MODELLING LEXICAL PSYCHOLOGY OF AN...

MODELLING LEXICAL PSYCHOLOGY OF AN INDIVIDUAL

Thesis submitted in partial fulfillmentof the requirements for the degree of

Master of Science (by Research)in

Exact Humanities

by

Shivani Poddar201156152

[email protected]

International Institute of Information TechnologyHyderabad - 500 032, INDIA

JUNE 2016

International Institute of Information TechnologyHyderabad, India

CERTIFICATE

It is certified that the work contained in this thesis, titled ’Modeling Lexical Psychology of an Individual’by Shivani Poddar, has been carried out under my supervision and is not submitted elsewhere for adegree.

Date Adviser: Prof. Navjyoti Singh

To Friends and Family,

Acknowledgments

First and foremost I wish to thank Prof. Navjyoti Singh for being my advisor and guide. His presencegave me immense support to venture out of my comfort zone and undertake a work as intensivelyinterdisciplinary as this. His constant ideas and faith gave me the motivation to pursue research withutmost dedication.

I thank Arpit Merchant, Akshay Minocha, Naman Govil and Sindhu Kiranmai for being the constantreviewers of my work and giving me the much needed feedback whenever in a flux.

I thank Venumadhav and Sindhu Kiranmai for being accomplices in learning about the many formidabletheories and technologies.

I thank my wingmates and my friends from IIIT for always being pillars of support in times ofuncertainty and making my stay here a memorable one. I thank my fellow EHD folks for keeping ourbrethren alive and scaling many unthinkable frontiers, inspiring me to keep achieving more.

Finally I thank my parents, Gunjan Poddar and Jaideep Poddar for being there in my support andhaving faith in me during my research. They gave me freedom and stood by my decisions. In the endtheir patience with me helped to give justice and quality to my research work.

v

Abstract

In today’s era social media is our window into creating contextual systems of advanced technology.These systems span a breadth of industries and domains. Whether it is contextual phone assistants likeSiri and Crotana or search based advertisements, all of them rely deeply on user data to teach themselveshuman-like responsiveness. A major chunk of the data available at our disposal is lexical data (primarilyfrom social media platforms). Thus, psychology infered via lexical markers online is an indispensablepart of the future of our technology.

Lexical psychology and personality are paradigms that have been given immense attention for overa decade. Most state of the art techniques use pre-existing features from social media functions (forinstance Number of followers, Likes on Facebook etc. to predict various personality models (such asBig Five, MBTI etc.) that have been around for over 70 years. Parts of our work attempt to tackle thisapproach of using pre-existing features (available for direct reference from social media). In doing so,we traced the conception of the Big Five model of personality to be concieved by means of extensivelexical analysis (both empirical and theoretical). Based on this knowledge, we illustrate the importanceof additional features (informed by the fundamentals of psychology) in training robust personality pre-diction models. Training one such model ourselves, the results we achieved were a clear indicator of theeffectiveness of lexical features versus stand-alone technological features to predict user personality.

We then move on to tackle the drawbacks of the continued usage of the Big Five model to understanduser personality since the last 70 years. At the time this model was proposed by Goldberg et al. theonly mode for obtaining relevant user data was by means of psychometric tests. While the databasesavailable to us have advanced significantly, the models used to capture lexical personality from thesedata reserves remain constant. They make available to us an overall persona of an individual. Contrary tothese, abundant literature in the domain of psychology suggests that the persona of an individual is everevolving. It is dynamic and changes temporally with the individual’s cognition, action and experiences.As opposed to this our technology today is still dependent on such overall personality models (staticpersonality traits from Big Five and other’s of it’s likes). We have by means of this study attemptedto discover, formulate and compute a dynamic model of representing the persona of an individual toovercome these limitations. In each of the domains described above an iterative model will elevatethe contextual information by a manifold as opposed to the approximate static summarization of anindividual’s persona.

vi

vii

Thus, inspired by the Abhidhamma scholarship of Buddhism, we proposed PACMAN - the psychocomputational model of an individual. The model theoretically represents a stochastic finite state ma-chine that encapsulates the moment by moment mental states of an individual. It is empowered by itsgenesis which lies in the foundations of Brentano’s western psychological theories, phenomenologicalthinking and strong parallels with the characteristic biomedical processes that define the psychology ofan individual. Another feature which helps this model to carve it’s own niche is it’s close resemblance toa formal system. Much like a formal model is defined by a set of states and their inter state transitions,this model is also defined by it’s moments and transitional rules between these moments. It is adept inestablishing a definite schema which can be ubiquitously populated by any given data reserve and helpsin visualising the persona depicted. Thus, we first categorize the temporal (time to time) social mediadata into pre-defined mental states (concieved via Abhidhamma). These mental states are then popu-lated into the respective temporal moments of our automaton. Finally, they are used to draw variousinferences about the lexical psychology and/or personalities of the respective social media users.

We bring closure to our work by illustrating the functionality of our model in three different use cases.We first use our model to formalise the abstract idea of an individual in the concept of social machines.We also exemplify how our model tackles sparsity and non-uniformity of data in social media. Thesecond use case is to model the psychological phenomenon of Anxiety, its conception, continuationetc. by means of lexical data available online. We finally introduce the lexico-psychological mentalengagement factor to capture the maximum engagement of each individual with any given mental state.As opposed to the psychometric methods used to capture such a factor, our model uses lexical features tocompute it. We discuss the importance of each of these in contributing towards a deeper understandingof the lexical persona of a an individual (primarily a social media user).

Contents

Chapter Page

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Big Five Model of Lexical User Personality . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.1 Lexical Psychology and Big Five . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.3 Problem Description & Challenges . . . . . . . . . . . . . . . . . . . . . . . 31.1.4 Domain Specific Feature Extraction via Psychology Fundamentals . . . . . . . 4

1.2 Critique of Big Five - Moving Towards PACMAN . . . . . . . . . . . . . . . . . . . . 41.2.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.2.2 Problem Description & Challenges . . . . . . . . . . . . . . . . . . . . . . . 61.2.3 Stochastic Modeling and Muti-Class Training Models . . . . . . . . . . . . . 7

1.3 Applications of the Formal Adaptation of Lexical Psychology . . . . . . . . . . . . . 71.3.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.2 Problem Description, Challenges & Solutions . . . . . . . . . . . . . . . . . . 8

1.4 Contributions of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Understanding the Big Five Model of Lexical Personality . . . . . . . . . . . . . . . . . . . 122.0.1 Lexical Hypothesis of Psychology . . . . . . . . . . . . . . . . . . . . . . . . 122.0.2 Peer Judgments to infer Psychology . . . . . . . . . . . . . . . . . . . . . . . 132.0.3 Big Five Model of Personality . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.1 Mining Lexical Big Five Personality from Biographical Data/ Peer Judgments . . . . . 142.1.1 Datasets Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.1.1.1 Biographical Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.1.1.2 Evaluation Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.1.1.3 Feature Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.1.2 Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.1.2.1 Training Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.1.2.2 Testing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1.3 Our Approach - Adjectival Marker Technique for Mining Lexical Personality . 182.1.4 Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2 Mining Big Five Personality from Mobile Data . . . . . . . . . . . . . . . . . . . . . 202.2.1 Dataset Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.2.2 Our Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2.2.1 Mobile Data Features . . . . . . . . . . . . . . . . . . . . . . . . . 222.2.2.2 Logistic Regression Model . . . . . . . . . . . . . . . . . . . . . . 22

viii

CONTENTS ix

2.2.3 Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.2.3.1 From Correlation Analysis . . . . . . . . . . . . . . . . . . . . . . 242.2.3.2 From Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . 25

2.3 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Beyond the Staticity of Big Five - Modeling the Dynamic Psyche of an Individual . . . . . . 273.1 Theoretical Psychology by Abhidhamma . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.1.1 Formal Ontology of an Individual by Abhidhamma Traditions . . . . . . . . . 273.1.2 Comparing Abhidhamma with Other Psychological Doctrines . . . . . . . . . 30

3.1.2.1 Abhidhamma and Brentano’s Philosophy . . . . . . . . . . . . . . . 303.1.2.2 Abhidhamma and the logics of Phenomenological Thought . . . . . 313.1.2.3 Abhidhamma and Empirical Cognitive Processes (Bio-Medical Pro-

cesses) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323.2 Stochastic-Computational Modeling of Lexical Psychology using Abhidhamma’s model

of Individual’s Psychology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2.1 Automaton Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2.2 Stochastic Finite State Automaton for an Individual . . . . . . . . . . . . . . . 353.2.3 Stochastic-Computational Lexical Model (PACMAN) . . . . . . . . . . . . . 38

3.2.3.1 Dataset Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.2.3.2 Pre-Processing Data . . . . . . . . . . . . . . . . . . . . . . . . . . 393.2.3.3 Features Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.2.3.4 Multi-Label Classification - Binary Relevance Method . . . . . . . . 403.2.3.5 Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . 413.2.3.6 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4 Applications of Stochastic-Computational Lexical Model of Psychology . . . . . . . . . . . 474.1 Social Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.1.1 Proof of Concept: Working Example . . . . . . . . . . . . . . . . . . . . . . 494.2 Modeling Lexical Psychological Phenomenon . . . . . . . . . . . . . . . . . . . . . . 524.3 Lexico-psychological Engagement Factor . . . . . . . . . . . . . . . . . . . . . . . . 554.4 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

List of Figures

Figure Page

1.1 Static versus Dynamic user personality from social media data. The blue bubbles rep-resent social media data di at any time instance ti. The left side represents the existingmodels spanning across the complete data to output one static user persona. The rightside illustrates our approach of modeling a temporal persona to capture the dynamics ofevery stage in an individual’s life. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Data Mobilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2 Methodology for training and testing the logistic regression model by the adjectivalmarker technique. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.3 Accuracy variation over word count of testing data . . . . . . . . . . . . . . . . . . . 20

2.4 Accuracy variation over adjective distribution (AC/TWC) in testing dataset . . . . . 21

2.5 Comparison of accuracy results of our technique (1st column) with other state of the arttechniques of Montjoye et al [26]. (2nd column) and Staiano et al [78] (3rd column). . 24

3.1 Represents a conceptualization of the mental states that populate the model of an indi-vidual. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2 Represents a conceptualization of the model of an individual. It handles the temporalcharacteristics with the sequencing of subsequent moments one after the other M i+1follows Mi. It also handles the sparsity of data by means of “Sleep States” (Translu-cent Blue Bubbles) accommodated amongst the populated states (Dark Blue Bubbles).Finally it attempts to illustrate the varied granularity which can be manifested in the re-spective moments. For instance multiple blue moments make up bigger mega moments(grey). This phenomenon can be witnessed both top-down and bottom-up. . . . . . . . 29

3.3 DFA accepting all strings with a substring 01, from [41] . . . . . . . . . . . . . . . . . 34

3.4 An NFA accepting all strings that end in 01, from [41] . . . . . . . . . . . . . . . . . 34

3.5 A stochastic finite state machine representation for the model of an individual. . . . . . 37

3.6 Maximum Moments reflect the number of statuses which have a consecutive set of sim-ilar mental factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.7 Maximum Moments reflect the number of statuses which have a consecutive set of sim-ilar mental factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

x

LIST OF FIGURES xi

4.1 The Worry mental state is either a conceived product of the previous mental factors oran induced mental state from external stimuli/object. Both of these instances can bemodeled as a cognitive/physical object (m). This state gives rise to the next subsequentstate Fear. These further manifestations are a product of both the personality of anindividual and the external factors as defined by Abhidhamma. . . . . . . . . . . . . . 54

4.2 State by state evolution of anxiety in performing musicians (adapted from [45]). Thesestates can inhere in a moment or across multiple moments based on the lexo-psychologicalengagement factor of an individual. . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.3 The figure represents the Lexico-Psychological Engagement factor for 50 users as ana-lyzed by the (mental factors mined from) statuses posted by these users over a span of 1year. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

List of Tables

Table Page

2.1 Comparison Notations for Personality Models . . . . . . . . . . . . . . . . . . . . . . 162.2 Correlations between Personality traits. The left hand column illustrates the MBTI per-

sonality set that has a high strength of correlation with the right hand column combina-tion of Global 5/Big 5 personality traits. These corroborate the comparisons illustratedin Table 2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.3 Factor Loadings of 5 of the 435 adjectives presented by Saucier et al (1996). (FactorI - Extraversion, Factor II - Agreeableness, Factor III - Conscientiousness, Factor IV -Emotional Stability, Factor V - Intellect/ Imagination) . . . . . . . . . . . . . . . . . . 18

2.4 Adjectival Marker samples for various traits. Samples with values > 0 in the SaucierGoldberg table have been given a binary count of 1, while those lower than 0 have beengiven 0. (*Decimal Values taken from Saucier et al (1996)). . . . . . . . . . . . . . . . 19

2.5 Table of Features aggregated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.1 Inter-annotator values. G is the labeled data by annotator 1, A is the labeled data setfrom annotator 2 and B is the labeled data set from annotator 3. . . . . . . . . . . . . . 39

3.2 Row of Matrix Mi,j which was used for training and testing the Multi-Label ClassifierModel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3 Analysis of Mental factors, (enaco) tuple represents the Big Five traits in the orderof Extraversion, Neuroticism, Agreeableness, Conscientiousness and Openness, here ymeans that the trait is present and n means it is absent. . . . . . . . . . . . . . . . . . 44

4.1 Observing and analyzing the role of an Individual in Facebook as a social machine. . . 504.2 Motivation and Mental Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514.3 This table illustrates the probable mental states (Mental Factors Extracted) mined from

the usual state of minds of performing musicians through their hypothetical Social Me-dia Statuses. The corresponding Lexical Engagement Factor µ from our analysis hasbeen listed in column 3 (and explained in section 4.3) . . . . . . . . . . . . . . . . . . 54

xii

Chapter 1

Introduction

Understanding psychology, emotion and personality profiles are a key to unlocking elusive humanqualities. These qualities provide valuable insights into the interests, experiences, behavior and opinionsof the respective individuals. Personality helps us in fingerprinting an individual, which in turn is usefulin decoding the human behavior, mental processes and affective reactions of people over time towardsvarious external stimuli. Contextual systems used in a multitude of domains for instance e-commerce,advertisements, e-learning etc. could greatly benefit from such user insights [59]. By means of thisresearch, we undertake a two fold research process. The first attempts to devise new lexical featuresand techniques leveraging the foundational theories of lexical psychology, which when modeled usingvarious machine learning algorithms strengthen the state-of-the-art personality predictions (specificallybig five personality of social media users). The second studies the shortcomings of these existing mod-els of static psychological analysis. We critique the existing models of static personality prediction bymeans of abundant literature advocating personality to be dynamic and ever evolving. So as to capturethis idea in its essentiality we propose a stochastic finite state machine model for capturing the moment-by-moment (over various time instances) lexical persona of an individual (from their social media data).We conclude the work by studying the various psychological phenomena which can be modeled usingthis formalaisation, leading to the proposition of a lexico-psychological engagement factor for the user.This model has shown to be an essential tool in recognition, observation and tackling of various psycho-logical phenomenon that can be captured using lexical cues of social media. We also tackle the idea of aubiquitous user in defining social machines so as to unify a multitude of definitions that have attemptedto do so. Thus, evolving lexical personality recognition has proved to be an important step for variousknowledge extraction, contextual assistance and cognitive understanding purposes.

1

1.1 Big Five Model of Lexical User Personality

The Big five model of user personality, is characterized by five basic traits as explained below insection 2.0.3. It is by far the most robust determinant of user personality from lexical cues. The foun-dational adaption of big five has been through lexical markers studied by psychologists in 1970s. Sincewe wanted to capture the lexical psychology of individuals, big five was the most obvious choice for usto start our work from. We thus, started our exploration into unraveling the lexical persona of a socialmedia user by analyzing their big five personality from various psychological features.

1.1.1 Lexical Psychology and Big Five

Concentrated research in the field of lexical psychology can be traced back to the origins of Big Five,whose growing acceptance by personality researchers profoundly influenced the scientific discourse inindividual differences. Directed efforts to organize the language of personality (for English language)began shortly after exposure to Klages’s work on examining personality terms commonly found inthe German language. [?] These were the primary influences on the work of Allport and Obert [3]on conceptualizing the idea of a ‘trait’ 1 and thus, undertaking their own examination of the Englishlanguage of personality. Their work had a direct impact on various studies that followed, including thatof Cattell’s [17]. His system, based on factor-analytic studies of peer ratings of college students, andlater extended to both the questionnaire and objective test realms, was a more objective approach tothe organization of the thousands of terms in the English (or any) language used to describe individualdifferences. The complexity of the system encouraged many scientists to work on simpler correlationsas opposed to the ones suggested by Cattell. Fiske, one of the researchers working on it found evidencefor a simple five factor solution using Cattells 21 bipolar scales [30]. These were the foundational factorsof the big five personality traits. These were then refined by researchers such as Norman [60], Goldberg[36] etc. to arrive at the big five model of user personality, as we know it today.

1.1.2 Literature Review

The last few years have witnessed a considerable escalation in studies which are directed at mininguser personalities from social media data. Those which are related to this work can be mined in mainly2 classes. (i) Studies which are based on lexical cues to mine author’s personality, (ii) Studies whichhave used social media based features to study the personality of the user. The former section includeswork by Tausczik and Pennebaker [80] wherein they mined author personality via LIWC (Linguistic In-quiry 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 neuroticismand extraversion (Argamon et al [5]). Oberlander & Nowson [62] classified extraversion, stability,

1discussed below, this would go on to become the building block of big five in the subsequent years

2

agreeableness and conscientiousness of blog authors using n-grams as features and Naive Bayes algo-rithms. Mairesse et al, [54] reported a long list of correlations between Big5 personality traits. Theyobtained those correlations from psychological factor analysis on a corpus of essays and audio cues[65] to develop a supervised system for personality recognition. Luyckx et al, (2008) [51] [52] built acorpus for stylometry and personality prediction from text in Dutch using n-grams of Part-Of-Speech(POS) and chunks as features. They used the MBTI schema in place of the Big5 (it includes 4 binarypersonality traits, see Briggs & Myers (1980)). Along the same lines, Iacobelli et al [42] used as fea-tures, word n-grams extracted from a large corpus of blogs, testing different extraction settings, such asthe presence/absence of stop words or inverse document frequency. They found that bi-grams, treatedas Boolean features and keeping stop words, gave substantial results using Support Vector Machines(SVM) as learning algorithm. Kermanidis [46] followed Mairesse et al. and developed a supervisedsystem for POS tagging in Modern Greek, based on low level linguistic features, such as Part-of-Speechtags, and psychological features, like words associated to psychological states like in LIWC. His re-search also somewhat operated along the lines of Lexical Hypothesis by mining author personalities viaKMeans clustering algorithms.

Personality Analysis in Social Media Analysis is a recently observed phenomenon. Herein, somesubstantial work was done by Goldbeck et al, [32] wherein the authors predicted the personality of 279users from Facebook, using either linguistic or social network specific features. Quercia et al [68] usednetwork features to predict the personality of 335 Twitter users, using M5 rules as learning algorithm.Various means of evaluation have been used by the above researchers, ranging from accuracy to AUC(Area Under the Curve) values so as to establish relative accuracies of models against each other. Theabove have been discussed and captured very effectively by Celli et al [18]. One important observationwhich comes to surface while analyzing relevant literature is that, none of the studies so far have ex-ploited the primordial lexical hypothesis and Adjectival Traits suggested by Saucier and Goldberg [72][37]. Our work carves a very different niche for itself by computing this very approach of personality ad-jectives, encapsulating the last 80 years of theoretical, and empirical lexical psychological research andmerging it with the latest computational techniques. This confluence has yielded encouraging resultsin predicting traits matching those predicted by a psychometric test. The accuracies are substantiallyabove that exhibited by the current state of the art models.

1.1.3 Problem Description & Challenges

While the work in the area undertaken to mine lexical user personality from social media is extensive,there is a dire scarcity of research deriving inspiration from theoretical foundations of psychology. Thefeatures extracted to train the various (state-of-the-art) machine learning models were selected solelybased on available information (for instance No. of likes on Facebook, No. of followers on Twitteretc.), as opposed to being aggressively filtered and pre-processed based on foundational psychologicaltheories developed for lexical psychology. As a result, despite of the abundance in user information

3

online, the accuracy achieved by various models has been sub par. Another area that has lacked attentionis the “social” aspect of social media. While our own activities on social media provide valuable insightinto our personalities, various psychology theories [71][48][61] suggest that the perception of us byour peers is also a viable indicator of our overall personalities. We have attempted to address theseissues in the first part of our research. An indepth domain exploration into lexical user psychology andpersonality suggested that the cues mined from an individuals vocabulary provides unparalleled insightsinto the psychology of an individual. We also learned that these lexical cues embedded in the languageof individuals affiliated to a person, lend equally important insights into an inference of their personality.

1.1.4 Domain Specific Feature Extraction via Psychology Fundamentals

Since all the existing models (in our knowledge), depended upon training based on a pre-selectedset of features directly from social media, with little or no domain knowledge, we attempted to bridgethis gap between these computational models and the psychological theories of personality so as toelevate the accuracy of the state of the art models (of personality prediction from lexical data). Ourresearch leverages from the age old theories of psychology, extracting adjectival markers as the lexicalcues embedded in the vocabulary of an individual to predict user personality. We also explore the roleof peers in determining the personality of an individual. We extracted these adjectival markers fromthe biographical data of an individual (i.e. descriptions of an individual by his/her peers) and trained alogistic regression model so as to predict their personality. The high accuracy that this approach yieldedin predicting big five user personality also helped us improve upon the theoretical details of the lexicalhypothesis of psychology. Finally, so as to evaluate the importance of lexical features to predict founda-tionally lexical models such as big five personality, we also used (non-lexical) features extracted frommobile phone data to train another logistic regression model. Much lower accuracy was achieved usingthis, which indicated the importance of lexical cues in predicting such models as well as the salience ofthe lexical hypothesis of psychology in the context of developing predictive models.

1.2 Critique of Big Five - Moving Towards PACMAN

Over extensive literature reviews of psychological studies, we discovered that the psychology andpersonality of an individual are not static singular overall personality, as illustrated by the big five modelbut is dynamic in nature. All the applications of contextual advertisements, human computer interactionsystems, sensitizing machines to their users, perceptual assistants depend upon the user psyche and per-sonality. [22][56][76] Thus, it is of paramount importance that the temporal nature of user personalitiesis captured by the models being used to train these systems. Much like a time-series data reserve, ourpsychology is also a collection and result of its past formations. By means of our research we attemptto model the lexical psychology of an individual as the combination of a set of mental factors which

4

have been dominant in the individual for a significant amount of time. The lexical psychology here,unlike state of the art models, is not static or of a specific type, but keeps evolving with the individualhimself. For instance, if a child demonstrated acute “Selfishness” in the early years, but grew out of iteventually, their psychology would manifest selfishness in the respective time span (namely childhood),and eventually evolve to get rid of deprecated traits. In essence our model attempts to capture a per-sonality trait (or a set of mental states) from the grassroot level (i.e. the beginning moments when theystart to manifest in a person) to the time when it matures and defines a person. (i.e. a series of sub-sequent moments when it starts recurring without fail). Thus, in keeping with this discovery, we nextattempted to solve this problem by adapting and computationally modeling, training and testing a dy-namic representation of a user’s persona. The gold standard we used to train this was largely lexical soas simultaneously to dig deeper into the lexical psychology of a social media user. By means of our workwe also attempt to present a discrete moment based model. In order to understand and model the clas-sical rules of user psychology we draw inspiration from traditional doctrines of Theravada Buddhism.Much like the foundational schools of thoughts of western psychology by Brentano et al. this doctrinealso captures concepts like perception, granularity of mental thought and intentionality of an individual.However, unlike western models the processes described in this model are in close conjunction withthe actual biomedical processes contained in mental transitions in the brain. We further reinforce themodel’s footing (in describing mental phenomenon) by drawing parallels in its conception with that ofmathematical logic. A central concept to this doctrine is that, there is a total ordered temporal sequenceof moments that captures the consciousness (or lexical psychology in our adaptation) of an individual.This is similar to the ordering of states in formal language models. This feature helps such models tobe rendered into a definitive domain schema. This can be populated with different types of data so as todraw needful inferences. Thus, we proceeded to model this sequence of moments (and their respectivemental-states) described in Abhidhamma as states of a finite state automaton. We could also define therespective transition functions and heuristics to compute the automaton and refine a working model ofthe same (from social media data).


Various studies of social media have attempted to capture the Big five traits (as suggested by section1.1.1) extensively from websites such as Twitter [32], Facebook [70], Blog data [67] etc. A recurrentunderlying theme that all the research in the domain has in common is that of a constant user personality.The suggested personality of a user mined by means of most of the state of the art techniques focus onextracting the overall personality of a person. For instance, Goldbeck et al. classify the subjects intoone of the Big Five categories by means of extensive feature extraction from Facebook (Page likes,comments etc.). Although initial literature in psychology did suggest that personality remained constantafter the age 30, many recent studies (as cited above) contradict this notion. More such studies thatencapsulate the importance of capturing temperamental changes in adolescence which later on can be

5

connected to adult behavior have been undertaken by McCrae [56] , Specht [76] etc. The importance offacet-level research for understanding life span age differences in personality has also been recognizedby Soto et al. [75] their work.

Figure 1.1 Static versus Dynamic user personality from social media data. The blue bubbles representsocial media data di at any time instance ti. The left side represents the existing models spanning acrossthe complete data to output one static user persona. The right side illustrates our approach of modelinga temporal persona to capture the dynamics of every stage in an individual’s life.

1.2.2 Problem Description & Challenges

There is no computational model that attempts to establish coherence with the psychological theoriesof variability of an individual’s personality (thus, the Big Five personality) across various age groups(starting from 18 towards 65) as suggested in previously cited studies. No present models as yet, at-tempt to tap into personality changes over any given period of time. Thus, capturing this dynamic ofpsychological transitions from one moment to another has been a long standing and critical problem.

Another issue that plagues the development of computational models to capture the variability ofdata is the lack of gold standard. This prevents models from training and testing moment by momentmodels that capture the temporal personality of an individual.

Subsequently, the lack of such dynamic personality models also prevents the current applications ofpersonality mining from social media to leverage the actual personality of an individual in a period oflife, as opposed to their approximate overall static personality.

6

1.2.3 Stochastic Modeling and Muti-Class Training Models

Our study attempts to tackle each of the problems stated in section 1.2.2. Firstly we adopted a pri-mordial psychological theory that conformed with the formal modelling theories, western psychologicaland biological descriptions of individual personality. We then translated the various heuristics of thistheory to a computationally salient system by defining a formal model for it. The model accommodatesseveral rules defined by this theory of Abhidhamma and supports data population so as to make infer-ences about the personality of a person based on their data. While initially we used manual methods tounderstand, populate and infer personalities of social media users, we finally transcended this theory toa computational framework (PACMAN).Thus, the first step of this part of the thesis was to create a stochastic finite state machine, to describemoment by moment states of an individual (by means of their time stamped social media data). The sec-ond step was to train a machine learning module (using mutli-class SVMs) enriching this state machinewith its moment by moment mental states. This was useful in learning the various rules of inferringpsychological states and personalities from the lexical cues populating the moments of the former statemachine. So as to evaluate our model, we also worked on creating a gold standard with the help ofannotators and domain experts who tagged the respective social media statuses and lexical markers withthe relevant mental states described by our model.

1.3 Applications of the Formal Adaptation of Lexical Psychology

The formal adaptation of the Abhidhamma Theory to infer the lexical psychology of an individualhas helped us to mold our model for various use cases in applications required to observe, analyse andformalise the psychology of an individual. The three major areas that we intend to cover by means of ourresearch are 1. Proposition of a ubiquitous model defining the role of an individual in social machinesalong with tackling the sparsity of data for the model available online. 2. Modeling Lexical Psychologi-cal Phenomenon (for users across demographics) using lexical cues available on social media websites.3. Formulating the Lexical User Engagement Factor to capture the immersion of a user in any mentalstate for a sustained time period.


Social Machines: There is a plethora of work revolving around systems between man-machine col-laborations, and interfaces facilitating this interaction. For instance, Human-Computer interaction [16]is an area of applied cognitive science and engineering design is concerned with understanding how peo-ple make use of devices and systems that incorporate computation. Social Computing is concerned aboutthe intersection of user behavior and computational systems. It finds inspiration in creating or recreatingsocial conventions or contexts with software systems and technology. Personality Mining [29][67] is

7

a newer domain of computer science which focuses on capturing the psychological processes and dis-positions of an individual through the person’s publicly available data. Encompassing these outlooksinto analyzing socio-technical systems is the concept of a social machine. Social machines are typi-cally presented as systems that combine some form of social participation with conventional forms ofmachine-based computation. There are a multitude of definitions attempting to encompass this idea ofsocial machine, and the role of the individual and computers within these systems. We unify these def-initions by means of a singular model for the participation of an individual in context of social machines.

Modeling Psychological Phenomenon Lexically: While there are many studies which revolve aroundthe need to recognise various psychological phenomenon from social media [25], there is a dearth ofresearch attempting to achieve this. The inverse of the former has been extensively researched upon.With studies pertaining to the differences in social media engagement between individuals sufferingfrom anxiety versus those who do not [85]. No work so far has attempted to model psychological phe-nomenon as a function of various social media markers.

Lexical User Engagement: There are various works that attempt to capture the cognitive and psy-chological engagement [4][44] by means of a psychometric measure based on questionnaire responses.Many other works attempting to exploit the social media reserves have proposed mechanisms to exploreuser engagement with a third person object. For instance, there are various works that address the cap-ture and improvement of consumer engagement on Social Media [9]. There are also studies that analyzethe engagement of individuals with communities or governments [69]. However, no work so far hasattempted to recognise psychological engagement from lexical cues on social media.

1.3.2 Problem Description, Challenges & Solutions

Social Machines: The multitude of definitions describing the role of an individual in a social machineposed an issue of unifying the conceptualization of a social machine. Our model helped by proposinga consolidated formal model which could fit each of these different perspectives and fit them togetherby means of theoretical formalisms and logic. Another issue that needed tackling in this domain wasthat of sparse descriptive data online. Our model proposed the population of states on various levels ofgranularity (explained further in Section 4.1) thus solving this problem effectively.

Modeling Psychological Phenomenon: Various studies as explained above, stressed on the need torecognise psychological phenomenon from the data available to us online. We leveraged the temporalaspect of our model to recognise mental states (and thus the person’s psychology) at any given instance.This helped us to model the onset, conception, persistence and recovery of any individual’s psychologi-cal phenomenon (here Anxiety) from their social media updates and data.

8

Lexical Engagement Factor: While there was abundant literature on determining psychological en-gagement by empirically fitted analysis of psychometric tests, we could not find any such studies whichhypothesized the same using lexical data online. We found no detailed works on understanding andmodeling of the engagement of a social media user with oneself. We thus, attempt to capture the depthof this engagement which is varied for each user by means of the Lexico-Psychological Factor (µ) asdescribed in Chapter 3. We leverage the continuous number of moments engaging in similar mentalstates to arrive at this hypothesis. Our work is validated and based on the theoretical constructs of userpsychology.

1.4 Contributions of this Thesis

The contributions (C) of this thesis can be divided into the following two major phases.

• Phase 1The first contributions of the thesis are, C1: to evaluate and improvise the “Lexical Hypothesisof Psychology”. C2: We propose the Adjectival Marker Technique inspired from the traditionaltheories of lexical psychology and formulate better features so as to effectively predict user per-sonalities from social media. This has helped in elevating the state of the art accuracy of predictingthe big five user personality significantly.

We have also attempted to verify the dependency of Big Five on lexical data cues as opposedto features obtained by other data sources. C3: We implemented a model that predicted theuser personality based in their mobile phone data. Although elevating the current state of the artmodels used to predict big five personality from this dataset, the model was not in par with theaccuracy various systems have achieved through lexical feature modeling.

• Phase 2As the second phase, we provided critiques of the static personality mining of users. In an attemptto move towards a more dynamic and formally adaptive user model we adopted the Abhidhammamodel of Buddhism. We discussed various logical stances about the salience of this model asopposed to any other theoretical model available. So as to formalize this conjecture, the thesiscontributes the following. C4: A Formalized Ontology of mental states (adapted from Abhid-hamma) and a Stochastic Model of an individual to capture the evolving user personality fromthese moment-by-moment mental states. C5: A Multi-Label machine learning module (PAC-MAN) to enable training of individual user social-media statuses with the respective mental fac-tors. C6: Labeled dataset of 4,179 Facebook statuses (of the mypersonality dataset [19]) with therespective annotated mental factors 2. C7: An indepth analysis of the mental factors and evolvingpersonalities of 50 users from the above mentioned dataset and the comparisons of these with theBig Five traits of the same users.

2https://researchweb.iiit.ac.in/˜shivani.poddar/PACMAN_Dataset

9

We furthered the above models and analysis by also studying some use cases and applications ofour proposed model of user personality. The first one of these was the proposition of a ubiqui-tous model of an individual in the context of social machines. The work contributed in C8: thedevelopment of a unified framework for modeling an individual participating in a social machine.This framework is useful not only for classification purposes of social machines, but will help usin undertaking thorough observation and analysis of the persona of the given individual. C9: Themodel also solves issues of temporal variability of processes and sparseness of individual dataencountered while observing and monitoring social machines.We also achieved an C10: expansive modeling of psychological phenomenons, Anxiety by meansof our model. As illustrated further on, our model can encapsulate the mental states characterizedat the initial conception of anxiety, persistence and inherence of anxiety oriented mental states inan individual. This model can thus be assistive in capturing anxiety related disorders of peopleactive on social media very early on. The final contribution of the work C11: is the conceptionand proposition of the Lexico-Psychological Engagement Factor for an individual. While thereare various works that attempt to capture the cognitive and psychological engagement [4][44] bymeans of a psychometric measure based on questionnaire responses, there is no literature that wecould find that defined psychological engagements of users based in their lexical data. Modelsto ascertain this engagement have so far been proposed based on the best empirically fitted anal-ysis of psychometric tests of individuals. Thus, our work stands out by calculating the mentalengagement of a social media user by means of only their lexical data.

1.5 Thesis Organization

The rest of the thesis is organized as follows. Chapter 2, Section 2.1 discusses the modeling of bigfive user personality by leveraging the fundamental theories of psychology. We also discuss validatingthe Peer Judgment theories and the Lexical Hypothesis of Psychology by means of this model. Section2.2 explores the modeling of the same personality traits using non-lexical mobile phone features.

Hereon, Chapter 3 recognizes the shortcomings of the current personality model as being static.Thus, so as to keep in line with the psychological literature which defines personality as dynamic andever evolving, we attempt to adopt a more dynamic model, and formulate a computational variant forthe same. Section 3.1 expatiates on the theoretical footings of this model, Abhidhamma inspired byBuddhist theory of the persona of an individual. Section 3.2 describes a formal mathematical formula-tion of this as a stochastic finite state automaton. In Section 3.2.2 we populate it with social data so asto train a multi-class classifier on top of this model and predict the moment by moment mental statesexhibited by an individual. We finally summarizes how these are adept in drawing inferences about thelexical psychology of an individual in Section 3.3.

Chapter 4 describes the applications of this stochastic finite state machine. Section 4.1 formulates itas the ubiquitous model of an individual in the context of social machines, also addressing the problem

10

of sparse data in social media. Section 4.2 illustrates the usefulness of this model to map psychologicalphenomenon such as Anxiety, solely using lexical cues from social media. Section 4.3 proposes theconception of the Lexico-psychological Engagement Factor, computed by populating our model.

Finally, Section 5 presents the conclusions drawn by means of this work. We also cover the futurework can be undertaken to extend this study, in this section.

11

Chapter 2

Understanding the Big Five Model of Lexical Personality

Psychology refers to individual differences in characteristic patterns of thinking, feeling and behav-ing. The study of psychology focuses on two broad areas: One is understanding individual differences inparticular personality characteristics, such as sociability or irritability. The other is understanding howthe various parts of a person come together as a whole. With the onset of social media, there is abundantinformation about users so as to capture these characteristic differences among people. An importantaspect of studying user psychology is capturing their personality. It functions as one of the buildingblocks to the larger psychology exhibited by an individual. Thus, so as to explore the psychology of anindividual we begin by first capturing their personalities. Personality recognition is the task of assigninga personality profile to people. Earlier, this was done by having the subjects manually fill out a question-naire and evaluating their answers. However, automatic recognition through Facebook, or video and/ortextual data has found remarkable success in recent years [32]. Lexical Psychology/Personality aims atleveraging this social media data (particularly lexical data) so as to make useful deductions about thepersonality of an individual.

2.0.1 Lexical Hypothesis of Psychology

The Lexical Hypothesis [3] (also the Fundamental Lexical Hypothesis, Lexical Approach or Sed-imentation Hypothesis) is one of the most important and widely used guiding scientific theories inpsychology. Despite some variation in its definition and application, the Lexical Hypothesis is generallydefined by two postulates. The first states that those personality characteristics that are most importantin people’s lives will eventually become a part of their language. The second follows from the first,stating that more important personality characteristics are more likely to be encoded into language as asingle word [17], [61], [43]. The Lexical Hypothesis is an important foundational unit of the Big Fivepersonality traits [34], the HEXACO model of personality structure, the 16PF Questionnaire and hasbeen used to study the structure of personality traits in a number of cultural and linguistic settings. Weleverage from this fundamental hypothesis, the big five model inspired by it and explore it further so asto validate it and also empirically accommodate the theoretical fundamentals discussed in Section 2.1.2.The theories of psychology were influenced by various revolutionary concepts, for instance, “trait” - a

12

theoretical construct which describes a basic dimension of a person’s personality. The idea of trait gavebirth to the “Lexical Hypothesis of Psychology”. The initial direction of this paper was solely governedby this exact hypothesis:

Those individual differences that are most salient and socially relevant in people’s liveswill 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 theone for Greek Language described by [?]. Motivated by the same inspiration, we too worked on extract-ing author’s personality traits from the English language text they wrote (as opposed to Greek text inthe cited study). This involved mobilizing huge datasets of web blogs and essays and extracting “traits”from them to determine the author’s personalities. However, by the course of our study, we found outthat this was not as effective as the initial hypothesis proposed [43]. The average accuracy of the initialexperimentation was less than 50%, which was as good as a randomly predicted personality set. Thescanty data on social media websites in user’s vocabulary defining themselves restricted this analysis.We realized that another unexplored aspect of any individual’s social media profile was the lexical cuesleft by their peers. Thus, as an effort to corroborate the previous analysis on lexical psychology we ex-plored deeper into the foundational theories of the subject arriving at the Peer Judgments of Psychology(described in Section 2.0.2).

2.0.2 Peer Judgments to infer Psychology

While Lexical Hypothesis described in section 2.0.1 attributes the personality of an individual to beembedded in his/her language, many other studies also infer the personality of an individual as a functionof their peer’s judgments about them. The very genesis of lexical psychology saw various researcherscapturing personality differences between individuals via peer descriptions. For instance, as dicussedin section 1.1.1, Norman [61] used the peer nomination ratings to arrive at a simpler version of the 31bipolar scales suggested by Cattell[17]. Another instantiation, supportive of this claim theoretically,is that of Brentano [13] . He postulated that all and only psychological states exhibit intentionality,and that in this way the subject matter of psychology could be demarcated. This intentionality of anindividual gets captured in the bubble of impressions maintained by their peers over time. Thus, whenthese peers describe the individual in question, the lexical cues (much like in the Lexical Hypothesisof Psychology) are a reflection of the overall personality of this individual. As a part of this work, weempirically evaluate this stance and capture how peer judgments over the lifetime of an individual arealso a good enough measure to determine the Big Five personality of a person.

Combining the lexical hypothesis of psychology (for predicting personality) with the findings ofpeer judgments to predict personality we thus hypothesized a modification to the Lexical Hypothesis inpsychology. This modification suggests that the personality of a person is predicted based on cumulative

13

judgments of their peer’s language cues (while describing the subject) along with the subject’s ownlanguage cues. The “Adjectival Marker” Technique helps us in inferring these judgments, and is derivedfrom the adjectival markers of Big Five personality traits as discussed by Goldberg & Saucier (1996)[43]. Thus, the modified Lexical Hypothesis of Psychology proposed and verified (in the subsequentsections) by means of this work is as follows:

Those individual differences that are most salient and socially relevant in people’s liveswill eventually (over time) become encoded into their language as well as that of peoplewho 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.

2.0.3 Big Five Model of Personality

One of the most popular models for such classifications is the Big-Five Model. This model character-izes human personality using a five dimensional vector with values corresponding to bipolar attributes.This model has been used among language and computer science researchers for personality traits iden-tification and simulations because of its strong hold and inherent foundations in lexical psychologicalanalysis. It is defined as follows:

• Openness to experience: Artistic, imaginative etc.

• Conscientiousness: Efficient, organized, etc.

• Extraversion: Energetic, assertive, etc.

• Agreeableness: Compassionate, cooperative, etc.

• Neuroticism: Anxious, tense, etc.

2.1 Mining Lexical Big Five Personality from Biographical Data/ Peer

Judgments

Building on the confluence of two major domains, the primordial theories of psychology and therecent techniques of computational data modeling, we implemented the following machine learningmodel for user personality prediction. While heuristically the model is informed by the fundamentals ofpsychology, empirically we have attempted to train, and test it on real world social media data.

2.1.1 Datasets Used

Following datasets were used to undertake the approach suggested in Section 2.3.

14

2.1.1.1 Biographical Data

The data collected as a part of this study was by means of a Python-based crawler. We first used asimple web crawler to get a list of web-pages with the name of the respective person as the argumentkeyword to the crawler. These web pages were then filtered based on their meta-tags. To boost truepositives, we only considered the pages which specified their content as biographical in the meta-tagdescriptors. This resulted in mobilization of few Wikipedia resources, blog mentions and descriptivebiographical websites. We then manually cleaned the noisy data to assure entity disambiguation andirrelevant mentions. The same has been illustrated by means of Figure 2.1.

Figure 2.1 Data Mobilization

2.1.1.2 Evaluation Data

There was a dearth of data corresponding to the big five traits of users whose biographical data wasavailable online. Following this scarcity we reverted to their Jungian model personality types (whichwere easily available). So as to evaluate our model (that predicted big five traits) against the availableJungian ones, we scaled the Jungian personalities into big 5 using the matrix illustrated in table 2.1.Thus, The Jungian Personality functions of 574 personalities were extracted from the personality re-source 1 to evaluate our model (as described in Section 2.1.3). Thus, we were able to train and evaluateour computed predictive model by scaling the Jungian Typology type to the closest traits of the Big5using correlation factors as shown in Table 2.2 [15] [1] [63]. Table 2.1 shows the supporting notationsof the personality systems.

As illustrated, we were able to scale 4 (of the 5) personality traits (each of which had medium to highcorrelation with the MBTI types) namely - Agreeableness (Accommodation - A/E), Extraversion (R/S),

1http://www.celebritytypes.com

15

Big5/ Global5 Jung/MBTI/Kiersey Strength of Correlation

Extraversion Introvert/Extrovert HighEmotional Stability Feeling/Thinking Very LowConscientiousness Judging/Percieving HighAccommodation/Agreeableness Feeling/Thinking MediumIntellect Sensing/Intuition Medium-High

Table 2.1 Comparison Notations for Personality Models

Conscientiousness (Orderliness - O/U) and Intellect (N/I).) These were also the final ones that we usedto ascertain the accuracy of lexical personality prediction of an individual by the proposed AdjectivalMarker Technique.

2.1.1.3 Feature Set

The Adjectival Markers that this chapter uses are a proven indicator to reflect the Big5 traits ofpersonality [72] [37]. The adjectives mined from the biographical data were refined to extract theseadjectival markers i.e. specific adjectives descriptive of the subject of the biographical data. They werethen used as features in the final LASSO logistic regression model. The adjectival markers extractedare based on the work of Saucier & Goldberg. Table 2.3 provides the factor loadings of few of the 435adjectives [72] on each of the five factors as discussed in their work. The order reflects the relative size(variance) of the factors (e.g. Factor II is the highest), and the sign reflects the relative size of the itemsubsets at each pole of the factor (e.g. the negative pole of Factor IV has more items). We have, as a partof our study, condensed this table to solely indicate whether or not the trait is descriptive of a particulartrait, so as to achieve a binary matrix for them (for the respective 4 of the Big 5 traits mentioned above).The binary equivalent for Table 2.3 is shown in Table 2.4.

2.1.2 Pre-Processing

Biographical data was mined for 574 personalities from online resources as discussed in the formerSection 2.1.1. This data was divided into 2 categories. Testing data and Training data. Users with nosubstantial data (< 100 words were discarded from the analysis). The final mined data used for thestudy is spread across various social media resources including Wikipedia articles, blog posts, socialQ & A sites and community media sites ( building datasets of word-count upto a maximum of 10, 000

words or less for each user)

16

Semi-Correlating DescriptionsJung/MBTI/Kiersey Global 5

INFP RCUAI, RLUAIINTP RCUEI, RLUEIINFJ RCOAI, RLOAIINTJ RCOEI, RLOEIISTJ RCOEN, RLOENISFJ RCOAN, RLOANISTP RCUEN, RLUENISFP RCUAN, RLUANENFP SCUAI, SLUAIENTP SCUEI, SLUEIENFJ SCOAI, SLOAIENTJ SCOEI, SLOEIESTJ SCOEN, SLOENESFJ SCOAN, SLOANESTP SCUEN, SLUENESFP SCUAN, SLUAN

Table 2.2 Correlations between Personality traits. The left hand column illustrates the MBTI personalityset that has a high strength of correlation with the right hand column combination of Global 5/Big 5personality traits. These corroborate the comparisons illustrated in Table 2.1

2.1.2.1 Training Data

The training data set used to mine adjectival markers, comprised of biographic data content of 283personalities. The word count of the dataset ranged from 500 - 10,000 words. The ratio of the numberof adjectives to the total number of words in the dataset ranged from 0 to 0.005. This data contentwas mined by means of a Python-based web crawler, which parsed biographic websites,Wikipedia, andsocial media mentions.

2.1.2.2 Testing Data

The testing dataset comprised of biographic data content of a set of 291 personalities (different fromthose used for training). These were mined using the same crawler from the social media reserves aswell, eg. Wikimedia articles, blog posts about the respective personalities, social Q & A sites etc. Theword count and the number of adjectives to the total number of words ratio ranged from 100 - 10,000words and 0.0001 to 0.003 respectively.

Due to the paucity of big five personality data for subjects whose biographical data was also available,we used the technique common in alternate augmentative communication. Vertanen and Kristensson[84] resorted to invented a dataset (of phrases used by mobile phone users) that was a subset of their

17

Adjectives* II I IV V IIISympathetic 0.62 0.02 0.07 0.03 -0.05Kind 0.60 0.07 0.02 0.00 0.06Sensitive 0.46 -0.10 0.35 0.23 0.00Rude -0.50 0.08 0.00 0.06 -0.15Adventurous 0.00 0.38 -0.19 0.10 -0.04

Table 2.3 Factor Loadings of 5 of the 435 adjectives presented by Saucier et al (1996). (Factor I -Extraversion, Factor II - Agreeableness, Factor III - Conscientiousness, Factor IV - Emotional Stability,Factor V - Intellect/ Imagination)

training data. It also followed close to the same distribution as their training data. We followed a similarapproach, with our testing data being a subset of our training data. We were also able to validate this bymeans of the overall accuracy of the model.

2.1.3 Our Approach - Adjectival Marker Technique for Mining Lexical Personality

Figure 2.2 Methodology for training and testing the logistic regression model by the adjectival markertechnique.

The training data (283 users) was mined for adjectival markers according to Saucier’s adjectivalmarker list [72]. Personality traits and their adjectival markers were represented as a sparse User-Trait Adjective Matrix for each of the 4 big traits to be predicted. The entries of the respective UserTrait Matrix (M ) were set to 1 if there existed an adjectival feature marker in the user’s descriptivebiographical data and 0 if it was not there. Thus, each personality trait was calculated by it’s matrixwherein the Row of the matrix M , consisted of adjectival-features and the corresponding column entryconsisted of the User (described by the row adjectival features). The matrix entity Mij was a binarynumber which was 1 if the adjectival marker in the jth column indicated the presence of the trait Tin the personality of the user contained in the ith row of the Matrix M . To predict the binary scoreof a given personality feature, we then performed a LASSO logistic regression modeling [81] [57]

18

using the Weka Toolkit [55]. A variety of regression algorithms were tested, each with a 10-fold cross-validation with 10 iterations. The best result out of all the algorithms was achieved using a binaryclassifier with Lasso regression (with 10 fold cross validation). Using the LASSO Technique ensuredthat there was no over-fitting. We used the remaining data of 291 personalities to evaluate our model.The testing biographical data was mined for adjectival trait markers and their respective traits werepredicted. The average accuracy of each of the Big Five traits evaluated against those achieved bypsychometric tests were: Extraversion 82.82%, Agreeableness 89.62%, Conscientiousness 92.48%and Imaginativeness/Intellect 81.67%. These results have been discussed elaborately in the nextsection. Figure 2.2, which can be found above, is also illustrative of the procedure define above.

Adjectives Agreeableness Conscientiousness Extraversion ImaginativeDecimal Binary Decimal Binary Decimal Binary Decimal Binary

Sympathetic 0.62 1 -0.05 0 0.02 1 0.03 1Kind 0.60 1 0.06 1 0.07 1 0.00 0

Sensitive 0.46 1 0.00 0 -0.10 0 0.23 1Rude -0.50 0 -0.15 0 0.08 1 0.06 1

Adventurous 0.00 0 -0.04 0 0.38 1 0.10 1

Table 2.4 Adjectival Marker samples for various traits. Samples with values> 0 in the Saucier Goldbergtable have been given a binary count of 1, while those lower than 0 have been given 0. (*Decimal Valuestaken from Saucier et al (1996)).

2.1.4 Evaluation and Results

The average accuracy of each of the Big Five traits evaluated against those achieved by psychometrictests were: Extraversion 82.82%, Agreeableness 89.62%, Conscientiousness 92.48% and Imag-inativeness/Intellect 81.67%. Due to the paucity of evaluation data via psychometric tests for theNeuroticism trait, the scope of the model is limited to the aforementioned four of the big five traits. Itis interesting to note that these readings do not necessarily demonstrate the prediction accuracy of theinnate personality of a person but match the personality reflected via psychometric tests with the givenaccuracies. They are at par with other techniques predicting personalities for instance, the work of Iaco-belli et al [42] attempted to decipher the personalities of bloggers has an average personality predictionaccuracy of ≈ 62.5%. While the results demonstrate a substantial elevation to the pre-existing methodsof big five personality determination, they are also a useful in studying the role of factors such as thenumber of words or word to adjective ratios that help in determining user personality. Thus, as a part ofthis study, we also attempted to capture the the variation in accuracy with the change in these factors,namely, word count (WC) of the corpus, and the ratio of the number of adjectives to the total numberof words (AC/TWC). These were primarily helpful in exploring and discovering a threshold for the

19

Figure 2.3 Accuracy variation over word count of testing data

word count and the adjective distribution in the given document set so as to substantially predict userpersonality via the Adjectival Marker Technique.

As indicated in Figure 2.4 the accuracy in the prediction of traits increases with the increase in thedata word count. We also compared the accuracy results in predicting the respective traits on the basisof varying distribution of adjectives in the training dataset (Figure 2.3). The accuracy in predicting thetraits is relatively low when the ratio of the AC/TWC is low and increases with a subsequent increasein the AC/TWC ratio.

2.2 Mining Big Five Personality from Mobile Data

So as to understand the dependence of Big Five (and in turn lexical psychology) on Lexical cues fromsocial media as opposed to other features that might illustrate user personality, we also experimented bymodeling personality using mobile phone data features. This primarily helped us to understand in depththat the foundation of big five were lexical cues exhibited by individuals.

20

Figure 2.4 Accuracy variation over adjective distribution (AC/TWC) in testing dataset

2.2.1 Dataset Used

In the Workshop on Computational Personality Recognition, two gold standard labeled datasets werereleased along with an evaluation tool. We used the Mobile Personality Dataset for the purposes of thecurrent paper. This dataset [18] contains the results of a study conducted in a major research universityin America wherein they collected the mobile phone data of about 130 participants all of whom aremarried couples and belong to young families. The data was continuously and non-intrusively, anony-mously collected over a period of three months and to preserve the privacy of the participants. Usinga mobile phone sensing platform they recorded, all Bluetooth scans, SMS profiles and logs of voicecalls (duration, timestamp, type etc.) between any and all two participants. For the purposes of ourwork, we divided the data into training and testing sets. The former contained data of 61 participantsthat we used to train our model. The remaining 69 participants were used for testing . Class labels forpersonality traits classification were also made available which we used for the distribution of the labelsin the corpus for different Big-Five categories. This features were passed into the logistic regressionmodel (10-fold cross validation) in WEKA [55] (Waikato Environment for Knowledge Analysis) so asto train the required model. Additionally, the dataset included the results of psychometric tests obtainedfrom manual questionnaires administered as part of the exit survey. These reflected the self-perceivedpersonality of the participants and were used as a benchmark against which the accuracy of our model

21

was tested. The evaluation tool provided a means of measuring the mean and standard deviations of ourtesting sets.

2.2.2 Our Approach

So as to further study the dependency of lexical psychology inferred by means of Big Five on lexicalfeatures alone, we explored predicting it using mobile phone data. We analyzed the relationship betweensmartphone usage (various features such as calls, messages, bluetooth pings and so on) and the Big-Fivetraits using common statistical analysis techniques of logistic regression.

2.2.2.1 Mobile Data Features

The data was made available to us after anonymization and thus the features extracted are privacysensitive. The features are extracted from four main modalities namely Voice Call logs, SMS logs,Relation Status and Bluetooth (BT) scans. Features pertaining to voice calls and SMS logs were obtainedby first extracting events from twelve-hour segments of time throughout the three months. Those fromthe Bluetooth scans on the other hand, were collected by probing the phones at an interval of every fiveminutes.

Since the dataset contained logs in such a longitudinal format, we aggregated the features longenough to capture the usage of phones while also providing us with sufficient data points for analy-sis. Hence, our features are on a daily basis. That is to say, for each user, we isolated all voice calls,SMS and BT scans for each day by automatically parsing the logs. Features based on BT scans capturethe number of different people that each participant interacts with and their frequencies. On the otherhand, we consider incoming and outgoing calls as well as average number of missed calls. The list offeatures used is given in the Table 2.5. We used a total of 15 features for training our model.

2.2.2.2 Logistic Regression Model

We first created a UserTrait-Component matrix where the rows represent the participants and thecolumns represent the features. The User Trait matrix represents all the Big Five traits in the entiredata population . Thus we have five such matrices which encapsulate the five traits of the completedataset population. The entries of the matrix at the ith row and jth column represent the value of the jthfeature for the ith participant. The features used in this matrix are our dependent variables. Initially, Wetook weighted dependent variables for all the features with the values being exact numbers obtained onparsing the dataset. We then carried out several other experiments to select the component matrix. Wechose different subsets of features and obtained results to identify which features are better predictorsof personality. Lastly, we replaced real-valued variables with binary valued indicator variables. Thiswas done by fixing a particular threshold value for each feature and marking all participants with values

22

Modality Feature Threshold- Average No. of unique IDs seen per day 62.38- Average No. of IDs seen significantly per day 9.38

Bluetooth- Average No. of unique IDs seen per day during odd hours. (Spe-cific time, say from 12:00 am - 5 am)

10.90

- No. of pings from Bluetooth Proximity 68.66- No. of ping to people in Bluetooth Proximity 104.47- No. of average BT ping 21.14- Average no. of outgoing messages per day 17.67

SMS - Average ratio of incoming pings to outgoing pings per day 0.86- Average ratio of incoming pings to outgoing pings per day 0.86- Average no. of incoming calls per day 5.76

Voice Calls - Average no. of outgoing calls per day 11.0- Average no. of missed calls per day 2.76- Pings from opposite gender if couples are in the bluetooth prox-imity

7600.76

Couples- Pings from opposite gender if couples are not in bluetooth prox-imity

0.67

- Couple in same occupation (Binary value) 0.42

Table 2.5 Table of Features aggregated

greater than the threshold (for each feature) as 1 and those below by 0. We trained five classifiers foreach of the five personality traits. Logistic regression has proved to be very effective for studying therelationship between dependent and independent variables [1,7]. We found regression algorithms to beeffective in categorization and robust over the varying feature spaces. In each experiment, the resultswere averaged over ten-fold stratified cross-validation. We used the Weka package with a linear kernel.In logistic regression, the personality trait was represented as a linear combination of the dependentvariables X given below:

Y = b0 +N∑i=1

bixi (2.1)

Where N is the number of features used for the particular experiment and X = x1, x2, ..xN . HereB = b1, b2, .., bN represent the regression coefficients.

23

Figure 2.5 Comparison of accuracy results of our technique (1st column) with other state of the art

techniques of Montjoye et al [26]. (2nd column) and Staiano et al [78] (3rd column).

2.2.3 Evaluation and Results

We calculated the accuracy of our predicted personality traits using the testing data of big five per-sonalities (via psychometric tests) available in the dataset described before. The results achieved by thenewer features we proposed via this study i.e. accuracies of Agreeableness ≈ 76%, Conscientiousness≈ 62%, Emotional/Neuroticism ≈ 68%, Extraversion ≈ 61% and Openness/Imaginativeness ≈ 59%

were in most cases better than the state of the art, however illustrated lesser accuracy than those achievedvia the lexical features from social media. We have curated these results together by means of Figure2.5.

2.2.3.1 From Correlation Analysis

We used standard techniques of statistics to understand the relationships between Big-Five traits andmobile phone usage. We begin by analyzing intra-feature and feature-trait dependencies with respect toour regression analysis results. We discuss them here based on past literature and statistical significance.

24

1. Bluetooth Logs: Several features obtained from this data were found to significantly explain thevariance in traits prediction. It was observed that individuals scoring high on openness and lowon agreeableness had larger number of unique IDs seen per day(p < 0.05, t = 2.53), but theywere less likely to have seen some IDs significantly during the day and during odd hours. Thebeta values for conscientious and extraversion were found to not contribute significantly to theregression function which is in line with a previous study [21]

2. SMS Logs: The features corresponding to the SMS logs were found to contribute significantly tothe accuracy parameters we obtained for the different traits. Number of incoming and outgoingmessages for agreeable persons was observed to be quite high (p < 0.05, t = 2.33). On theother hand, the unconscientious (p < 0.05, t = −2.29), low on openness (p < 0.05, t = −3.54)participants were seen to have sent more messages than received.

3. Voice Call Logs: The features corresponding to missed calls (p < 0.05, t = 0.19) did not suf-ficiently impact the variation in the traits. Open, neurotic users who were disagree- able andextravert were more likely to use SMS as compared to voice calls. Additionally, it was found thatthe number of unique contacts associated with outgoing calls was also more likely to be higherfor extraverted, agreeable and non- conscientious users.

2.2.3.2 From Regression Analysis

Our study lends further support to the relatively unexplored field of predicting personality from datacollection restricted to only mobile phone logs. Using a set of features based on past personality researchalong with others that we developed, we were able to predict whether the participants were low or highor average on each of the Big-Five traits much better than random. In fact, to our knowledge, thesepredictions outperform previous research [26][11][14][2] in this specific field (as shown in the Figure2.5).

Interestingly, Agreeableness and Conscientiousness were the traits that received the highest accura-cies in our study with 80.95% and 66.67%. These are the traits that are associated with the ability ofpeople to be able to work with others. In particular, they deal with the self-discipline and dutiful action.This leads us to hypothesize that our features deal with the impulses behind actions of individuals. Ap-plying a greedy method (similar to the one in [53]) to choose those subset of features that lead to betteraccuracies, we also evaluate the correlations between features. The Pearson coefficients between theBluetooth and SMS features (eg. Average no. of unique IDs seen per day and Average no. of IDs seensignificantly - 0.79) depict that these are strongly correlated and indicative of the traits.

While Average Number of Unique IDs seen by a person could substantially quantify the number ofpeople our subject came in contact with throughout the day, we wanted to tap in deeper and find out whatnumber of people the subject chose to come in contact through the day. This would thus illustrate if theperson was meeting a bigger group of colleagues/friends or a smaller one on daily basis. This, in ouropinion, was a significant contributor in judging the extraversion and agreeableness the person had with

25

other people. Why these features, among others, are good predictors of personality is an interesting pointof investigation. It can possibly lead to contrasting studies with other models such SVM to measure theasymmetries between them [38]. Research on large-scale datasets is needed to test the robustness ofour results, but we believe that this is an exciting area of research to enable non-intrusive personalityinvestigations.

2.3 Summary and Conclusions

In the aforementioned part of this thesis we attempted to capture the lexical personality of an in-dividual using the Big Five personality traits as the baseline ontology that represented the personalityof an individual. We also tried to explore alternative methodologies and features so as to capture thispersonality. We discovered that while there may be numerous other tools to achieve insight into thelexical persona of an individual, the most effective were lexical cues to predict the most fundamental ofall personality models.We also realized that while big five may be an important personality model which has prevailed overseveral years, it only limits to assigning a static overall personality to an individual over a range of time.This is opposed to a multitude of studies which suggest that personality evolves with time [76][75].Thus, at any given time, t instance, the personality p demonstrated by person A might not be the sameas that demonstrated by A at t+ n (where n can be any time +t/− t.)Our work hereon thus revolves on exploring in-depth these insights. We worked on developing a dy-namic model to represent the personality of an individual. Further more we populated this model withlexical cues from social media so as to render the model predictive.

26

Chapter 3

Beyond the Staticity of Big Five - Modeling the Dynamic Psyche of an

Individual

Present theorizations and work undertaken in the domain of lexical personality and psychology ofan individual are dominated by state of the art models such as Big 5 [35] , MBTI [63] etc. While thesemodels have been insightful in understanding various traits of user persona and psychology throughvarious lexical markers from their social media profiles, they remain static in nature. For instance, amultitude of studies which have attempted to capture the Big 5 traits from various social media websitesof an individual over time, are successful in assigning a specific personality type to each user [33][67]. Contrary to this practice theoretical psychology along with its empirical validations states that anindividuals persona evolves in time with respect to their situations, context, age, cultures and many suchstimuli (both external and internal) [56] [76]. Thus, capturing an overall psychology of an individual asopposed to a more realistic and dynamic one results in a severe loss of information about the evolutionof this psychology. By means of the following work we introduce and describe a theoretical model(Buddhist Model of Psychology as proposed by Abhidhamma Traditions) which is adept in handlingthis shortcoming of our existing models of psychology (enabling us to achieve a lossless compression oflexical psychology in it’s computational formulations). We also attempt to draw various parallels of thismodel with Western Psychology, functioning of Bio-medical Processes and logics of Phenomenology.An in-depth discourse that follows in Section 3 of this chapter helps us to validate the salience of thismodel as opposed to any other model in our present knowledge.

3.1 Theoretical Psychology by Abhidhamma

3.1.1 Formal Ontology of an Individual by Abhidhamma Traditions

Formal systems and machines are inherently isomorphic. “Provability” of the existance of an inputin a formal system is equivalent to “Computability” of the same on a machine simulated in accordanceto this system. Thus, isomorphically if computers could think their brain states would match the se-quential ones in their formal counter-parts. Provability of an individual’s psyche in this formal system,

27

Figure 3.1 Represents a conceptualization of the mental states that populate the model of an individual.

would thus ensure its computability in a machine. Abhidhamma provides us a framework to formulate amachine adept at ensuring provability of a thought process/personality trait/psyche inputed into it. Ab-hidhamma is also the only doctorine that provides us with a model capable of capturing these subsequentstates of human thought (and psychology) while also maintaining their temporality. The schema pro-posed by Abhidhamma is renderable into a formal model (as illustrated in the following chapters) andthus can be furthered to contrive various computational systems. It also preserves the structural proper-ties of a formal model of constancy across the varied range of content parsed through it. This makes it aubiquitous model, true across data from a multitude of sources. Thus, Buddhism encompasses the ideaof a formal model by means of its ’moment by moment’ explanation of individual’s psychology. Wehave attempted to explore this further by populating these states by the categorized lexical data madeavailable through social media.

Abhidhamma scholarship in Theravada Buddhism [58] has long deliberated on the mechanisms ofreality. In the Abhidhamma both mind and matter, which constitute the complex machinery of man,are microscopically analyzed. The analysis provide descriptions of sentient experience as a successionof physical and mental processes that arise and cease subject to various causes and conditions. Thesesequential processes (mental states) formulated as discrete, momentary events are referred to as tropes(defined as dhammas in the original text). Tropes are thus seen as psycho-physical events that providemental cognitive awareness. The doctrine also presents the concept of a moment (khana) which is akind of synchronic duration of each such event. These moments are made up of tropes. In this sense,

28

Figure 3.2 Represents a conceptualization of the model of an individual. It handles the temporal charac-teristics with the sequencing of subsequent moments one after the otherM i+1 followsMi. It also handlesthe sparsity of data by means of “Sleep States” (Translucent Blue Bubbles) accommodated amongst thepopulated states (Dark Blue Bubbles). Finally it attempts to illustrate the varied granularity which canbe manifested in the respective moments. For instance multiple blue moments make up bigger megamoments (grey). This phenomenon can be witnessed both top-down and bottom-up.

Abhidhamma visualizes the time scale of these mental/physcical processes so they are now seen asoperating from moment to moment. The Abhidhamma thus attempts to provide an exhaustive accountof every possible type of experience, every type of occurrence that may possibly present itself in one’sconsciousness in terms of its constituent tropes [24]. Further, the doctrine provides a taxonomy of tropesand their relational schema whereby each acknowledged experience, phenomenon, or occurrence canbe determined and identified by particular definition and function. There are two kinds of tropes thatconstitute reality according to this doctrine - ultimate tropes (paramattha dhamma) and conventionaltropes (samutti dhamma). Conventional tropes are complexes constituted by ultimate tropes and includesocial and psychological reality. Ultimate tropes are organized into a fourfold categorization. The firstthree categories include 1) the bare phenomenon of consciousness (citta) that encompasses a single tropetype and of which the essential characteristic is the cognizing of an object; 2) associated mental factors(cetasika) that encompasses fifty-two trope types; and 3) materiality or physical phenomena (rupa) thatinclude twenty-eight trope types that make up all physical occurrences (including the human bodilyones). The fourth category that neither arises nor ceases through causal interaction is nibbana. Forour conception of modeling an individual based on Abhidhamma, we build a discrete line of moments,wherein each moment stands for a consciousness trope or citta. An individual is then conceived as aformal arrangement of these conscious tropes on a discrete line. This line of moments compulsivelypasses to the next moment as a result of previous cognition and action. Each moment has 2 categoriesof tropes embedded in it. 1) cetasikas or mental factors related to the cognition and 2) rupa or materialcognition and actions (referred to as Objects in this work).

29

3.1.2 Comparing Abhidhamma with Other Psychological Doctrines

By means of this section we intend to compare Abhidhamma with the doctrines proposed by var-ious other psychological schools of thoughts (primarily the originating ideas of Western psychologythrough philosophers of the likes of Brentano etc.). We also attempt to explore the sanctity of thismodel by drawing parallels of its conception with Mathematics, more broadly with the idea of trainedintrospection. We primarily attempt to percieve the model in contrast with some of the major subclassesof the conception of psychology i.e. Western Thought, and Phenomenological Meditations. We alsodiscuss the importance and usefulness of this model as opposed to any other model for psychologicaltheorisations.

3.1.2.1 Abhidhamma and Brentano’s Philosophy

Franz Brentano was an influential German philosopher and psychologist whose work strongly in-fluenced psychologists like Sigmund Freud, Kazimierz Twardowski, Alexius Meinong, Carl Stumpf,Anton Marty, Christian von Ehrenfels, and Tom Masaryk. He was one of the pioneering thinkers wholaid foundation to the modern western psychology. While Brentano proposed various concepts in hiswork Psychology, his major goal was to lay a foundation for a the “Science of Mental Phenomenon”.By means of this idea he intended to capture the science of psychology (mental phenomenon) and dif-ferentiate it effectively from that of the physical phenomenon. Amongst the criterions he proposedto distinguish these two, the most fundamental of these, which can also be traced into the theory ofAbhidhmma are as follows:

• Mental phenomenon are exclusive object of inner perception : Inner perception is often timesengendered by the mental phenomenon that it perceives from the previous mental states. Sucha pattern (as also described by Abhidhamma) leads to a trail of uninterrupted perception risingsequentially from the chain of previous mental phenomenon. Brentano primarily argues herethat unlike physical phenomenon, mental phenomenon are perceived solely by means of innerperception. A similar distinguishing is also observed in the Abhidhamma theory, which arguethat the action is the sole communicator of physical phenomenon. It is the feedback of theseactions into the inner perception (i.e. the way in which we perceive our or other’s actions) whichlater alter the mental phenomenon.

• Mental Phenomenon always appear as unity: Much like Abhidhmma which argues that any mo-ment is divisible into various other atomic moments (as defined in Section 3.1.1), Brentano tooconveys a similar idea. He argues that, while we can perceive various physical phenomenon in agiven moment, we can only perceive one mental phenomenon at any given time. To exemplify,when we seem to have more than one mental phenomenon occurring at a given time, for instance,enjoying a musical melody while working. All the atomic mental phenomenon are perceived asthough melded into one, they become moments or in Brentano’s words, divisives of a collective.

30

Even when one of these divisives comes to an end, for instance, I stop studying, the collective stillcontinues to exist. One can perceive this collective to exist with nothingness. Thus, convergingthis theory into that propagated by the Abhidhamma tradition of Buddhism. The structure of ourunitary moments as defined by Buddhism (or states in the perceived finite state automaton) havethe ability to combine, or meld together into bigger moments, thereby combining the dynamicelements and revealing connects beyond the unitary level. Their transcendence into a state devoidof any mental states gives rise to a continuum of existence called bhavanga. Thus, not only doesBuddhism has a similar idea of mental states and their existence, it also has defined terminologywhich parallels the foundational theory of western psychology. Brentano also suggests that whileone can remember the mental state one was in a moment earlier, or expect future mental states,one cannot have two simultaneous mental states. The theory of Abhidhamma resonates this viewby suggesting that the mental states of each given moment are affected by those in the previousmoments, their resultant actions and the possible future mental states.

• They are always intentionally directed towards an object: Brentano also ascertains that each ofthese mental states we experience in a moment can be directed towards one and the same objectin different ways. There are various ways in which he defines that this contact occurs. This ideais resonant to the idea of the mental states in each moment being directed to the physical/mentalobjects from the previous moment. Brentano along with Husserl also furthered their work on‘intentionality’ by diversifying the explicit differences between intrinsic and extrinsic intentionsby means of Teleology1. The differences illustrated by these philosophical concepts have alsobeen captured by means of the various mental states covered in Buddhism (Figure 3.1). This is onethe most important parallels that we also use to capture deep rooted psychological phenomenonsuch as anxiety etc. are further elaborated in Chapter 4.

3.1.2.2 Abhidhamma and the logics of Phenomenological Thought

By means of this section we intend to draw parallels between the conception of Mathematical logicsand that of the Buddhist Theories of Abhidhamma. Phenomenology (from Greek phainomenon - “thatwhich appears” and logos - “study”) is the philosophical study of the structures of experience andconsciousness. Phenomenology studies conscious experience as experienced from the subjective or firstperson point of view. As elucidated by various works through time [40][82] while Mathematics doesbase its footing on logical derivations and calculations, the validation of these proofs require mentaljudgement. A subjective inference based on the structures of those calculations and proofs is what helpsus to define the various constructs of Mathematics. Thus, much like what was emphasized by Husserl,Godel, Bernays and more of their contemporaries, Mathematics is perceived as a phenomenologicalenterprise.

1Teleology is the philosophical study of nature by attempting to describe things in terms of their apparent purpose, directiveprinciple, or goal. A purpose that is imposed by a human use, such as that of a fork, is called extrinsic.

31

The Psychological Theory of Abhidhamma as explained in Section 3.1 and Figure 3.2 can also beviewed as a static finite state machine, moving from one moment state to the next by a set of welldefined rules. While the conception of these rules of translation can be attributed to trained meditativephenomenological conceptions, the model itself is in line with the formal methodologies of stochasticfinite state machines [66]. We can thus speculate that just like logic is nothing but meditative conception,the theory of recognizing the contact and transition of one mental state into another is also achieved bytrained meditative phenomenology and lies in the likes of logic.

3.1.2.3 Abhidhamma and Empirical Cognitive Processes (Bio-Medical Processes)

Apart from the above defined parallelisms, we also sought to study the theory of Buddhism withrespect to various scientific explanations of cognitive processes. One such phenomenon is that of dis-crete consciousness. As opposed to some traditional school of thought, many researchers advocated thepresence of consciousness as being composed to various atomic units, similar to the idea of momentsin Buddhism. Some empirical evidence of this was illustrated by means of the work undertaken byLehmann et. al. [50][49][47]. In this research, the investigators studied multichannel EEG by making amap of the isopotential lines. It was observed during this analysis that the “landscape” made a suddenchange every 100 milliseconds. Not only studies relating to the structure of consciousness but also thoserevolving around the structure of object perception and subsequent translation of one mental state intoanother have been undertaken. For instance, the “Synchronicity Hypothesis” of Von Der Malsburg et.al. [86][27] states that so as to achieve a representational object in the brain, a population of neuronsare fired simultaneously onto the physical object in consideration. The study then goes on to explain theserial way in which this object is perceived. This research has a stark resemblance to the flow of mentalstates which facilitates the perception of an object in the Abhidhamma theory of psychology.

Abhidhamma also states that there are primarily seven steps to achieve the successful conception ofan object. These queue of seven “Universal” mental states (illustrated in Figure 3.1) help us in success-fully grasping an object. These occur in the order of :

Contact→ Feeling → Perception→ V olition→ Intention→ Concentration→ Attention

Thus, we can hypothesize that the sequence of this order preserves the inferences we have from theabove stated bio-medical empirical processes. In keeping with the above explanations and parallelswith various established theories in the field, we have in the best of our understanding established thatBuddhism’s Abhidhamma is one of the few theories that comes close to preserving both the traditional,modern and the empirical standpoints on an individual’s psychology. It is important to note that thescope of this research, limits us to only explain the expressed mental states of an individual (and in turntheir lexical psychology as captured by these expressed mental states from Social Media platforms) as

32

opposed to the felt/unexpressed mental states. We undertake further discussion on our idea of “LexicalPsychology” in the subsequent section.

3.2 Stochastic-Computational Modeling of Lexical Psychology using Ab-

hidhamma’s model of Individual’s Psychology

As discussed in the previous section, we adapt the Buddhist Model of Psychology and attempt tomathematically define it, so as to use it to computationally capture the psychology of various individualsfrom their social media data. The depiction of consciousness as a single line of moments helps us incasting the Abhidhamma model of an individual as a stochastic automaton. We attempt to comprehendthe stream of consciousness of an individual as transitions in this automaton from one state to another.We thus start with a brief introduction of Automaton theory, followed the work we undertook in orderto construct our model and automate it computationally.

3.2.1 Automaton Theory

Automatons are abstract models of machines that perform computations on an input by movingthrough a series of states or configurations [41]. At each state of the computation, a transition functiondetermines the next configuration on the basis of the present configuration. As a result, once the compu-tation reaches an accepting configuration, it accepts that input. The major objective of automata theoryis to develop methods by which computer scientists can describe and analyze the dynamic behavior ofdiscrete systems. Characteristics of such machines include:

• Inputs: assumed to be sequences of symbols selected from a finite set I of input signals. Namely,set I is the set x1,x2,x3 ...,xk where k is the number of inputs.

• Outputs:sequences of symbols selected from a finite set Z. Z is the set y1, y2, y3..., ym where mis the number of outputs.

• States: finite set Q, whose definition depends on the type of automaton.

There are 2 types of basic automatons - Deterministic, Non-Deterministic that can be extrapolatedinto a stochastic automaton. The deterministic automaton is defined by it’s lexical meaning which meansthat after a state takes an input, there is one and only one state (determined state) to which the automatoncan transition from this state. On the contrary, in a non-deterministic finite state automaton, the nextstate transition can go to none, one or more possible states.

33

Figure 3.3 DFA accepting all strings with a substring 01, from [41]

Definition 1. Deterministic Finite State Automaton (DFA: A Deterministic Finite Automaton is afive tuple consisting of:

1. A finite set of states, often denoted Q

2. A finite set of input symbols, often denoted by Σ.

3. A transition function that takes as arguments a state and an input symbol and returns a state.The transition function is commonly denoted by δ.

4. A start state, one of the states in Q.

5. A set of final or accepting states F . The set F is a subset of set Q.

A deterministic finite automaton is often referred to by its acronym DFA with the notationA = (Q,Σ, δ, q0, F )

where A is the name of the DFA, Q is its set of states, Σ its input symbols, δ its transition function, q0its start state and F its set of accepting states. An example of a DFA accepting all binary strings with asubstring “01” is as shown in the Figure 3.3.

Figure 3.4 An NFA accepting all strings that end in 01, from [41]

Like DFA, a Non-deterministic Finite Automaton (NFA) has a finite set of states, a finite set ofinput symbols, one start state and a set of accepting states. The difference between the DFA and theNFA is in the type of the transition function δ. For the NFA, δ is a function that takes a state and inputsymbols as arguments (like the DFA’s transition function), but returns a set of zero, one, or more states(rather than returning exactly one state, as the DFA must). Figure 3.4 is an example of an NFA acceptingall strings ending with “01”

A Probabilistic Automaton (PA) is a generalization of the NFA. It is inclusive of the probabilityof a given transition into the transition function, turning it into a transition matrix or stochastic matrix.

34

Thus, it is essentially a general automaton, wherein the next transition state is a probabilistic choice overseveral next states.

Definition 2. Probabilistic/Stochastic Automaton: A Probabilistic/Stochastic Automaton is a systemPA = (Q,Σ, δ, q0, F ), where:

1. Q = q1, q2, q3..., qn is its set of states.

2. Σ its input symbols of alphabet.

3. q0 its start state.

4. δ its transition function from Q × δ into [0, 1]n+1 (the transition probabilities matrix) suchthat for (q, σ) ∈ Q× Σ, δ (q, σ = (p0(q, σ), ..., pn(q, σ))

where, 0 ≤ pi(q, σ),∑n

i=0 pi(q, σ) = 1

5. F its set of accepting/final states.

Our model of an individual finds its footing in the idea that the human psychology is inherently non-deterministic. It is governed both by the intrinsic personalities and external stimuli. The situations thatindividuals find themselves in render a certain probability to each of the subsequent states of being. Wefurther discuss this in detail through the following section.

3.2.2 Stochastic Finite State Automaton for an Individual

In this section we formally define a stochastic automaton of an individual based on the conceptionof a formal model of individual as described in the previous section. A central concept to this doctrineis that, there is a total ordered temporal sequence of moments that captures the consciousness of anindividual. We model this sequence of moments as states of a finite state automaton. Each state is atemporal moment defined in terms of the mental factors and actions embedded in it. This embedding ofa particular set of mental factors and actions in each moment is defined through transition functions ofthe automata. Upon this basic architecture, to populate each moment as a bag of word representationfrom individuals web data, we write stochastic processes to help in modeling, predicting and refiningrules governing the persona of the individual.Formally speaking we define our automaton as a finite state machine. Let Q = {Q1, Q2, Q3 . . . } be aset of symbols that represent moment states, A = {A1, A2, A3 . . . } be a set of symbols that representactions and material cognition, and T = {T1, T2, T3 . . . } be a set of symbols that represent the mentalconcomitants of an individual. We define our stochastic automaton whose internal state space is Q andwhose input and output spaces as a Cartesian product AxT .

I(r, f) = {Q,A, T, r, f, (f, r, ., .),M(f, ., .), AT (r, ., .), E}

r ∈ [0, 1])D, f ∈ [0, 1]

35

AT : [0, 1]D ×Q×A× T → [0, 1]

AT (ri, Qi, Aj × Tj) :

Probability that the output is Aj × Tj when the internal

state is Qi.

It is important to note here that which Q (a moment state) is an embedding of A (action and materialcognition) and T (mental concomitants of the social machine), it’s structure varies by means of it’stemporality and the personality/persona (f, r) of an individual.

M : [0, 1]× (A× T )× (A× T )×Q→ [0, 1]

M(f,Aj × Tj , Ak × Tk, Ql) :

Probability that the next moment state is Ql when the input

is Aj × Tjand the output is Ak × Tk

E(∈ Q) : Halting state

i.e. when the moment state moves on to *empty state*

π(f, r,Qi) :

Probability that the initial state (after *empty state*) is Qi

n∑j=1

AT (r,Qi, Aj × Tj) = 1

m∑l=1

M(f,Aj × Tj , Ak × Tk, Ql) = 1

m∑l=1

π(f, r,Ql) = 1

Here, f represents the personality parameter and r represents the attitude of the given individual towardsan object for output.Let m(t) ∈ Q be a moment state at any discrete time ’t’, at out(t) be any output set of A&T at time’t’ and at in(t) be any input set of A&T at ’t’. Then the relation m(t), at out(t) and at in(t) share isas follows:

Prob(em(0) = Qi) = π(f, r,Ql) (3.1)

Prob(em(t+ 1) = Qi) = M(f, at in(t), at out(t), Qi)

36

Figure 3.5 A stochastic finite state machine representation for the model of an individual.

Prob(ac out(t) = Aj × Tj) = AT (r, em(t), Aj × Tj)

Let TRMk(f, r) ∈Matm(R) be the state transition probability matrix in the case the input is Ak×Tk.From (1), we can get TRMk(f, r) as follows:

TRMk(f, r) = (trmk(f, r)ij) ∈Matm(R)

trmk(f, r)ij = Prob(Ei → Ej |input = Ak × Tk) (3.2)

=

m∑l=1

AT (r, f,Qi, Aj × Tj).M(f,Ak × Tk, Aj × Tj , Ql)

TRMk(f, r) = AT (r)Mk(f)(k = 1, 2, 3, . . . ),

AT (r) ∈Mat(R : m,n),Mk(f) ∈Mat(R : n,m),

(AT (r))(ij) = AT (r,Mi, Aj × Tj),(Mk(f))(ij) = M(f,Ak × Tk, Ai × Ti,Mj)

Here, Matij is the (i, j) component of the matrix Mat. Therefore, upon availability of individual data(lexical data), the state transition matrix can begin to learn and refine the probabilities of our individualautomaton going from one state to another. With a good volume of data, it also accommodates forpredictive analysis of the psychology of an individual.

37

3.2.3 Stochastic-Computational Lexical Model (PACMAN)

By means of this section we aim to present the machine learning module which will help in tran-scending the above defined theoretical constructs (of an evolving personality) into a usable model forpersonality observation and prediction. This model is envisioned to be competent in handling lexicalfeature population (to different subsequent states) so as to eventually predict the personality of an in-dividually temporally, given any social media data (Facebook statuses here) in any specific range oftime. We also handled the issue of sparsity of data by modeling the atomic units of this model homo-geneously with the bigger molecular ones constituted by these unitary units. For instance, while themodel can handle descriptive data from a user daily for (say) a week, it is also adept in handling a datapoint describing (say) ’childhood’. Figure 3.2 illustrates this concept by the terminology of momentsand mega moments. We empirically verify our model using the dataset described in section 3.1. Thefollowing section 3.2.2.1 illustrates the methodology used to train a multi-label classifier to predict the40 mental states (Figure 3.1) populating moment-by-moment states of an individual (here, consecutiveuser statuses on social media). Our aim is to use these mental states as descriptors of the change in userpersonality over time.

3.2.3.1 Dataset Used

We used the myPersonality [18] dataset. It is a sample of personality scores and Facebook profiledata that has been used in recent years for several different researches [7]. It has been collected by DavidStillwell and Michal Kosinski by means of a Facebook (FB) application that implements the Big5 test[23], among other psychological tests. The application obtained the consent from its users to recordtheir data and use it for the research purposes. We randomly picked a set of 50 users from this dataset(who had more than 20 status updates) to analyse and validate PACMAN. As the first attempt to solvethe problem of evolving personality explicitly, we contributed a dataset named PACMAN Dataset 2, withannotated statuses of users. These annotations cover all the 40 mental states discussed by Abhidhammaas described above. So as to achieve an extensive and unbiased set of annotations, we had a set of 3independent annotators tag each of the FB statuses of a random user in our dataset with a set of relevantmental factors (out of the 40 factors suggested in Figure 3.1). We then computed the MASI (MeasuringAgreement in Set-Valued Items) to evaluate the disagreement amongst these annotations. Given twosets, A and B, the formula for MASI is:

1− JA,B ×MA,B

where J is the Jaccard metric [12] for comparing two sets: a ratio of the cardinality of the intersectionof two sets to their union. M (for monotonicity) is a four-point scale that takes on the value 1 when twosets are identical, 2/3 when one is a subset of the other, 1/3 when the intersection and both set differencesare non-null, and 0 when the sets are disjoint. MASI ranges from zero to one. It approaches 0 as two

2https://researchweb.iiit.ac.in/˜shivani.poddar/PACMAN_Dataset

38

sets have more members in common and are more nearly equal in size. An average value of 0.376 ascomputed in Table 3.1 reflects that the sets of the labels under consideration are a close intersection ofone-another. For the purpose of training and testing our model (PACMAN), as expatiated in section3.2.2.4, we split this annotated dataset of 101 statuses into two parts. The first 50 were used for trainingthe model, and the next 51 to test the accuracy of the trained model.

G↔ A G↔ B A↔ B Avg. MASI0.306 0.386 0.435 0.376

Table 3.1 Inter-annotator values. G is the labeled data by annotator 1, A is the labeled data set fromannotator 2 and B is the labeled data set from annotator 3.

3.2.3.2 Pre-Processing Data

To preprocess the user-focused data available to us from the myPersonality dataset [18], we extractedeach individual’s data based on the unique authentication ID provided in the dataset. This data is inclu-sive of the statuses posted by the user, the dates of these posts and his/her Big-Five3 personality traits.For our analysis we chronologically sorted these extracted FB statuses based on their correspondingdates. By means of language filtering, we then processed this dataset to retain only those statuses thatwere using English language (as opposed to multi-language statuses). So as to feed the statuses intothe LIWC API (described in section 3.2.2.3), we were then required to also (for an improved analysis)determine the gender of the given user. We extracted Pronouns (such as herself, himself, hers etc.)from the Stanford POS tagger and mapped these pronouns to their respective gender usage as definedin English Language. This helped us to heuristically determine the grammatical gender of each usereffectively. We left the gender of the users without any gender specific pronoun usage to be 0 in theLIWC API.

3.2.3.3 Features Used

Feature extraction from short texts such as FB statuses, requires extensive linguistic analysis. Soas to achieve an effective feature generation, we leverage the psycho-linguistic tool, Linguistic Inquiryand Word Count (LIWC) [64]. It is adept in reflecting the various features relevant to the linguisticand psychological processes of a user in the context of social media (meaning shorter texts and morenoise pertaining to faulty usage of English). The LIWC API includes a text analysis module alongwith a group of built-in dictionaries. The dictionaries are used to identify which words are associatedwith which psychologically-relevant categories. These categories include psychological features such as

3These were mined to illustrate comparisons of our dynamic model versus the static personalities proposed by Big Five

39

Analytic thinking, Emotional Tone, Social words and Informal speech as well as linguistic features suchas Functional Words, Personal Pronouns and Punctuation. We use this API to extract the respectivepsycho-linguistic features for each FB status of a given user. To enhance the predicted 180 LIWCfeatures by means of this API, we also specified (as additional parameters) the content-type as “SocialMedia” and the user’s gender obtained via pre-processing.

3.2.3.4 Multi-Label Classification - Binary Relevance Method

Determining 40 mental factors from a linguistic unit, such as a Facebook status, can be cast as amulti-label classification problem. Traditional single label classification problems are concerned withlearning from a set of examples that are associated with a single label l from a set of disjoint labelsL, |L| > 1. If |L| = 2, then the learning problem is called a binary classification problem, while if|L| > 2, then it is called a multi-class classification problem. In these problems each of the examplescan be assigned one or more labels from L. Formally speaking, Consider S as the training set withN examples Ei = (xi, Yi), where i = 1..N . Each instance Ei is associated with a feature vectorxi = (xi1, xi2, ..., xiM ) and a subset of labels Yi ⊆ L, where L = yj : j = 1..q is a set of q possiblelabels. Considering this scenario, the task of a multi-label learning algorithm is to generate a classifierH that, given an unlabeled instanceE = (x, ?), is capable of accurately predicting its subset of labels Y, i.e., H(E)Y , where Y is composed by the labels associated with the instance E (from [20]). So as totrain a classifier H for our multi-label classification problem we used the Binary relevance(BR) methodwhich belongs to the problem transformation category of methods for multi-label learning. The BinaryRelevance method decomposes a multi-label classification problem into several distinct single-labelbinary classification problems, one for each of the q labels in the set L = y1, y2, ..., yq . This approachinitially transforms the original multi-label training dataset into q binary datasets Dyj , j = 1..q, whereeach Dyj contains all examples of the original multi-label dataset, but with a single positive or negativelabel related to the single label yj according to the true label subset associated with the example, i.e.,positive if the label set contains label yj and negative otherwise. After the multi-label data has beentransformed, a set of q binary classifiers Hj(E), j = 1..q is constructed using the respective trainingdataset Dyj . This set of q classifiers that are constructed by means of the BR approach can be summedas follows -

HBR = Cyj((x, yj))→ yj ∈ 0, 1 | yj ∈ L : j = 1...q

To classify a new multi-label instance, BR outputs the aggregation of the labels positively predictedby all the independent binary classifiers. We used the above described BR approach to train our multi-label classifier H . For training purposes, we transformed the extracted LIWC features for each status asa Mi,j matrix, where i ∈ (0, Length of LIWC feature vector f ) and j ∈ (0, No. of FB statuses of eachuser). We then appended this matrix with a set of 40 columns each that represented each of the mentalfactors mf by a value of 0 (for marking absence of mf ) and 1 (for the presence of the mf ).

40

f1f1f1 f2f2f2 .. f180f180f180 mf1mf1mf1 mf2mf2mf2 .. mf40mf40mf4025.7 0.2 .. 7 1 0 .. 1

Table 3.2 Row of MatrixMi,j which was used for training and testing the Multi-Label Classifier Model.

Adapting to the Problem Transformation Method [83], this problem is then approached as a jointset of binary classification tasks. These are expressed with label binary indicator array: each sample isone row of a 2D array of shape (nsamples, nclasses) with binary values: the one, i.e. the non zeroelements, corresponding to the subset of labels. Using the binary relevance approach, we then use theOne-vs-All SVM classifier to discriminate the data points of one class versus the others. Since our labelsare not exclusive this works well for us since each classifier essential would answer the question “Doesit contain mental state x ? ” and so on for all x ∈ (total 40 mental states). A brief representation of theMi,j matrix is as illustrated in Table 3.2.

3.2.3.5 Evaluation and Results

Since each instance in the multilabel data is not a single label but a vector of different labels, estab-lished evaluation metrics such as accuracy, precision-recall, f-measure etc. cannot be used directly [31].Based on the learning problem we are addressing, Hamming loss was used instead. It can be defined asthe fraction of the wrong labels to the total number of labels, i.e.

HammingLoss(xi, yi) =1

|D|

|D|∑l=1

xor(xi, yi)

|L|

where |D| is the number of samples, |L| is the number of labels, yi is the ground truth and xi is theprediction.

The average value of Hamming Loss is 10.455, which means that approximately every 10 out of 100labels are predicted wrongly. Our results show that we can analyse and predict the evolving mentalstates contributing to the composition/change in the personality of an individual. We can also predictmental states to within just over 10%, a resolution that is likely fine-grained enough for many applica-tions. A loss of 0.1 labels in a dataset which is being analysed moment by moment will not have manyimplications in various practical applications of our framework.

3.2.3.6 Discussions

Since this research relied heavily on studying the mental states w.r.t Buddhist tradition of Abhid-hamma, we define our heuristics for these analysis inspired by the same doctrine. Figure 3.6 illustrates

41

Figure 3.6 Maximum Moments reflect the number of statuses which have a consecutive set of similarmental factors.

the moment by moment representation of 40 mental states for the given users from their FB statuses.The mental states unlike the Big 5 traits of these users evolve over time. While few can be mapped toappear occasionally, some mental states also appear inherently in the users. The states are representedby the following values: 0: Aspiration, 1: Discursive Thinking, 2: Effort/Energy, 3: Desire, 4: DecisionMaking, 5: Greed, 6: Hate, 7: Dullness/Wavering, 8: Error, 9: Selfishness, 10: Worry, 11: Con-ceit/Pride, 12: Envy, 13: Shamelessness, 14: Recklessness, 15: Restlessness, 16: Sloth, 17: Torpor, 18:Skepticism/Doubt/Perplexity, 19: Generosity, 20: Faith/Confidence, 21: Discretion, 22: Equanimity,23: Tranquility, 24: Lightness, 25: Adaptability, 26: Elasticity, 27: Proficiency, 28: Right Speech, 29:Right Action, 30: Right Livelihood, 31: Wisdom, 32: Goodwill, 33: Insight, 34: Sympathetic Joy, 35:Compassion, 36: Ignorance, 37: Attentiveness, 38: Modesty, 39: Uprightness, 40: Interest. Tappinginto the dynamic nature of user persona would require us to study the different persistent mental stateswhich dominate various timespans in a user’s stream of consciousness, changes in these mental states,and finally new emerging mental states. Drawing on these research ideas, the work also chalks potentialin the field of studying external situational conditions which affect the presence and frequency of cer-tain mental states affecting user personality. By means of this section we attempt to present a two-foldanalysis. Firstly, elaborate on the in-depth insights per user for a small subset of users (4/50 users) thatwe analysed as a part of this study. Second, discuss some inferences which we found salient for all the

42

users we analysed by means of a bigger subset (50 users).

As illustrated in Figure 3.6 we present the maximum number of occurrences of particular mental factorsin consecutive moments. This provides a statistical estimate of the number of times a mental factor hasto occur to become part of a personality trait. Over a sample of 50 users and their FB data over an year,we find this estimate to average out at 6.02 moments (say λ is approximately 6 moments). We use thisin identifying two important properties of the occurrence of mental factors leading to changes in per-sonality. First are the ones which occur consecutively upto λ are identified as ‘Inherent mental states’of an individual. These states represented as a bag of words form a static image of our individual’spersonality traits. Secondly, we further identify ‘dynamic mental states’ as those mental states thatoccur in bursts and are contributory to the evolving personality of an individual. From the sample data,we find that empirically the distance between the previous occurrence of a mental state and its currentoccurrence is at 1/5th of the total number of statuses/data points. The value was found by workingout the closest formula to the sequence of numbers that our dataset generated for dynamic mental statedistances4. With this value, for each of the four users we identify their dynamic occurring mental statesas described in Figure 3.6 and Table 3.3. For example, For User 6 we found the mental states: Deci-sion Making (4, Yellow), Greed (5, Orange), Sloth (16, Orange), Torpor (17, Yellow), to be persistentand thus contributing to their personality. Whereas states such as Worry (10, Yellow), Confidence (20,Blue), Equanimity (22, Yellow), Tranquility (23, Orange) occurring in bursts causing the dynamicallychanging attributes of his/her personality to vary.

Along with these individual analysis, an extensive exploration of the dataset of an additional 50randomly selected users helped us encounter some interesting findings that further helped in validatingsome of the claims made by the Abhidhamma tradition. For instance, the doctrine suggests that the un-pleasant and the pleasant mental factors occur exclusive of one another. The mental factors predicted bymeans of PACMAN adhered to this theory. For example, in one of the users (from the PACMAN datasetof predicted user states), while we did see an overall fluctuation in factors such as “faith” (belongingto pleasant mental factors) and “skepticism” (belonging to unpleasant mental factors), they never oc-curred at same instance (moment/status). Another interesting observation that can be made on the basisof the inherent and sporadic mental factors of all the users (like those covered in Table 3.3) are theco-occurrence of certain mental factors with one another. For instance, “sloth” always accompanies“torpor”, “selfishness” is usually present with an inherent state of “greed” and so on. Figure 3.6. is anillustration of the analysis of these basic phenomenon shown for 4 out of the 50 users we analysed.

4http://math.cmu.edu/ bkell/21110-2010s/formula.html

43

UserIDMaximum/Total

momentsPACMAN Inherent States PACMAN Dynamic States

Big-Five

(enaco)

User 1 5/101Desire(3, Yellow), Interest(40,

Orange)

Aspiration (1, Blue), Decision

Making (4, Orange)nynny

User 6 7/194

Decision Making (4, Yellow),

Greed (5, Orange), Sloth (16, Or-

ange), Torpor(17, Yellow),

Worry (10, Yellow), Confidence

(20, Blue), Equanimity (22, Yel-

low), Tranquility (23, Orange)

nnnny

User 7 5/215

Hate (6, Red), Decision Making

(4, Yellow), Sloth (16, Orange),

Torpor(17, Yellow), Restlessness

(15, Green)

Desire (3, Green), Shamelessness

(13, Violet)nynny

User 10 5/72Greed (5, Orange), Restlessness

(15, Green)

Sloth (16, Orange), Torpor(17,

Yellow), Envy (12, Red), Skepti-

cism (18, Red), Faith (20, Blue)

nynny

Table 3.3 Analysis of Mental factors, (enaco) tuple represents the Big Five traits in the order of Ex-

traversion, Neuroticism, Agreeableness, Conscientiousness and Openness, here y means that the trait is

present and n means it is absent.

In comparison to the state of art, we observe that while the Big Five characteristics (Col. 5 in Table3.3) of the user remain constant over the course of this year, PACMAN helps us in mining both theinherent and dynamic mental states (Col. 3 and 4 in Table 3.3) for the user’s persona. For instance, forUser 1, while Interest (40, Orange) might be an inherent mental factor for the user, we do encounter asudden change in the presence of other mental factors such as Aspiration (1, Blue), Decision Making(4, Orange). These are factors directly contributory (by definition) to one of the Big Five traits such asAgreeable defined to be absent in User 1 (this absence is perceived to be constant for the personality ofthe user). Various modern literature suggests that personality is a construct of various external stimuliand a different adaptive process for each one of us. Since, by most means the experiences we have arestarkly different from one another, our personalities are also varied. In keeping with this theoreticalfoundation, we observe that while the Big Five personalities for 3 of the 4 randomly chosen usersshown in Table 3 are the same, the PACMAN model accommodates different inherent and dynamicstates for each one of them. We witnessed such changes in all the users we analysed by means ofour experimentation. These time-spans which record the continuous presence of these dynamic mentalfactors (thus transcending them into inherent factors) have also been illustrated by means of Figure 3.4.

44

We analyse each individual and assert that any mental states which have a sustained cognitionfor more than a threshold x of the states is contributing to the personality of an individual. Soas to arrive at the threshold x, we analyse the temporal mental states of n individuals and work out theintersection of states which imply sustained mental states in a given time span. This time span would becontributory to the defining personality of an individual.

Figure 3.7 Maximum Moments reflect the number of statuses which have a consecutive set of similarmental factors.

Definition 1. Sustained Mental States (smc): Let m be the mental state that occurs contiguously insubsequent moments constituting the stream of consciousness of an individual. Then m can bedefined as a sustained mental state.

Definition 2. Threshold (λ): Threshold is the average of all the maximum moments of the all sus-tained mental states (smc, of the listed users). Here, empirically our threshold came out to beapproximately 6.02 moments.

Threshold(x) =

∑Ui=1

∑|smc|l=1 f(l)

smc

U

Here, U is the total number of users. f(l) is the function that returns the number of the maximumnumber of moments wherein l occurs contiguously. M is the total number of mental states for thegiven user.

Here, Figure 3.7 illustrates this calculation of the threshold λ for the 50 users. The average of all thevalues of y over x in the graph come out to be 6.02.


These results of our initial investigation in dynamic personality analysis from social media are en-couraging evidence which back the theoretical foothold of evolving user persona in psychology. Ex-tracting and modeling mental states from lexical resources are just the beginning of our explorationinto the plausible dynamics of personality change over time. By means of this study we attempt topropose an initial stochastic model of an individual, a theoretical foundation inspired from the Abhid-hamma theory of Buddhism to ascertain the transitional heuristics of the model (transition matrix and

45

so on), and a machine learning framework to populate and analyse the dynamic personality model ofa social media user. We believe that our model will eventually accommodate not only applications fo-cused on observing dynamic user persona, but also those which want to leverage from predicting userbehavior, mentality, actions, and thoughts. The dataset we contribute by means of this work, a firstannotated dataset for user mental states based on the Buddhist Model of Personality, would also be auseful resource helping researchers to conduct explorations in the domain. As a part of our future effortshereon, we worked on further extending and refining this model so as to understand and infer variouspsychological phenomenon and social applications.

46

Chapter 4

Applications of Stochastic-Computational Lexical Model of Psychology

There are various applications that we found where our model of lexical psychology (described inChapter 3) would be useful. Through the course of our research, we used PACMAN to: 1) proposea ubiquitous model of an individual for Social Machines (elaborated further in Section 4.1), 2) modelpsychological phenomenon such as Anxiety (Section 4.2) and 3) propose a Lexico-Psychological En-gagement Factor to capture the temporal mental engagement of users from Social Media Data.

4.1 Social Machines

The emergence of personal computing and internet has made everyone in the world potential tech-nological contributors, elevating the ever increasing interaction between humans and technology. Thereis a plethora of work studying systems and interfaces facilitating this interaction. For instance, Human-Computer interaction [16] as an area of applied cognitive science and engineering design is concernedwith understanding how people make use of devices and systems that incorporate computation. So-cial Computing as the area of Computer Science is concerned about examining the intersection of userbehavior and computational systems. It finds inspiration in creating or recreating social conventionsor contexts with software systems and technology. Personality Mining [29, 67] is a newer domain ofcomputer science which focuses on capturing the psychological processes and dispositions of an indi-vidual through the person’s publicly available data. Encompassing these outlooks into analyzing socio-technical systems is the concept of a social machine. Social machines are typically presented as systemsthat combine some form of social participation with conventional forms of machine-based computation.There are numerous views on what defines a social machine. The first attempt at defining the conceptwas provided by Berners-Lee and Fischetti [10] in their book “Weaving the Web: The Original Designand Ultimate Destiny of the World Wide Web”:

Real life is and must be full of all kinds of social constraint – the very processes fromwhich ‘society’ arises. Computers can help if we use them to create abstract social ma-chines on the Web: processes in which people do the creative work and the machine doesthe administration.

47

Although this characterization is a valuable tool in identifying social machines, it is extensively focusedon the design and engineering aspects of the mechanisms that are found in social machines and over-looks the social dynamics that inhabit them. Attempting to rectify these shortcomings, Smart, Simperland Shadbolt [74] proposed the following definition which brought a major conceptual shift. It demo-cratically allowed all involved components, animate or inanimate, to have a participatory role in socialmachinery [79].

Social machines are Web-based socio-technical systems in which the human and tech-nological elements play the role of participant machinery with respect to the mechanisticrealization of system-level processes.

In parallel to the above, another tangential conception tends to see social machines as socio-computationalsystems. This view entails processing social machines as “socially-extended computational systems” inwhich some aspects of the computational parts are delegated to multiple human individuals. While thereis a clear common ground among the above stated definitions, they do harbor significant differences interms of the scope of conceptualizations entertained by each of these perspectives.

Social Machine, at present is a conjectural notion with various ideas and definitions attempting toelucidate the idea it represents. However, we believe that the convoluted image of the entire systemcan be simplified by understanding its individual components and would lead to an inductive processof discovering the whole itself. We are, by means of our research, attempting to formalise one suchpart of this whole. The human agent or individual in a social machine has one of the most complicatedparticipatory roles in terms of their relations, goals, contributions to the system, and is one such facetwhich is invariant across the diverse perceptions of a social machine. The individual is the buildingblock to the “social” aspect of a socio-technical machine. Thus, understanding one individual is thekey to unlocking the social interaction between multiple individuals and finally their interactions withtechnology in a given interface.

Thus, by means of this work and our model we can propose a unified framework of any individ-ual who is participatory in a social machine, modeled based on Buddhist psychological tradition, as astochastic finite state machine. Further, we address the issue of temporal variability of processes in a so-cial machine and how our model of an individual can accommodate visualizing the machine at differenttemporal granules. By temporal variability, we express the differences between processes that are rela-tively short-lived, fleeting (for instance in the case of a social machine that supports social coordinationwith respect to a specific event) and ones that are enduring (for instance a social machine which keepsrecord of a relatively longer span of time). We also briefly discuss how various other characteristics likesociability, visibility of user contributions, variability in goals and evolving personality of an individualcan be captured with the proposed model of an individual.

48

4.1.1 Proof of Concept: Working Example

To illustrate how our proposed model fits into the mould of an individual in the framework of a so-cial machine we present an example derived from the social media platform Facebook. This model isa spectacle of thorough insight into the persona of an individual. It handles the varied temporal char-acteristics and sparseness of data from social media platforms, accommodating it seamlessly into oursystem. The structure of our unitary moments (or states in the automaton) have the ability to combine,or mould together into bigger moments, thereby combining the dynamic atomic elements and revealingconnects beyond the unitary level. These also help us to observe the motivations that are inherent in themulti-individual social machine systems, thus, coming in handy for studying individual incentives forsocially motivated observations in web sciences.

One of the frameworks which attempts to address a definition space for social machine has beencovered in paper [10]. Here they discuss 31 constructs clustered according to the main components ofsocial machine: social and machine driven services and the interactions between them. We are by meansof the following example trying to expatiate on the social aspect of these constructs. We are not onlyunraveling the motivation (one of the 6 motivations discussed in the paper) but also the mental proce-dures concomitant to each of those motivations. This can also be seen as one of the major advantagesof the proposed techniques.

As a working example, we consider the case of an online social machine scenario of an individualupdating his/her status on Facebook. Here the programmed technological element of posting a statusprompts the user with questions such that “X happened today, what do you think”, or “what is on yourmind”, the user in turn participates with the interface, with a motivation to also engage in multi-userstatus sharing with his/her social network. While the present work deals with the motivations and men-tal states of one individual, it can by means of its mutability be scaled to a group of individuals. Eachmoment represented below is in chronological order with the temporal granularity set to a day. Each mo-ment is populated from status updates by the individual in consideration who we refer to as Person. Themental procedures and action procedures of this Person might express themselves at a singular moment,or across a combination of several Moments. Thus, preserving the handling of variable granularity inour model. The dataset used is the same as described in Chapter 2 for all our applications here.

49

Table 4.1 Observing and analyzing the role of an Individual in Facebook as a social machine.

MOMENT DATE STATUSMENTAL PROCE-

DURES

Moment 106/15/09

01:15 PM

is sore and wants the knot

of muscles at the base of

her neck to stop hurting.

On the other hand, YAY

I’M IN ILLINOIS! <3

CFPIC*A, Desire,

Interest

Moment 206/16/09

04:52 AM

is watching cousin play

computer game on tele-

vison box thing. Also,

sleepy.

CFPIC*A, Torpor,

Sloth

Moment 3 06/17/09 - Sleep State


Moment 506/19/09

03:21 PM

likes the sound of thunder.CFPIC*A , Interest



Moment 806/22/09

04:48 AM

likes how the day sounds

in this new song.Interest, CFPIC*A



Moment 1106/25/09

04:36 AM

saw Transformers, Up,

and Year One this week.

Good movie overload. :D

CFPIC*A, Interest,

Decision Making


Moment 1306/27/09

05:41 AM

saw a nun zombie, and

liked it. Also, *PROP-

NAME* + Tentacle!Man

+ Psychic Powers =

GREAT Party.

CFPIC*A, Interest,

Discursive Think-

ing






Moment 1907/11/2009

05:44:00

is tired. *PROPNAME*,

let me go to sleep pl0x.Sloth, Torpor


50

In Table 4.1 are the sequence of statuses updated by a given user. Following each status is a bag ofwords representations of mental states captured in the respective moment. We intend to capture thesesets of mental states to understand and subsequently predict the individual’s participation in a socialmachine. Here the mental procedure labeled CFPIC*A is a set of 7 cognitive procedures namely : Initialcontact with the object: Feeling or Sensation < Perception < Intention < Concentration < Vitality <Attention or Advertance. Sleep State represents the empty states i.e. time-periods (here, days) for whichwe have no data available. These mental procedures are helpful in tracing the motivation that sustainsprolonged and continuous participation of human-counterparts in the system. The 6 different types ofmotivations listed in the study by Shadbolt et al [73] can be easily mapped onto the respective set ofmental procedures covered as a part of Abhidhamma theory (Table 4.2) Thus, our model is not onlycomplementary to the work undertaken in the domain so far, it also helps us in carrying out a morein-depth analysis of the intentions which constitute each of the given motivations.

Table 4.2 Motivation and Mental Procedures

Motivation Mental Procedures

Participation is funPerceive, Energy, Interest, Deci-

sion Making

Accomplishment of an activity

that the participant enjoys or

wants to finish

Desire, Decision Making

Participation satisfies the desire to

gain/share knowledge

Wisdom, Insight, Proficiency,

Discursive Thinking, Decision

Making

Participation to satisfy the desire

to be socialIntention, Desire, Interest

Philanthropic

Compassion, Sympathetic, Joy,

Right Action, Right Speech,

Generosity

Induction is regarded as the transition from plenty of particulars to the general. The individual weare meaning to formalise here is one of these particulars which will lead us to the general idea of a socialmachine. It is a unifying construct which formalizes the conception of an individual in the plethora ofideologies and the different examples of social machines across domains. The power of our momentcentric model as a descriptive tool lies in its inherent ability to combine and influence multiple models

51

much like itself, allowing for inter-personal dynamics. This makes it easier for us to envision and hereon,formulate a multi-agent social machine in the future.

Another idea that we intend to scale our present framework to is ensuring enough ubiquity in themodel/definition of a social machine which can help us in maintaining inter-operability across variouswebsites. For instance, to get the real sense of an individual engaged with the Web, we would need allhis/her activities across various social machines to understand their motivations/mental procedures intheir entirety. For instance, a Person could be demonstrating X mental procedures on (say) Facebookwhile being engaged in an entirely different set of mental procedures Y on (say) Quora. To be ableto capture these and infuse them into the profile of a single human agent (at a given moment in time)is our vision for the model’s furture applicability. Our work thus paves way for a universal model ofa social machine, across a multitude of social media platforms. In conclusion, we not only propose amodel to extensively analyse an individual engaging in a social machine, we also introduce the ideaof furthering this model to incorporate muti-user scenarios and inter-operability across various socialmedia platforms.

4.2 Modeling Lexical Psychological Phenomenon

Through the multifarious parallels that one can draw between Abhidhammas theory of psychologyand that of western ideology of individuals psychology, it can be inferred that many of the psychologi-cal phenomenon explained by the latter can also be expatiated by the former. That is any psychologicalphenomenon which encapsulates contact with a physical or mental object and involves a series of transi-tions to transpire within an individual’s psychology can as well be explained by means of Abhidhamma.These include, but are not limited to various psychological phenomenon associated with the constantmanifestations of one or a combination of mental states (leading to the establishment of a recurrentfeeling). We attempted to explore various such phenomenon by modeling their parent mental states bymeans of our model - PACMAN (Chapter 3). These were Anxiety [77], Impulsiveness [8] (also relatedto recognizing teenage behavior [28], that gut feeling [39] and so on. One of the psychological phe-nomena which can be captures, recognized, observed and studied by means of our model is Anxiety. Asquoted in the psychological western literatures anxiety can be defined as:

Anxiety is a natural response and a necessary warning adaptation in humans. Anx-iety can become a pathologic disorder when it is excessive and uncontrollable, requiresno specific external stimulus, and manifests with a wide range of physical and affectivesymptoms as well as changes in behavior and cognition. As outlined in the Diagnostic andStatistical Manual of Mental Disorders, fourth edition, text revision (DSM IV-TR), anxietydisorders include generalized anxiety disorder (GAD), social anxiety disorder (also knownas social phobia), specific phobia, panic disorder with and without agoraphobia, obsessive-compulsive disorder (OCD), post traumatic stress disorder (PTSD), anxiety secondary tomedical condition, acute stress disorder (ASD), and substance-induced anxiety disorder

52

Another more biologically salient interpretation can be mapped to the following from the Diagnosticand statistical manual of mental disorders [6]:

Generalized anxiety disorder has been linked to disrupted functional connectivity of theamygdala and its processing of fear and anxiety. Sensory information enters the amygdalathrough the nuclei of the basolateral complex (consisting of lateral, basal, and accessorybasal nuclei). The basolateral complex processes sensory-related fear memories and com-municate their threat importance to memory and sensory processing elsewhere in the brainsuch as the medial prefrontal cortex and sensory cortices. Another area the adjacent centralnucleus of the amygdala that controls species-specific fear responses with its connectionsbrain stem, hypothalamus, and cerebellum areas. In those with general anxiety disorderthese connections functionally seem to be less distinct and there is greater gray matter inthe central nucleus. Another difference is that the amygdala areas have decreased con-nectivity with the insula and cingulate areas that control general stimulus salience whilehaving greater connectivity with the parietal cortex and prefrontal cortex circuits that un-derlie executive functions. The latter suggests a compensation strategy for dysfunctionalamygdala processing of anxiety.

This is consistent with cognitive theories that suggest the use in this disorder of attempts to reduce theinvolvement of emotions with compensatory cognitive strategies. In keeping with both these theories,we can mould the psychological phenomenon of Anxiety by means of the Abhidhamma model of lexicalpsychology. The engendering of anxiety is to be attributed to the conception of an “unpleasant” men-tal factor on encountering an unpleasant object (mental/physical). This object could be a product of theprevious chain of mental factors or something that is perceived by an individual externally. For instance,as illustrated in Figure 4.1, the unpleasant object from a previous mental state, mp (or an external factoraffecting the present one) is the cause of an unpleasant mental factor (here) “Worry”. The mental factorsucceeding the state, mp addresses the unpleasant object along with it’s citta (in which this mental stateis embedded, as explained in Section 4.2) also cognizing it. Subsequently, the next state of “Fear” is aproduct of this previous object and mental state “Worry”. This state of unpleasant mental factors tendsto recurrently manifest themselves and arouse a state of “Anxiety”. There are two ways in which wecan recognize anxiety in an individual: Expressed and Felt (un-expressed). While Expressed Anxiety(or any other psychological phenomenon) is easily deducible from the information an individual will-ingly divulges (vocal expression, social media updates, emotional expressions and so on), Felt Anxietyrequired further reading in between lines and deeper psychological engagement with an individual at amore personal level. For the scope of this thesis, we intend to solely recognize Expressed psychologicalphenomenon. These are both conscious and unconscious expressions of certain mental states that a userconveys via their social media presence.

So as to elaborate on this further, we consider a few hypothetical statuses formulated by the charac-teristics of performance anxiety in musicians by Dianna Kenny [45]. We also extrapolate this to alsoexplain and utilize the Lexical Engagement Factor µ (defined in Section 4.3) Kenny’s book talks about

53

Figure 4.1 The Worry mental state is either a conceived product of the previous mental factors or aninduced mental state from external stimuli/object. Both of these instances can be modeled as a cogni-tive/physical object (m). This state gives rise to the next subsequent state Fear. These further mani-festations are a product of both the personality of an individual and the external factors as defined byAbhidhamma.

the various major factors which were experienced in reaching high anxiety levels for 357 professionalmusicians, all members of one of the eight premier state orchestras in Australia. We notice that all thesestates subsequently leading towards performance anxiety are mental states also conceivable in Bud-dhism’s Abhidhamma. The results in this book are descriptive of the structure of performance anxietywhich is consistent with the emotion-based theory of anxiety described above.

Table 4.3 This table illustrates the probable mental states (Mental Factors Extracted) mined from theusual state of minds of performing musicians through their hypothetical Social Media Statuses. The cor-responding Lexical Engagement Factor µ from our analysis has been listed in column 3 (and explainedin section 4.3)

Social Media Status Metal Factors Extracted µ

Have a big performance tonight! Neutral Universals () 1I have a weird tingling feeling in my stomach! I hope things are okay tonight Skepticism (Somatic Anxiety) 1

I usually screw up big nights! I have a feeling I might mess up this time too! Worry (Worry/Dread) 1

I dont see this ending well for me, its scary and I might not go Worry, Restlessness (Hopelessness) 2

.... Worry x

.... Worry x+ +

There are various factors which lead to, are in function and are a product of performance anxieties.Not only their semantics but also the order in which these factors spew anxiety disorders in the psycheof an individual are consistent with those studied by Abhidhamma meditations. A detailed descriptionof these factors, their descriptions and parallels with those in Abhidhamma are illustrated by meansof Figure 4.2. While the present systems of psychology help identifying the state of an individualsuffering from these disorders, there are no such models that can imitate and capture the evolutionary

54

nature of psychological phenomenon at a temporal level as our model does for Anxiety. Since themodel is in line with the western theories of psychology and also draws salience from the biomedicalaspects of individual psychology, we can draw a reasonable inference that it facilitates similar modelingfor other psychological phenomenon as well. The identification of early symptoms of anxiety such assomatic anxiety, the consecutive states associated with worry etc. through lexical features from the socialmedia websites hypothesizes a newer method of early diagnosis and treatment of such psychologicalphenomenon using solely social media data.

4.3 Lexico-psychological Engagement Factor

Various psychological studies have followed on the premise that people use varying degrees of theirselves physically, cognitively and emotionally, in their day to day engagements, performance at work-place, relationships, self reflections and so on. These studies are grounded in both empirical results (byengaging with subjects on a one-on-one basis) and theoretical frameworks of psychology [9]. Most ofthese studies attempt to capture the engagement, of workers [9] and students [10] in various conditionspresented by the principle investigators of these studies. While there are abundant works in the domainthat attempt to address the performance or engagement of workers, students, users in day to day activi-ties by monitoring their actions empirically, there is a dearth of such mechanisms that can capture thistemporal engagement from social media platforms (and data).

In the age where people update their day-to-day activities (sometimes on an hourly basis) to socialmedia platforms 1, the lexical information evinced by these activities is of unparalleled importance andabundance. So as to capture the psychological engagement of users with a particular mental state (forinstance Worry) over a consecutive set of moments we undertook an in-depth analysis of the Facebookstatuses of 50 users. By means of the PACMAN model we extracted the mental states from each of thestatuses for all these 50 users. We then analyzed the moment by moment mental state of each user so asto capture the depth of the most recurrent mental states. We define the Lexico-Psychological Engage-ment Factor (µ) as the ratio of the maximum depth (no. of moments) of recurrence of a particular mentalstate (xi) to the total number of moments. We can mathematically express this ratio by the following:

µ =max(x1, x2, x3...)

total number of moments(4.1)

Here, in equation 4.1, xi is the number of consecutive moments where the mental state i occurs withoutinterruption from any other mental state. The maximum of these occurrences defines the maximumdepth of lexical engagement of the user towards a mental state. The ratio of these consecutive numberof moments with the total number of moments defines the Lexico-Psychological Engagement Factor ofan individual as analyzed from their social media data.

1http://www.pewinternet.org/fact-sheets/social-networking-fact-sheet/

55

Figure 4.2 State by state evolution of anxiety in performing musicians (adapted from [45]). These statescan inhere in a moment or across multiple moments based on the lexo-psychological engagement factorof an individual.

56

Figure 4.3 The figure represents the Lexico-Psychological Engagement factor for 50 users as analyzedby the (mental factors mined from) statuses posted by these users over a span of 1 year.

The values computed for all of these 50 users are illustrated in Figure 4.3. The range of this engage-ment factor as observed is 0.023 ≤ µ ≤ 0.25. The average is computed to be 0.085, which means thatthe individuals with the maximum engagement factor have a much greater sensitivity and involved men-tal states as opposed to the average of these users. There are various applications particularly revolvingaround early diagnosis of various psychological phenomenon from social media cues of a person thatone can extend this system to predict. The study of this engagement factor can also be furthered toclassify individuals into the following categories, thus analyzing the mental factors evinced by individu-als in particular categories differently. For instance Age based demographical classifications can lead todifferent results for pre-teens, teens, adults and older individuals. We can also undertake individual anal-ysis differentiating different people on the basis of gender, locality, cultural values etc. For example,studies suggest that some parts of Eastern Asia have more reserved cultures than Western societies2.This would imply that while the engagement of the individual might be less on a publicly accessiblesocial media platform, it might not be an absolute reflector of their real engagement towards a givenmental state. There are also various studies which discuss the ease of expression of positive emotionsfor some sections of the society as opposed to the negative ones. Thus, the expressions of the pleasantand unpleasant mental factors if captured in conjugation with the cultural context of an individual wouldhelp us in arriving at deeper insights solely from their lexical information.

2https://www.american.edu/soc/communication/upload/Shani-Lewiscapstone.pdf

57


The work undertaken as this part of the research has thus validated the various use-cases of ourstochastic model of an individual. Foremost we used our model to mould the idea of an individual insocial machines. This helped in unifying the multiple definitions trying to encapsulate the various ideasof such socio-computational systems. The model, because of its temporal malleability, was also handyin handling the issue of data sparsity in social media. Next, we modeled a psychological phenomenon-Anxiety, recognizing the mental states associated with its conception, growth and finally leading on toterminal depression. We illustrated the usefulness of this model in capturing such instances of manymore such psychological phenomenon among social media users. Finally, we introduced the Lexico-psychological Engagement Factor (µ) to encapsulate the mental engagement that any given social mediauser exhibited. While there exist many such theoretical concepts in psychology, this is the first concep-tualization of it in computational psychology from social media. Unlike the state of the art factors ofuser engagement, µ does not depend upon psychometric tests. It is instead informed by the social mediafeatures demonstrated by a given user.

58

Chapter 5

Conclusions and Future Work

5.1 Conclusions

In this thesis we explored the theoretical as well as computational aspects of lexical psychology andpersonality of an individual. We began our work by improving the state of the art techniques to pre-dict the big five model of personality using social media data. Mining user personality is an importantpart of creating futuristic contextual systems across domains eg. advertisements, medical assistants,office assistants, augmented reality applications, contextual search and so on. Big five is a model ofpersonality which helps categorizing a users into the categories: Extraversion, Neuroticism, Opennessto experience, Agreeableness and Conscientiousness. We helped elevating the state of the art accuracyof predicting big five user personality. Our features were largely informed by the domain knowledge oftheoretical psychology (unlike all of the other works we came across). By means of our model we werenot only successful in elevating personality prediction accuracy but it also helped validate and improvethe Lexical Hypothesis of Psychology. We proposed a set of features called the Adjectival Markers thatwere adept in predicting user personality from peer reviews of the user as opposed to his/her own shareddata. This has helped in bettering the current system of personality mining and given a methodology tomake use of the peripheral data that was discarded earlier to be used in an efficient manner. Next, toverify the dependency of big five solely on lexical markers, as illustrated by the theoretical origins of thepersonality model, we explored mobile data features instead, thus, seeking to compare the accuracies ofboth feature sets. Although the enhanced logistic regression model we proposed was better than the cur-rent state of the art techniques for mobile data, it did not outdo any of the existing lexical models beingused to predict big five. We thus corroborated the salience of lexical data in predicting an enriched bigfive user personality as opposed to the scattered mobile phone features such as calls, SMS frequenciesetc.

A deeper exploration into the domain of lexical psychology and extensive literature helped us dis-cover that while big five personality was an overall static image of a user, real world applications de-manded a more dynamic one. Literature in the domain suggested that the user psychology was moulded

59

subsequently in time, depending on the life experiences and innate tendencies of individuals. The pref-erences exercised by an individual kept changing and so did their psychologies. We thus needed a modelto encapsulate this change from the temporal (time to time) information available in social media. Weexplored multiple theories of various fundamental ideas of psychology. We finally arrived at the Ab-hidhmma tradition’s model of an individual. Unlike any other model we came across, this model couldbe mapped to a formal theorization, extended to mathematical explanations, paralleled with biologicalexplanations about various psychological phenomenon and also adhered to the logic of phenomenolog-ical derivations.

Thus, so as to computationally implement it, we formulated a stochastic finite state machine repre-sentation for the various functions defined in the model. The model not only accommodates the dynamicmovement of the subsequent states as a function of the previous ones, but also helps in defining a feed-back loop that illustrates how a moment affects the inherent psyche of an individual. In the scope ofthis thesis we concentrated ourselves only on the mental states discussed in the model. These mentalstates provided a window into understanding the moment by moment persona of an individual. So as toscale this model to a computationally predictive one, we then worked on accumulating a domain specificdataset. We achieved this by having three annotators with sufficient domain knowledge to tag our socialmedia data with various mental state descriptors defined in Abhidhamma. We used this dataset to thentrain and test out multi-class SVM model predicting state by state mental features of an individual. Wecould then cumulatively use these mental features over a period of time to infer the respective personaexhibited by an individual during that time.

We finally used this theoretical and computational model to exhibit several use cases. First was topropose a ubiquitous model of an individual for social machines. This model helped in streamlining therole of an individual in a scattered notion of social machines. We also modeled the psychological phe-nomenon of Anxiety using this model. We were able to observe the conception of anxiety related mentalstates, their continued inhering in an individual and their growth with and without the interference of anexternal stimuli or positive mental states. Lastly, we proposed the Lexico-psychological mental engage-ment factor for a social media user. While most psychological engagements of users are captured overa series of psychological tests, we theoretically define this over mental states mined by computationalmeans through social media user. This would help ease the understanding about each user and theirengagement with a respective mental state. For instance, if a user has a high engagement factor they areprone to deeper reflections and can thus be understood in that context.

60

5.2 Future Work

Since this work is the first of its kind across a multitude of domains namely: psychology, mathematicsand computer science, in our opinion it has a great potential to engage researchers in furthering what wehave started. Few problems that I personally encountered in extending the current work were as follows:

• Handle Data Limitations: There is a dearth of data related to works dealing with user psychol-ogy. Especially for Abhidhamma traditions, scanty data often times meant that we had to exhaustgreat amount of resources in curating, cleaning, annotating, verifying large datasets.

• Probabilistic Modeling: There is an expansive possibility to map the current models as prob-abilistic stochastic models, by populating it with abundant data and computing the probabilityvalues.

• Feedback Loop Accommodation: The model currently is descriptive of the mental states andlexical psychology of an individual. This can be augmented to incorporate various actions whichin turn can feedback into the mental states of the model.

61

Related Publications

• Shivani Poddar, VenuMadhav Kattagoni, and Navjyoti Singh.”Personality Mining from Biograph-ical Data with the Adjectival Marker”. Biographical Data in a Digital World 2015. Vol-139939-47 [67]

• Shivani Poddar, Arpit Merchant and Navjyoti Singh. ”Personality Recognition from MobilePhone Data”. Mobile HCI 2015 (Weak Reject)

• Shivani Poddar, Sindhu Kiranmai, Najyoti Singh and Rev. Ashin Samvara. ”Towards a Ubiqui-tous Model of an Individual in Social Machines”. SOCM, WWW 2016 ACM. ISBN 123-4567-24-567/08/06. DOI: 10.475/123 4 [66]

• Shivani Poddar, Sindhu Kiranmai and Navjyoti Singh. ”PACMAN: Psycho and ComputationalFramework of an Individual (Man)”. ESA, LREC 2016

• Shivani Poddar and Navjyoti Singh. ”Towards Understanding Lexical Psychological Phenomenonand Mental Engagement via the Stochastic Model of Buddhism Psychology”. SCAP 2016

62

Bibliography

[1] F. Adrian. The big five versus the big four: the relationship between the myers-briggs type indicator (mbti)

and neo-pi five factor model of personality. Personality and Individual Differences, 1996.

[2] F. Alam, E. A. Stepanov, and G. Riccardi. Personality traits recognition on social network-facebook. WCPR

(ICWSM-13), Cambridge, MA, USA, 2013.

[3] Allport. Traitnames. A psycho-lexical study, Psychological Monographs, 47, 1936.

[4] J. J. Appleton, S. L. Christenson, D. Kim, and A. L. Reschly. Measuring cognitive and psychological

engagement: Validation of the student engagement instrument. Journal of School Psychology, 44(5):427–

445, 2006.

[5] S. Argamon, M. Koppel, J. W. Pennebaker, and J. Schler. Automatically profiling the author of an anony-

mous text. Communications of the ACM, 52(2):119–123, 2009.

[6] A. P. Association, A. P. Association, et al. Diagnostic and statistical manual of mental disorders (dsm).

Washington, DC: American psychiatric association, pages 143–7, 1994.

[7] Y. Bachrach, M. Kosinski, T. Graepel, P. Kohli, and D. Stillwell. Personality and patterns of facebook usage.

In Proceedings of the 4th Annual ACM Web Science Conference, pages 24–32. ACM, 2012.

[8] E. S. Barratt. Impulsiveness subtraits: Arousal and information processing. Motivation, emotion, and

personality, 5:137–146, 1985.

[9] M. R. Barrick and M. K. Mount. The big five personality dimensions and job performance: a meta-analysis.

Personnel psychology, 44(1):1–26, 1991.

[10] T. Berners-Lee, M. Fischetti, and M. L. Foreword By-Dertouzos. Weaving the Web: The original design

and ultimate destiny of the World Wide Web by its inventor. HarperInformation, 2000.

[11] A. Bianchi and J. G. Phillips. Psychological predictors of problem mobile phone use. CyberPsychology &

Behavior, 8(1):39–51, 2005.

[12] D. T. Blackburn. A generalized distance metric for the analysis of variable taxa. Botanical Gazette, pages

325–335, 1980.

[13] F. Brentano. Psychology from an empirical standpoint. Routledge, 2014.

[14] S. Butt and J. G. Phillips. Personality and self reported mobile phone use. Computers in Human Behavior,

24(2):346–360, 2008.

63

[15] R. Capraro. Myers-briggs type indicator score reliability across: Studies a meta-analytic reliability gener-

alization study. Educational and Psychological Measurement, 2002.

[16] J. M. Carroll. Human–computer interaction. Encyclopedia of Cognitive Science, 2009.

[17] R. Cattell. The description of personality: basic traits resolved into clusters. The Journal of Abnormal and

Social Psychology, 38(4):476–506, 1943.

[18] F. Celli, F. Pianesi, D. Stillwell, and M. Kosinski. Workshop on computational personality recognition

(shared task). In Proceedings of the Workshop on Computational Personality Recognition, 2013.

[19] P. L. Celli F. Relationships between Personality and Interactions in Facebook, pages 41–54. Nova Science

Publishers, 2013.

[20] E. A. Cherman, M. C. Monard, and J. Metz. Multi-label problem transformation methods: a case study.

CLEI Electronic Journal, 14(1):4–4, 2011.

[21] G. Chittaranjan, J. Blom, and D. Gatica-Perez. Who’s who with big-five: Analyzing and classifying per-

sonality traits with smartphones. In 2011 15th Annual International Symposium on Wearable Computers,

pages 29–36. IEEE, 2011.

[22] P. T. Costa Jr and R. R. McCrae. Still stable after all these years: Personality as a key to some issues in

adulthood and old age. Life-span development and behavior, 1980.

[23] P. T. Costa Jr and R. R. McCrae. Domains and facets: Hierarchical personality assessment using the revised

neo personality inventory. Journal of personality assessment, 64(1):21–50, 1995.

[24] C. Cox. From category to ontology: The changing role of dharma in sarvastivada abhidharma. Journal of

Indian philosophy, 32(5):543–597, 2004.

[25] L. Curtis, C. Edwards, K. L. Fraser, S. Gudelsky, J. Holmquist, K. Thornton, and K. D. Sweetser. Adoption

of social media for public relations by nonprofit organizations. Public Relations Review, 36(1):90–92, 2010.

[26] Y.-A. de Montjoye, J. Quoidbach, F. Robic, and A. S. Pentland. Predicting personality using novel mobile

phone-based metrics. In International Conference on Social Computing, Behavioral-Cultural Modeling,

and Prediction, pages 48–55. Springer, 2013.

[27] R. J. Edelmann. Anxiety: Theory, research and intervention in clinical and health psychology. John Wiley

& Sons, 1992.

[28] W. N. Evans, W. E. Oates, and R. M. Schwab. Measuring peer group effects: A study of teenage behavior.

Journal of Political Economy, pages 966–991, 1992.

[29] E. Faliagka, L. Kozanidis, S. Stamou, A. Tsakalidis, and G. Tzimas. A personality mining system for auto-

mated applicant ranking in online recruitment systems. In Proceedings of the 11th International Conference

on Web Engineering, ICWE’11, pages 379–382, Berlin, Heidelberg, 2011. Springer-Verlag.

[30] D. W. Fiske. Consistency of the factorial structures of personality ratings from different sources. The

Journal of Abnormal and Social Psychology, 44(3):329, 1949.

[31] W. Gao and Z.-H. Zhou. On the consistency of multi-label learning. Artificial Intelligence, 199:22–44,

2013.

64

[32] J. Golbeck, C. Robles, M. Edmondson, and K. Turner. Predicting personality from twitter. In Privacy,

Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Social Computing

(SocialCom), 2011 IEEE Third International Conference on, pages 149–156. IEEE, 2011.

[33] J. Goldbeck, C. Robles, and K. Turner. Predicting personality with social media. In Proceedings of the

annual conference extended abstracts on Human factors in computing systems, 2011.

[34] L. Goldberg. The development of markers for the big five factor structure. Psychological Assessment,

4(1):26–42, 1992.

[35] L. Goldberg. The development of markers for the big-five factor structure. Psychological assessment,

4(1):26, 1992.

[36] L. R. Goldberg. Language and individual differences: The search for universals in personality lexicons.

Review of personality and social psychology, 2(1):141–165, 1981.

[37] L. R. Goldberg. An alternative” description of personality”: the big-five factor structure. Journal of person-

ality and social psychology, 59(6):1216, 1990.

[38] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik. Gene selection for cancer classification using support

vector machines. Machine learning, 46(1-3):389–422, 2002.

[39] S. P. Hams. A gut feeling? intuition and critical care nursing. Intensive and Critical Care Nursing,

16(5):310–318, 2000.

[40] M. Hartimo. Phenomenology and mathematics, volume 195. Springer Science & Business Media, 2010.

[41] J. E. Hopcroft. Introduction to Automata Theory, Languages and Computation: For VTU, 3/e. Pearson

Education India, 1979.

[42] F. Iacobelli, A. J. Gill, S. Nowson, and J. Oberlander. Large scale personality classification of bloggers. In

Affective Computing and Intelligent Interaction, pages 568–577. Springer, 2011.

[43] M. K. Jill Goldberg. Operational efficiency and the growth of short-term memory span. Journal of Experi-

mental Child Psychology, 33(3):386–404, 1982.

[44] W. A. Kahn. Psychological conditions of personal engagement and disengagement at work. Academy of

management journal, 33(4):692–724, 1990.

[45] D. T. Kenny and M. S. Osborne. Music performance anxiety: New insights from young musicians. Advances

in Cognitive Psychology, 2(2-3):103–112, 2006.

[46] K. L. Kermanidis. Mining authors personality traits from modern greek spontaneous text. In 4th Interna-

tional Workshop on Corpora for Research on Emotion Sentiment and Social Signals, in conjunction with

LREC12, pages 90–94. Citeseer, 2012.

[47] T. Koenig, L. Prichep, D. Lehmann, P. V. Sosa, E. Braeker, H. Kleinlogel, R. Isenhart, and E. R. John. Mil-

lisecond by millisecond, year by year: normative eeg microstates and developmental stages. Neuroimage,

16(1):41–48, 2002.

[48] D. W. Kolar, D. C. Funder, and C. R. Colvin. Comparing the accuracy of personality judgments by the self

and knowledgeable others. Journal of personality, 64(2):311–337, 1996.

65

[49] D. Lehmann. Multichannel topography of human alpha eeg fields. Electroencephalography and clinical

neurophysiology, 31(5):439–449, 1971.

[50] D. Lehmann, H. Ozaki, and I. Pal. Eeg alpha map series: brain micro-states by space-oriented adaptive

segmentation. Electroencephalography and clinical neurophysiology, 67(3):271–288, 1987.

[51] K. Luyckx and W. Daelemans. Authorship attribution and verification with many authors and limited data.

In Proceedings of the 22nd International Conference on Computational Linguistics-Volume 1, pages 513–

520. Association for Computational Linguistics, 2008.

[52] K. Luyckx and W. Daelemans. Personae: a corpus for author and personality prediction from text. In LREC,

2008.

[53] C. MacCann, A. L. Duckworth, and R. D. Roberts. Empirical identification of the major facets of conscien-

tiousness. Learning and Individual Differences, 19(4):451–458, 2009.

[54] F. Mairesse, M. A. Walker, M. R. Mehl, and R. K. Moore. Using linguistic cues for the automatic recognition

of personality in conversation and text. Journal of artificial intelligence research, pages 457–500, 2007.

[55] G. H. B. P. P. R. Mark Hall, Eibe Frank and I. H. Witten. The weka data mining software. An update. ACM

SIGKDD Explorations Newsletter, 11(1):10–18, 2009.

[56] R. R. McCrae, P. T. Costa Jr, A. Terracciano, W. D. Parker, C. J. Mills, F. De Fruyt, and I. Mervielde. Per-

sonality trait development from age 12 to age 18: Longitudinal, cross-sectional and cross-cultural analyses.

Journal of personality and social psychology, 83(6):1456, 2002.

[57] L. Meier, S. Van De Geer, and P. Buhlmann. The group lasso for logistic regression. Journal of the Royal

Statistical Society: Series B (Statistical Methodology), 70(1):53–71, 2008.

[58] M. T. Mon. Buddha abhidhamma: Ultimate science. 1995.

[59] S. Moscoso and J. F. Salgado. ‘dark side’ personality styles as predictors of task, contextual, and job

performance. International Journal of Selection and Assessment, 12(4):356–362, 2004.

[60] T. Nomura. Generation of relations between individuals based on a stochastic automaton and an analogy

from social psychology. Proceedings of the ALIFE V Poster Presentations, pages 125–132, 1996.

[61] W. T. Norman. Toward an adequate taxonomy of personality attributes: Replicated factor structure in peer

nomination personality ratings. The Journal of Abnormal and Social Psychology, 66(6):574, 1963.

[62] J. Oberlander and S. Nowson. Whose thumb is it anyway?: classifying author personality from weblog text.

In Proceedings of the COLING/ACL on Main conference poster sessions, pages 627–634. Association for

Computational Linguistics, 2006.

[63] M. R. . C. P. Reinterpreting the myers-briggs type indicator from the perspective of the five-factor model of

personality. Journal of Personality, 1989.

[64] J. W. Pennebaker, M. E. Francis, and R. J. Booth. Linguistic inquiry and word count: Liwc 2001. Mahway:

Lawrence Erlbaum Associates, 71:2001, 2001.

[65] J. W. Pennebaker and L. A. King. Linguistic styles: language use as an individual difference. Journal of

personality and social psychology, 77(6):1296, 1999.

66

[66] S. Poddar, S. K. Ernala, N. Singh, and A. Samvara. Towards a ubiquitous model of an individual in social

machines. In Proceedings of the 25th International Conference Companion on World Wide Web, pages

793–798. International World Wide Web Conferences Steering Committee, 2016.

[67] S. Poddar, V. Kattagoni, and N. Singh. Personality mining from biographical data with the adjectival marker.

[68] D. Quercia, M. Kosinski, D. Stillwell, and J. Crowcroft. Our twitter profiles, our selves: Predicting per-

sonality with twitter. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational

Conference on Social Computing (SocialCom), 2011 IEEE Third International Conference on, pages 180–

185. IEEE, 2011.

[69] R. M. R and P. T. Costa. Reinterpreting the myers-briggs type indicator from the perspective of the five-

factor model of personality. Journal of personality, 57(1):17–40, 1989.

[70] C. Ross, E. S. Orr, M. Sisic, J. M. Arseneault, M. G. Simmering, and R. R. Orr. Personality and motivations

associated with facebook use. Computers in human behavior, 25(2):578–586, 2009.

[71] J. Sanz, M. L. Sanchez-Bernardos, and M. D. Avia. Self-monitoring and the prediction of one’s own and

others’ personality test scores. European Journal of Personality, 10(3):173–184, 1996.

[72] G. Saucier and L. R. Goldberg. The language of personality: Lexical perspectives. The five-factor model of

personality: Theoretical perspectives, page 21, 1996.

[73] N. R. Shadbolt, D. A. Smith, E. Simperl, M. Van Kleek, Y. Yang, and W. Hall. Towards a classification

framework for social machines. In Proceedings of the 22nd international conference on World Wide Web

companion, pages 905–912. International World Wide Web Conferences Steering Committee, 2013.

[74] P. Smart, E. Simperl, and N. Shadbolt. A taxonomic framework for social machines. In Social Collective

Intelligence, pages 51–85. Springer, 2014.

[75] C. J. Soto, O. P. John, S. D. Gosling, and J. Potter. Age differences in personality traits from 10 to 65: Big

five domains and facets in a large cross-sectional sample. Journal of personality and social psychology,

100(2):330, 2011.

[76] J. Specht, B. Egloff, and S. C. Schmukle. Stability and change of personality across the life course: the

impact of age and major life events on mean-level and rank-order stability of the big five. Journal of

personality and social psychology, 101(4):862, 2011.

[77] C. D. Spielberger and C. D. Spielberger. Theory and research on anxiety. Anxiety and behavior, 1, 1966.

[78] J. Staiano, B. Lepri, N. Aharony, F. Pianesi, N. Sebe, and A. Pentland. Friends don’t lie: inferring personality

traits from social network structure. In Proceedings of the 2012 ACM conference on ubiquitous computing,

pages 321–330. ACM, 2012.

[79] S. Tarte, P. Wilcox, H. Glaser, and D. De Roure. Archetypal narratives in social machines: Approaching

sociality through prosopography. page 24, 2015.

[80] Y. R. Tausczik and J. W. Pennebaker. The psychological meaning of words: Liwc and computerized text

analysis methods. Journal of language and social psychology, 29(1):24–54, 2010.

67

[81] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society.

Series B (Methodological), 58(1):267–288, 1996.

[82] R. L. Tieszen. Phenomenology, logic, and the philosophy of mathematics, volume 203. Cambridge Univer-

sity Press, 2005.

[83] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. Dept. of Informatics, Aristotle

University of Thessaloniki, Greece, 2006.

[84] K. Vertanen and P. O. Kristensson. A versatile dataset for text entry evaluations based on genuine mobile

emails. In Proceedings of the 13th International Conference on Human Computer Interaction with Mobile

Devices and Services, pages 295–298. ACM, 2011.

[85] A. C. Weidman, K. C. Fernandez, C. A. Levinson, A. A. Augustine, R. J. Larsen, and T. L. Rodebaugh.

Compensatory internet use among individuals higher in social anxiety and its implications for well-being.

Personality and individual differences, 53(3):191–195, 2012.

[86] D. J. Willshaw and C. Von Der Malsburg. How patterned neural connections can be set up by self-

organization. Proceedings of the Royal Society of London B: Biological Sciences, 194(1117):431–445,

1976.

68

MODELLING LEXICAL PSYCHOLOGY OF AN...

Documents

Transcript of MODELLING LEXICAL PSYCHOLOGY OF AN...