) most optimized and accurate when
tested on bigger data samples. This research is thus
intended to pave way for extrapolating itself to smaller data
reserves and microblogs. We intend to apply the same
technique on not just adjectives but various other parts of speech
(POS) in the near future. There are various studies which
discuss the role of a person's personality in the development
of diseases