1 personality assessment from text for machine ...1 personality assessment from text for machine...
TRANSCRIPT
1
Personality Assessment from Text for MachineCommonsense Reasoning
Niloofar Hezarjaribi, Zhila Esna Ashari, James F. Frenzel, Hassan Ghasemzadeh, and Saied Hemati
Abstract—This article presents PerSense, a framework to estimate human personality traits based on expressed texts and to use themfor commonsense reasoning analysis. The personality assessment approaches include an aggregated Probability Density Functions(PDF), and Machine Learning (ML) models. Our goal is to demonstrate the feasibility of using machine learning algorithms onpersonality trait data to predict humans’ responses to open-ended commonsense questions. We assess the performance of thePerSense algorithms for personality assessment by conducting an experiment focused on Neuroticism, an important personality traitcrucial in mental health analysis and suicide prevention by collecting data from a diverse population with different Neuroticism scores.Our analysis shows that the algorithms achieve comparable results to the ground truth data. Specifically, the PDF approach achieves97% accuracy when the confidence factor, the logarithmic ratio of the first to the second guess probability, is greater than 3.Additionally, ML approach obtains its highest accuracy, 82.2%, with a multilayer Perceptron classifier. To assess the feasibility ofcommonsense reasoning analysis, we train ML algorithms to predict responses to commonsense questions. Our analysis of datacollected with 300 participants demonstrate that PerSense predicts answers to commonsense questions with 82.3% accuracy using aRandom Forest classifier.
Index Terms—Personality assessment, commonsense reasoning, machine learning, probability density function, personality trait.
F
1 INTRODUCTION
The demand for personalized services and adaptive tech-nologies such as driver-less cars are projected to growrapidly and will soon be integrated into our daily life. Thecommunication of these technologies with human-beingsneeds to be proper and well-trained in order to serve,explain the facts, and understand different mindsets better.Automating the process of understanding individuals’ pri-orities and their perception of the world can potentially beused to address specific needs of individuals and to estab-lish efficient interactions among human and technologies.Human commonsense reasoning involves human factorsand personality traits in the process of judgment and deci-sion making, and it goes beyond the conventional logical de-ductive or experimental inductive reasoning. These humanfactors include perceptions, desires, herd mentality, men-tal stability, prejudice or other biases, behavior traits, andpersonal/ethical/religious values. Therefore, commonsensereasoning is difficult to formulate for smart technologies.As an example, the complexity of predicting stock marketfluctuations is due to human factor or market psychology,and is not well understood by human-beings [1]. Similarly,the rise and fall of political figures in a democratic society isnot necessarily merit-based and not even deterministic andtime-invariant [2]. In marketing, other than functionalityand appearance of a product, consumers’ evolving taste
This work was supported in part by a grant awarded by Samsung GlobalResearch Outreach (GRO) program in 2016. Niloofar Hezarjaribi, Zhila EsnaAshari and Hassan Ghasemzadeh are with the School of Electrical En-gineering and Computer Science, Washington State University, Pullman,WA 99164-2752, USA. Emails {n.hezarjaribi, z.esnaashariesfahan, has-san.ghasemzadeh}@wsu.edu. James F. Frenzel and Saied Hemati are with theDepartment of Electrical and Computer Engineering, University of Idaho,Moscow, ID, 83844-1023, USA. Emails: {jfrenzel, shemati}@uidaho.edu.
contributes to a brand popularity and its success or failure.Understanding the limitations and ethical consequences ofcommonsense reasoning is paramount and requires devel-opment of machine commonsense reasoning.
In philosophy, the distinction between something rea-sonable and not reasonable is highly challenging to formu-late, especially when a person should choose between twomorally questionable acts [3]. The trolley problem, whichis a well-known ethical thought experiment discussed byPhilippa Foot, is an example of a highly challenging task.In the trolley problem, it is presumed justifiable for a driverof a runaway trolley that is moving towards five railroadworkers to divert the trolley to a branch where only oneworker is on the track. 86% of participants in the surveybelieved saving five lives is more important than saving asingle life [4]. However, a surgeon may not kill a patientor even let the patient die to distribute the patient’s vitalorgans to five other patients that would otherwise die. Themath seems identical, sacrificing one to save five, but onedecision seems reasonable and justifiable to many peopleand the second case is not. The hypothetical trolley problemcan be extended to realistic scenarios with driver-less cars[5]. A driver-less car should be equipped with commonsensereasoning for simply maximizing the gain and minimizingthe loss in its decision making process.
The complexity of commonsense reasoning problem canbe further explained by IBM Watson, the question answer-ing computing system that defeated Jeopardy championsin 2011. Watson had special tools for analyzing naturallanguage and had access to 200 million pages of informationincluding the Wikipedia encyclopedia with no Internet ac-cess. Jeopardy questions were facts not human perceptions.However, in response to the question that in which US city,the largest airport was named for a World War II hero and its
arX
iv:2
004.
0927
5v1
[cs
.CY
] 1
5 A
pr 2
020
2
second largest airport was named for a World War II battle?Watson answered: “What is Toronto?”. The answer was avery uneducated guess as Toronto is not even located inUS! The supercomputer had no human-like commonsensewhen answered this question. Obviously, not all differentscenarios can be programmed or thought through aheadof time. If we ask the same question from a large groupof children, the likelihood of hearing “Toronto” as the an-swer is projected to be extremely low, even though theyprobably will not know the answer. Although IBM Watsoneasily defeated the Jeopardy champions, the real problemarises when IBM Watson is asked about human perceptionsregarding issues that are not necessarily factual or for whichthere exists multiple correct answers for one question. As anexample, it is trivial to list the name of all animals whichwere mentioned in the Bible. However, it is challenging forWatson to guess the most common answer of general publicin US to this question.
In this paper, we hypothesize that “commonsense rea-soning” is mostly grounded in human personality traits. Todate, human personality traits have remained a virtually un-explored sources of information for understanding humanbeings’ priorities and their perception of the world. Thereare numerous standardized questionnaires that have beendeveloped for personality trait analysis [6]–[8]. However,this traditional approach is based on the active participationof subjects in the conducted study. The validity of theseapproaches also depends on the subjects’ honestly anddedication to correctly answer to the questions. Recently,a passive form of personality profiling is gaining interestin which the subjects’ natural behavior is examined, ratherthan the choices that they intentionally make in fillingquestionnaires. The passive approach is projected to moreaccurately reflect the truth if the subjects’ behavior over anextended period of time is available. The passive approachcan also exploit vast information that is readily available onsocial media, from product reviews on Amazon to Twitterand Facebook posts [9], [10]. The passive form of personalityprofiling is projected to shed more light on human mindsetby exploiting the massive information that is readily avail-able on the Internet.
In an entirely new direction, we propose PerSense, anovel framework for personality assessment and common-sense reasoning. The overall approach to depict the de-scribed hypothesis is shown in Figure 1. The relationshipbetween the personality and language is presumed to bepresent in two different directions and personality trait isalso hypothesized to be related with commonsense reason-ing. Therefore, our innovative work in this project aims to:
1) develop computationally-simple algorithms to es-timate human personality traits based on user’sdigital footprint on social media,
2) exploit the connection between personality trait andcommonsense decision making, and
3) demonstrate the feasibility of predicting human be-ings’ response to open-ended questions that have nounique factual answer.
Therefore, PerSense is a framework for commonsense rea-soning based on personality traits. It first employs texts fromdifferent sources such as Amazon reviews for analyzing
individuals’ personality. We designed two low complexityapproaches for estimating personality traits based on textdata in PerSense. In the absence of a large scale surveybased on standard personality traits questionnaires, thepersonality scores generated by the IBM Watson tool areused for ground truth labeling. These scores, which are realnumbers between 0 and 1, are then utilized for statisticalanalysis and machine learning model training focused onpersonality analysis. Finally, we exploit human personalitytraits and their distributions in a society to develop machinecommonsense reasoning. To this end, we use a machinelearning based approach where we train the algorithmsusing data collected from hundreds of subjects throughAmazon Mechanical Turk.
2 RELATED WORK
In this section, we discuss related research on personalityassessment via language and text. We also discuss previousresearch on commonsense knowledge and reasoning.
2.1 Personality Assessment
Personality has been found to significantly correlate witha number of real-world behaviors. It correlates with musictaste: popular music tends to be significantly liked by ex-troverts, while people with a tendency to be less open toexperience tend to prefer religious music and dislike rockmusic [32]. Personality also impacts the formation of socialrelations [36]: friends tend to be, to a very similar extent,open to experience and extrovert [35].
Human personality traits, thoughts, emotions, and judg-ments are amongst the important topics in psychology [11],[12]. Since the most universal way of expressing humanfeelings and thoughts is via language, it is reasonable toassume personality trait and language are related [13]–[15].The lexical hypothesis, which is one of the well-establishedpropositions in psychology, upholds this assumption [8].The lexical hypothesis states that any language provides anunbiased source of different personality types. Thus, it pos-tulates the feasibility of extracting out personality types andpsychometric features by examining the linguistic represen-tation of people [16], [17]. The measurement of psychometricfeatures, which highlight individual differences [18], mustbe reliable in the sense that the measurement method andtools provide stable results across different circumstances.The measurement must also be valid in the sense that theresults remain correct for various purposes [19].
The Big Five model or the five factor model (FFM) [20],is a scheme of personality traits classification [21]–[23]. Thismodel utilizes the evaluation of individuals’ personalitytraits based on their ranking on five bipolar factors calledOCEAN (Openness, Contentiousness, Extroversion, Agree-ableness, and Neuroticism) [24], which is traditionally doneusing standard questionnaires. The validity of the Big Fivemodel depends on the validity of the lexical hypothesis[25], [26]. Accordingly, the validity of the lexical hypotheses,depends on its initial claim that personality is accuratelyrepresented within the structure of the language. There areinteresting applications for the Big Five model. For example,It has been shown that extroverted individuals have more
3
LanguagePersonality
TriatCommonsense
Reasoning
Fig. 1: Utilizing personality traits for commonsense reasoning from language.
Facebook friends [27]. The users with high Neuroticismscores, on the contrary, post more on Facebook wall [28].Another application of Big Five model is in prediction of jobperformance [29].
There are a few technical reports and papers on ex-tracting out personality from lexicon which is called theFactor Analysis. There are different methods for factor anal-ysis. The first and conventional methods where based onthe personality traits description questionnaire [30]–[32]. Itextracted out relevant information about personality fromdescriptive adjectives and different questions and phrasesrelated to personality. The automatic recognition of person-ality traits was first presented in [33]. Recent developmentsin Artificial Intelligence (AI) explore new methods for au-tomatic detection of linguistic features in text. For example,Roderick Hart developed the DICTION program in 2001,which counts the words in a text related to a particulartheme [34]. The DICTION was a word count program toreveal the verbal tone of political statements. The five mastervariables (activity, optimism, certainty, realism, and com-monality) that provide a robust understanding of a text wereused as a basic characterization of a text. A well-known toolfor linguistic analysis is Linguistic Inquiry and Word Count(LIWC) [16], [35], [36], which is an a priori word-categoryapproach. The LIWC method relies on an internal defaultdictionary (a priori word-category) that defines which wordsshould be counted in the target text files. For example,the word “cried” is part of five word categories: sadness,negative emotion, overall effect, verb, and past tense verb.
There are many available user text datasets such as socialmedia, on-line reviews, and essays which can be utilizedin inferring personality and building models for predictingan individual’s behavior in society [37]–[39]. In [40], anapproach was proposed for capturing physical proximity,location, and movement. Later, these data were utilized formodeling human behavior. Phone and email logs can pro-vide some patterns about human communication. A studywas provided by Onnela et al., in which the communicationpatterns of millions of mobile phone users were utilized forobserving the coupling between the strength of interactionsand the structure of the network [41]. In a work proposedby Eagle et al., physical proximities were collected over timefor inferring cognitive relationships, such as friendship [42].
The Internet is the main source for understanding thebehaviors and connections of the people [43], [44]. Personal-ity has impacts on people’s online interaction. In a research
study, it is claimed that user’s personality can be predictedusing Facebook data, accurately [45]. They utilized linearregression for predicting user’s personalities. The inputfeatures included word counts, long words, words abouttime, health, etc. Mean absolute error for this predictionwas 11%. In another study, the impact of users’ personalityon Twitter activities was investigated [46]. In this research,the relationship between big five personality traits and dif-ferent types of Twitter users was studied. Moreover, user’spersonality traits were predicted based on the number offollowings, followers, and times the user has been listed onother’s reading list. For predicting each personality trait,they utilized a 10-fold cross validation with 10 iterationsusing M5′ rules. The maximum root-mean-square error(RMSE) is 0.88. In this work they utilized basic networkproperties in user’ Twitter account. In our research, wecollected data from different sources and text analysis wasperformed. Moreover, we utilized two different approaches,statistical and probabilistic, for predicting personality traits.In another research, they proposed that Instagram picturefeatures can be correlated with personality [47]. They triedto predict personalities based on the way the pictures weretaken and filters applied to them. Features were extractedfrom 22,398 pictures and mean values were calculated foreach feature to create a measurement of central tendency.Later, a correlation matrix was created based on the centraltendencies. Their findings suggested a relationship betweenusers’ personality traits and their posted pictures.
2.2 Commonsense Reasoning
Knowledge representation and reasoning are importantproblems in artificial intelligence. There are different rulesand methods for knowledge representation [48]. In orderfor a computer to act intelligently, it should have a wideknowledge representation of the world. Designing such acomputer, requires information about the type of knowledgeand the ways of obtaining it. The reasoning problem isa philosophical problem that can be solved clearly usingintelligent systems. In reasoning programs, interaction withthe world are through inputs and outputs that can berepresented in different forms. One of these representationsutilizes a set of sentences in the form of logical language.
Commonsense knowledge is the most general kind ofknowledge that belongs to all people [49]. There exist sev-eral properties such as defeasibility and context sensitivity
4
[50], which distinguish logic reasoning from human com-monsense reasoning. Therefore, the use of natural languageis considered to be the central component for commonsenseknowledge representation [51]. In order for a system toprocess the natural language similar to human beings, itmust find the relations and information about the wordsand their meanings. In a study by Liu et al. a systemcalled ConceptNet is proposed wherein natural language isexplored as a tool for commonsense representation [49]. Ina work performed by George A. Miller, a framework calledWordNet is proposed which benefits from the relationshipbetween words [52]. In another work by Singh et al. a systemcalled Open Mind is designed that obtains commonsenseknowledge from people over the web [53]. A calender withcommonsense called SensiCal is presented by Mueller thatgenerates warnings for avoiding errors [54]. For example,SensiCal sends a message to the user not to take a friendthat is a vegetarian to a steakhouse.
3 PERSENSE FRAMEWORK DESIGN
PerSense aims to perform commonsense reasoning based onan individual’s personality. An overview of the PerSenseframework for algorithm design is shown in Figure 2. De-signing the PerSense framework takes three phases including1) training data construction, 2) personality assessment,and 3) commonsense reasoning. In order to construct thetraining data, text data from different sources are gathered,natural language processing is applied on the data to extractimportant information from the text, and the personalitytraits associated with each text sample are scored using theIBM Watson Personality Insights engine. The text data, theinformation extracted using the NLP algorithms, and thescores achieved through the IBM Watson engine are thenstored in a database. After gathering the labeled trainingdata, PerSense estimates personality traits for each text sam-ple. Two models are proposed for estimating personalityscores including 1) a probabilistic approach based on Aggre-gated Probability Density Functions (PDF); and 2) an ML-based approach that utilizes machine learning algorithms ina supervised fashion to learn a model that classifies text datainto personality classes. After performing personality as-sessment, PerSense performs machine commonsense reason-ing based on each individual’s different personality traits.PerSense uses a supervised learning based methodology forcommonsense reasoning. In what follows the computationalalgorithms utilized in PerSense framework are explained inmore detail.
3.1 Probabilistic-Based Personality Assessment
Let us assume N text samples have been processed and forthe score s, a real number between 0 and 1 ( 0 ≤ s ≤ 1),is associated with a given personality trait T , such as Neu-roticism. We assume g(s1, s2) is the function that shows thenumber of text samples, among N processed text samples,for which the personality trait score is between s1 and s2,where s1 ≤ s2 (g(0, 1) = N ). The score interval [0, 1]is divided equally into n sub-intervals of [s0, s1], [s1, s2],· · · , [sn−1, sn], where s0 = 0 and sn = 1. In a uniformdistribution of samples, g(si, si+1) = N
n , where i is any
Natural Language Processing
IBM Watson Personality Traits
Labeled Training Data
Machine Learning Model Training
Probability Density Function Model
Text Data Acquisition
Personality Assessment
Trai
nin
g D
ata
Co
nst
ruct
ion
Pe
rso
nal
ity
An
alys
is
Machine Learning Model Training
Commonsense Reasoning Co
mm
on
sen
se
An
alys
is
Fig. 2: An overview of the PerSense framework.
positive integer number between 0 and n − 1. In practice,it is unlikely to have a uniform distribution of samples asthe chance of observing people with very high or very lowscores is less than observing people with moderate scores.In any statistical analysis, increasing the number of sampleswill increase the accuracy of the results. Consequently, thenon-uniform distribution of samples will adversely impactthe accuracy of the results. However, in a sufficiently largesurvey there will be a sufficient number of samples in anyinterval.
Any collected data in the k-th sub-interval, where 1 ≤k ≤ n, will be divided by the number of text samples inthat interval (g(k−1
n , kn )) to generate a normalized metric.This metric indicates the likelihood of using a specific word(in this study an adjective) by authors whose personalitytrait score is in the k-th interval for the personality trait T .To achieve this objective, we create a histogram for eachadjective with regard to a particular personality trait overpersonality trait intervals [s0, s1], [s1, s2], · · · , [sn−1, sn]. Forexample, if a word w occurred x times in a text sample andthe personality score for that text sample for a trait T is sy ,which is in the k-th interval, i.e., k−1
n ≤ sy ≤kn , a count of x
will be added to the number of times w has been observedin the text samples authored by subjects whose scores arelocated in the k-th interval. By processing a large number oftext samples, i.e., N,n→∞ and normalizing the generatedhistogram by its integral from 0 to 1 to have an area equal to1, the histogram will be converted to a probability densityfunction, fw,T (s). The probability that an author that uses win her writing has a specific personality trait score betweenα and β is calculated by
∫ βα fw,T (s)ds.
It is equally important in probabilistic-based and ML-based approaches to find words that convey distinctive
5
information regarding personality traits, i.e., their proba-bility density function is not uniform, i.e., fw,T (s) 6= 1, atleast for one personality trait. The words with the smallestvariance in their probability density functions are moreselective and convey more information. By developing thedatabase and creating the probability density functions forinformative words, it becomes possible to analyze new testsamples without relying on IBM Watson. Any new textsample is processed and its informative words are extracted.The related probability density functions for informativewords, which have already been developed, are combined tocreate an aggregated probability density function (PDF) foreach personality trait. The aggregated probability densityfunctions are then used to estimate the personality trait ofthe author.
Let us assume, w1, w2, · · · , wm are m informative wordsin a sample text, where m is a positive integer and thenumber of distinct informative words can be less than m,which means some informative words can be used morethan one time in a sample text. The corresponding probabil-ity density functions are fw1,T (s), fw2,T (s), · · · , fwm,T (s),respectively. The aggregated probability density functionfor the personality trait T of the text sample, ΦT (s), iscalculated by the following equation.
ΦT (s) =
m∏i=1
fwi,T (s)∫ 1x=0
m∏i=1
fwi,T (x)dx(1)
Here, we made a simplifying assumption that utilizingan informative word is a completely independent eventand conveys no information regarding any other infor-mative words in the text sample, not even regarding thesame informative word if it has been used multiple times.In other words, the fact that the subject author used w1
in the first sentence of the text sample, has no infor-mation about using wi in the text sample, where i isan integer number between 1 and m. As an example, iffw1,T (s), fw2,T (s), · · · , fwm,T (s) = 1 and convey no in-formation, ΦT (s) will also be equal to 1 and will conveyno information regarding the personality trait T . The sametechnique can be expanded to additional informative words(verbs, pronouns, etc.) as well as additional text features,such as the use of passive and active voice, capitaliza-tion, punctuation, spelling or grammatical errors, sentencelength, and word position.
3.2 Machine Learning for Personality AssessmentAlternatively, we developed a second approach for person-ality assessment based on supervised learning methodolo-gies. There are a large number of methods for classificationand regression in supervised learning research. Examplesinclude tree-based, naive-based, support vector machines,etc. The classifiers and regressors do the predictions basedon the input features to the system and their specific pat-terns. Supervised learning methodologies are useful tech-niques when a straightforward mathematical model cannotbe extracted or building a statistical model is expensive.These techniques infer decisions based on a set of trainingdata. The training data are divided into two sets:
1) X is a finite set of input features,2) Y is a finite set of output labels associated with each
input feature instance, and3) F (X) is a function that maps input features to
output labels, i.e., XF (X)→ Y
In this work, the input features are a set of adjectivesutilized across the text samples and their frequencies in eachtext sample. The output labels are the personality scoresassociated with each text sample. F (xi) is a classificationalgorithm that generates an output label yi as the predictedscore/label for samplei. The purpose of training is to findthe mapping function from the input features to the outputlabels; therefore, if we feed a set of input features into ourdecision making function, it will provide a prediction of theoutput label. The schematic of the machine learning methodis shown in Figure 3.
3.3 Machine Learning for Commonsense ReasoningWe utilized a supervised learning based methodology forcommonsense reasoning. Several classification and regres-sion algorithms are utilized for modeling the problem.As mentioned previously, a supervised learning methodderives the decisions based on the inputs that are givento the model. These inputs are represented as features inthe context of machine learning modeling. For the pur-pose of commonsense reasoning, we designed a way ofhuman-computer interaction wherein the input features areresponses to a standard FFM personality trait questionnaire[55]. This online FFM questionnaire was developed from theInternational Personality Item Pool (IPIP). A revised versionof this was utilized for this work which was created withacceptable reliability and factor uni-vocal scales. FFM modelwas proved to be acceptable for use in internet-mediatedresearch via different experiments. We perform classificationfor each individual commonsense question. The predefinedresponses to commonsense questions are output labels forthat question. The schematic of the supervised learningmethod utilized for commonsense reasoning is shown inFigure 4.
4 EXPERIMENTAL SETUP
This section describes the steps that are taken for evaluatingPerSense. In order to prevent propagation of errors from per-sonality assessment algorithms to commonsense reasoningtechniques, we evaluated these two algorithmic modulesindividually. The final dataset for personality analysis usingML contains 16, 934 data points, 345 input features, and 8class labels for each personality trait. The dataset for com-monsense reasoning consists of 300 data points, 50 inputfeatures, and depending on each commonsense questionthere are different numbers of input features ranging from 2to 7.
In this work, for evaluating the proposed framework,two sets of experiments where designed. First experimentis designed for evaluating personality assessment. For thispurpose IBM Watson’s text-based personality trait estima-tion tool was utilized for ground truth labeling. The ac-curacy of the proposed method was thus limited by IBMWatsons accuracy. However, recruiting tens of thousands of
6
Data Sources
Data sanitizer
Data Fusion
Am
azo
n r
evie
ws
Ess
ays
Blo
gs
POS tagging Feature
Extraction
Labels
Feature vector
Predictions Features
New data
. . .
. . .
. . .
Supervised Learning
Model
Input Data
Data sanitizer
Data Fusion
Per
son
alit
y t
rait
s
Feature Extraction
Labels
Feature vector
Predictions Features
New Question
. . .
. . .
Supervised Learning
Model
Fig. 3: The pipeline of the personality assessment framework using supervised learning methodology.Data Sources
Data sanitizer
Data Fusion
Am
azo
n r
evie
ws
Ess
ays
Blo
gs
POS tagging Feature
Extraction
Labels
Feature vector
Predictions Features
New data
. . .
. . .
. . .
Supervised Learning
Model
Input Data
Data sanitizer
Data Fusion
Per
son
alit
y t
rait
s
Feature Extraction
Labels
Feature vector
Predictions Features
New Question
. . .
. . .
Supervised Learning
Model
Fig. 4: The pipeline of the commonsense reasoning framework.
volunteers to provide text samples and take standard per-sonality tests was not a feasible option for this preliminarywork. In the second part, standard personality tests resultswere utilized for commonsense reasoning, and the groundtruth is generated by questionnaires filled through AmazonMechanical Turk.
In what follows, the steps for data collection and ex-perimental setup are explained in detail. We will discussthe obtained results based on these experiments in the nextsection.
4.1 Experiments for Personality Assessment
For evaluating the performance of the personality assess-ment algorithms, we construct a dataset of text samples toevaluate both approaches (i.e., PDF and ML) by utilizingexisting texts from Amazon movie reviews [37], stream-of-consciousness essays [56], and blog posts from sites hostedby WordPress. The mined text sample dataset contained17, 128 data points. These text samples are given a score foreach personality by the IBM Watson Personality Insights,a commercialized text-based personality assessment toolrunning on a big supercomputer and providing other fee-based services. In this work we utilized the IBM Watsonservice for assessing the personality characteristics based ona user’s digital footprint and communications [57]. This toolhas limitations on the text inputs. Moreover, some of theadjectives are not descriptive due to having a low frequencyin the samples. Therefore, the text samples that don’t meetthe following criteria are filtered out:
1) text samples should be in English;
2) the number of words in each sample should begreater than 600; and
3) each adjective frequency should be greater than 25throughout all the samples.
The mined text sample dataset contained 16, 834 data pointsafter data filtering.
The scores given by IBM Watson are real values in therange of 0 to 1, with 0 being the lowest in that personalitytrait. These scores are used as ground-truth labeling for test-ing our PDF and ML approaches for text-based personalitytrait assessment. The collected data were stored in a localdatabase that consists of two tables. The first table has thelist of all the adjectives, their frequencies, and a vector ofscores associated with the samples that they were found in.The second table has all the text samples, a vector of ad-jectives and their frequencies in that sample, and the scoresgiven by IBM Watson. Input features for training models forpersonality assessment are frequencies associated with eachadjective. The output labels are the scores given to each textsample by IBM Watson.
To build the training data in PerSense, the aggregateddata are passed through Natural Language Processing(NLP) tools for extracting important data and their frequen-cies. For this purpose, the Natural Language Toolkit (NLTK)[58] is utilized, which uses large bodies of linguistic data,or corpus. The NLTK is written in Python and distributedunder the GPL open source license. We utilized existing cor-pora in the NLTK which are large text bodies for aggregatinga list of existing adjectives, such as Gutenberg, web and chattext, Brown, and Reuters. A total of 26, 414 adjectives werefound in the corpora. Utilizing this tool, we tagged the sen-
7
TABLE 1: The distribution of the data for each class label.
Labels O C E A N0-0.1 0.3% 35% 31% 60% 6%
0.1-0.2 0.6% 26% 20% 17% 6%0.2-0.3 0.8% 15% 15% 8% 11%0.3-0.4 1% 9% 11% 4% 17%0.4-0.5 1% 5% 8% 3% 18%0.5-0.6 1% 4% 5% 2% 16%0.6-0.7 1% 2% 4% 2% 11%0.7-0.8 2% 2% 3% 1.5% 7%0.8-0.9 2% 1% 2% 1% 4%0.9-1 90% 1% 1% 0.4% 3%
tences in all of the text samples and extracted the adjectivesand their associated frequencies in the text samples. For thepurpose of ground-truth labeling, the dataset is analyzedby IBM Watson for assigning personality scores to each textsample. The data obtained from these two sources, NLP andIBM Watson, are then stored in a database. In what follows,the dataset for each personality assessment approach andstatistics of the data are explained in more detail.
4.1.1 Distribution of Scores
To obtain ground-truth labeling, the dataset was fed tothe IBM Watson and scores were assigned to each textsample in the dataset. Out of different personality traits,OCEAN traits were extracted from the results. The resultedscores by IBM Watson were divided into 10 class labels inorder to see the distribution of data. As shown in Table 2,Openness score for 90% of the data was above 0.9, whilethe remaining 10% of the data had a score between 0 and0.9. Although all personality traits are equally important,the most uniform distribution of the mined samples wasobserved for Neuroticism. This personality trait is importantin mental health care and suicide prevention. Therefore, ourtext-based analysis was focused on this personality trait.
4.1.2 Dataset Preparation for PDF Analysis
While 26, 414 adjectives were obtained from the NLTK cor-pora, only 17, 771 were used in our text samples. At first, wedivided the personality scores into 10 subsets (n = 10). TheIBM Watson scores that were below 0.1 and higher than 0.9seemed dubious and were typically from short text samples.Therefore, we limit the analysis to scores between 0.1 (10%)and 0.9 (90%). The labels for each subset are shown inTable 2.
The frequency of all the adjectives were computed sepa-rately for each sub-interval. The adjectives with frequenciesless than 300 over all the text samples were counted asunreliable data as there were zeros in the sub-intervals,resulting in Φ(s) = 0 in the related sub-interval based on (1).The final number of adjectives used in this study was 868.Figure 5 shows examples of PDF vectors for two adjectives:Major and Happy.
After forming probability density function vectors foreach adjective, we calculated the aggregated probabilitydensity function (PDF) for each text sample. The text sam-ples with less than 1000 and more than 6000 words wereeliminated. Text samples with less than 1000 words were tooshort to have reliable results using IBM Watson. A few text
(a)
(b)
Fig. 5: The PDF values for two adjectives and the correspondingNeuroticism score: (a) Major — (0.4-0.5), and (b) Happy — (0.2-0.3).
samples larger than 6000 words, up to one million words,were mined. These samples turned out not to be a usualtext sample written by a subject reviewing a product orcommenting about something and thus were discarded.
Furthermore, we noted that we have more samples at thecenter and fewer samples at the corners, so the g functionwas not uniform. In order to calculate each text sample’sPDF vector, we extract the adjectives in each text sampleand multiply their vectors. Finally, the PDF values are scaledto have an integral equal to one. Then the peak of eachvector shows the personality score for that text sample.An example of a PDF output for Neuroticism is shown inFigure 6. This figure shows that the Neuroticism score forthis specific text sample is in the range of 0.2 to 0.3.
4.1.3 Dataset Preparation for ML AnalysisThe initial set of input features are the total number of adjec-tives that were identified through the NLTK. After taggingthe adjectives throughout all the text samples, the numberof adjectives came down to 3, 834. Thereafter, adjectives thathad a frequency of less than 10% of the total number ofoccurrences of all the adjectives throughout the data pointswere removed from the feature set. The number of adjectivesafter data filtering was 345. Other than filtering the feature
8
Fig. 6: Neuroticism PDF for a specific text sample.
TABLE 2: The division of the personality scores and labels foreach section.
Personality Score Range Label0.1-0.2 0.150.2-0.3 0.250.3-0.4 0.350.4-0.5 0.450.5-0.6 0.550.6-0.7 0.650.7-0.8 0.750.8-0.9 0.85
set, the data points were filtered as well. The number of datapoints before data filtering was 17, 127. The frequencies ofadjectives in each data point were enumerated, and datapoints with fewer than 5.5% of the total number of featureswere removed from the dataset.
The final dataset contains 16, 934 data points and 345input features. Each xij in the feature set is the number ofadjective i in data point j.
4.1.4 Regression and Classification ModelsClassification is a form of predictive modeling and it is fo-cused on learning a mapping function from input features tooutput labels. The classification model is used for predictinga discrete class label output. The regression model, on theother hand, is used for predicting a continuous set of classlabels.
The main motivation of this analysis is to quantify thecorrelation between adjective usage in each text sampleand the personality scores associated with that sample. Thecorrelation of the adjectives and the personality scores isestimated using regression models. As mentioned above,the dataset contains 345 features after the process of featureextraction and feature selection. The regression models uti-lized in this work are Support Vector Regression (SVR) withradial basis function (rbf) kernel [59], Linear Regression [60],and Random Forest Regression [61].
Moreover, we tried to predict the output labels usingclassifiers as well. Six popular supervised learning classifierswere utilized for modeling the personality predictor, namelySupport Vector Machine (rbf, degree 2 polynomial, andlinear kernels) [62], Decision Tree [63], K-nearest neighbors[64], Perceptron [65], Multi-layer Perceptron [66], and Ran-dom Forest classifier [67].
4.2 Experiments for Commonsense Reasoning
In this section, the setup of the experiments for common-sense reasoning based on the users’ personality traits areelaborated. For the purpose of training a prediction model,data were collected using Amazon Mechanical Turk andFacebook. A test was designed for assuring the consis-tency of answers to survey questions. After passing theconsistency test, we obtained 300 survey response and eachresponse was considered as a data point. In what followsthe data analysis steps are explained in detail.
4.2.1 Commonsense Dataset
For building the dataset for commonsense reasoning, wedesigned a survey with two parts. The first section of thesurvey carries 50 personality trait questions. This question-naire contains standardized five factor personality test. Theresponses are in the form of a Likert scale which is com-monly involved in research studies that employs question-naires or surveys: very inaccurate (1), moderately inaccurate(2), neither inaccurate nor accurate (3), moderately accurate(4), very accurate (5) [68]. The first set of questionnaire(personality questions) are input features for training com-monsense reasoning models. In this article, input featuresare a set of scores given to each personality question by theuser. There are 550 possible outputs for standard personalityquestionnaire. Questions are expected to be independentand orthogonal, resembling eigenvectors in a vector space,each targeting a unique dimension of human personality.The FFM scores, historically were designed to be concise andhandy for general statements not necessarily an adequatetool in big-data era and do not convey equal informationfor common sense analysis.
The second section of the survey contains 67 questions,75% from the game show Family Feud [69], and 35% de-veloped in our lab. The developed questions are based onSamsung’s interest, popular political questions, and health-care problems. A model is trained for each commonsensequestion; therefore, the output labels are the answers to eachquestion (e.g. Question: “If money is not an issue, you willbuy a driver-less car ...”, Answers: 1. Right away, 2. Onlywhen 25% of cars are driver-less 3. Only when you have noother option 4. Only when 50% of cars are driver-less).
The distribution of class labels was calculated for eachquestion. If the class labels were not well distributed fora question, we tried to merge related labels and make theresults uniformly distributed. The labels for each questionwere merged based on the relevance of each label to an-other on for a particular question. This relevance of eachlabel to another one was determined by perception of theexperimenter of those labels to be positive or negative. Forinstance in the question about free health-care (What isyour opinion about creating united health care insuranceby slightly increasing the sales/income tax?), two answersagreed and two answers disagreed with this. Therefore,the experimenter merged two negative answers and twopositive answers.
The label distributions of a subset of commonsensequestions are shown in Table 3. This table shows examplesof commonsense questions and the distribution of classlabels for each question. As an example, row two in Table 3
9
(“President Trump’s travel ban is”), shows a question withfour possible choices (1. “Great and improves the safetyof legal and patriotic US citizens”, 2. “Not good, becauseit is discriminatory”, 3.“Good, but should include morecountries”, 4. “Not good, since not all Muslim people in thebanned countries should be punished because of the Sept.11 attack”). The distribution of responses to this questionfor each choice was 25%, 17%, 10%, and 48%, for each of thefour above-mentioned questions. Based on the distributionand relevance of these options, responses for choices 1 and3 were combined and responses selecting choices 2 and 4were combined.
4.2.2 Prediction ModelEight popular supervised learning classifiers were utilizedfor developing a machine commonsense reasoning modelthrough human-computer interactions: three Support VectorMachine with three different kernels (rbf, degree 2 poly-nomial, and linear), Decision Tree, K-nearest neighbors,Perceptron, Multi-layer Perceptron, and a Random Forestclassifier.
5 RESULTS
We conducted two sets of experiments for testing our pro-posed methods. The results of two personality analysis ap-proaches utilized in this research were compared with IBMWatson personality insights. IBM Watson itself utilized thepersonality score that was obtained from personality survey,as ground truth labeling. This comparison is performed dueto the similarity between our framework and IBM Watsonpersonality insight service. The Watson Personality Insightsservice extracts information from text data in user’s textmessages, tweets, and posts using linguistic analysis andperforms personality analysis. In what follows, we demon-strate the detailed results for each part of the data analysis.
5.1 Performance of PDF ApproachFor evaluating the performance of the PDF approach, wecalculated the PDF vector of each data point. Thereupon,the peak value of the vector was compared with the groundtruth label. The results show that the Mean Absolute Error(MAE) and the Root Mean Square Error (RMSE) of the PDFmethod are 15.5% and 19.5%, respectively. The logarithmicratio of the first to the second maximum was then calculatedfor each text sample for use as a measure of “confidence”.The confidence scores were from 0 to 10 with 0 being thelowest confidence and 10 the highest. In some cases, thefirst and second peaks are comparable, while in other casesthey are significantly different. An example of each is shownin Figure 7a and Figure 7b. The number of words in the textsample shown in Figure 7a are 3, 079 and the score givenby IBM Watson is 0.44. One of the peaks given by PDFapproach shows the Neuroticism value in the range of 0.4and 0.5. In Figure 7b, the number of words are 3, 500 and theNeuroticism score given by IBM Watson is 0.18. The scoregiven by the PDF approach is in the range of 0.2 and 0.3which is close to IBM score.
The MAE of the samples with confidence values morethan two is 10.5%. Figure 8 shows the MAE based ondifferent confidence levels.
(a) An example of a low confidence data point.
(b) An example of a high confidence data point.
Fig. 7: Two data points with different confidence scores.
Fig. 8: The Mean Absolute Error-Confidence plot shows byincreasing the confidence metric the MAE drops below 4%.
5.2 Performance of ML-Based Personality AssessmentSince the amount of the mined data are large, we initiallyevaluated the accuracy using a train-test split (67% training,33% testing). After choosing the best classifier, we used a 10-fold cross validation method for better accuracy evaluation.As it is shown in Table 1, the data are not well distributedover all the scores. Therefore, we chose the most evenlydistributed personality trait (Neuroticism) for performingour experiments and training models. For training ML clas-
10
TABLE 3: The distributions of the answers to some of the commonsense questions.
Question without fusion with fusionName a subject that sons discuss with their fathers rather than withtheir mothers. 5%,33%,46%,16%, 54% ,46%
President Trump’s travel ban is 25%,17%,10%,48%, 35% ,65%You share values with 26%,39%,19%,16%, 26% 39%,35%What is your opinion about creating united healthcare by insurancingthe sales/income tax? 54%,20%,14%,12%, 54% 46%
sifiers, we utilized real value scores and converted them tointeger percentages. While for training ML regressors weutilized real values as ground truth labeling, and outputof the regressors were real values. Therefore, the accuracyof a classifier was measured based on the differences ofthe output labels which were numbers in range of 1 to8, with the ground-truth labels generated by IBM Watson.The accuracy of a regressor was measured based on thedifference between the actual score in range of 0-1 and theassociated ground-truth labe.
Scikit provides the root-mean-square deviation (RMSD)or root-mean-square error (RMSE) of each trained modelwhich are frequently used as measures of the differencesbetween actual values and the values predicted by a model.RMSD is based on the standard deviation of the predic-tions from ground-truth labels. While RMSE based on theaverage differences of predictions with ground-truth labels.The main purpose of this study is to predict a score asclose as possible to the actual personality scores. There-fore, we defined an error margin for calculating the ac-curacy. If a predicted score is within the range of margin(-margin,+margin) away from the actual score, then it iscounted as a correct prediction. Otherwise, it is counted asa mistake. The margin of error is set to 10% for calculatingthe performance. It means if the score should be 90% and theprediction is in the range of 80% to 100%, this is counted as acorrect prediction. Results are shown in Table 4. The settingfor each classifier and regressor is explained as follows:
• SVM-Linear: A support vector machine with linearkernel.
• SVM-Polynomial: A support vector machine withpolynomial degree 2 kernel.
• SVM-rbf: A support vector machine with RadialBasis Function kernel. This is one of the most pop-ular kernels. For distance metric squared euclideandistance was used.
• Decision Tree Classifier: A supervised learning clas-sifier that trains decision rules for classifying thedata. In this work, the random state was set to zero.
• KNN: K-Nearest Neighbor is a supervised learningclassifier. The classification is done based on the voteof its neighbors. The class label of each datapointwill be the most common class among its k nearestneighbor. In this paper, k was assigned to 5.
• Perceptron: A supervised learning classifier, gener-ally used for binary classification.
• MLP: Multi-layer Perceptron a artificial neural net-work classifier. We utilized a log-loss function as asolver. L2 regularization penalty is set to 10−5. Thesize of hidden layers was 15.
• SVR-rbf: A support vector regressor with rbf kernel.
TABLE 4: The error metrics for each classifier
Classifiers and Regressors Marginal Error RMSE
SVM-Linear 37% 27%
SVM-Polynomial 50% 47%
SVM-rbf 39% 30%
Decision Tree 50% 34%
KNN 45% 40%
Perceptron 43% 29%
MLP 30% 22%
SVR-rbf 43% 26%
Linear Regression 40% 24%
RFR 46% 27%
RFC 35% 27%
Kernel coefficient was set to 0.1 and the penaltyparameter was set to 103.
• Linear Regression: Attempted to model a relation-ship in data by fitting a linear equation to observeddata.
• RFR: Random forest fits a number of classifyingdecision trees on the samples of the dataset and usesaveraging to improve the predictive accuracy andcontrol over-fitting. This is a random forest regressorwith the maximum depth set to two.
• RFC: Random forest classifier with the number oftrees set to 1000.
The highlighted cells are the lowest error values, obtainedby the multi-layer perceptron.
5.3 Performance of ML-Based Commonsense Reason-ingWe initially calculated the correlation between each inputfeature and the output labels and the features with low cor-relations were filtered out in building the prediction models.This results in different feature sets for each commonsensequestion. A model was trained for each commonsense ques-tion using each training model (8×67 models). For evaluat-ing the performance, a 10-fold cross validation method wasutilized. The non-uniform distribution of the class labels insome questions made the problem too easy and resultedin very high accuracies; therefore, these questions are not agood criterion for the evaluation of the models. On the otherhand, some questions are well distributed between morethan 3 class labels. Due to the low number of data points,these questions lead to significantly low accuracies. The bestcriterion for evaluating the classifiers and regressors are thequestions with uniform distribution among few numbers ofclass labels. We compared the accuracies of the classifiersfor some of the questions before and after the data fusion.
11
The results are shown in Figure 9. For instance, in Figure 9bthe choices of the questions were: 1) Great and improvesthe safety of legal and patriotic US citizens (25%); 2) Notgood, since not all Muslim people in the banned countriesshould be punished because their fellow citizens committedthe Sep.11 attack (17%); 3) Good, but should include morecountries (10%); and 4) Not good, because it is discrimina-tory (48%). We merged options 1 with 3 and options 2 with4, the distribution of the data became (65% and 35%). Theaverage accuracy before and after the data fusion 52% and76%, respectively.
(a)
(b)
(c)
Fig. 9: Accuracy values of models before and after data fusion.
TABLE 5: Confusion matrix for the commonsense question:“What is your opinion about creating united healthcare insur-ance by increasing the tax?”
XXXXXXXXXXCorrectPredicted
Agree Disagree
Agree 60 20Disagree 19 41
TABLE 6: Confusion matrix for the commonsense question:“What do you think about Trump’s travel ban?”
XXXXXXXXXXCorrectPredicted
Agree Disagree
Agree 31 19Disagree 9 81
TABLE 7: Confusion matrix for the commonsense question:“Tell me the most important habit a mom should teach herson?”
XXXXXXXXXXCorrectPredicted
Respect Others
Respect 47 21Others 30 41
The confusion matrix for each commonsense reasoningquestion is shown in Table 5 to Table 7.
6 CONCLUDING REMARKS
The methodologies proposed in this article address impor-tant problems in the area of behavioral analysis throughtwo novel approaches based on machine learning and prob-ability density functions. The proposed system may havemany different applications, including mental health careand suicide/homicide prevention. In this work, we appliedour methodologies on personality assessment and common-sense reasoning. Scores from IBM Watson Personality In-sights were utilized for ground-truth labeling. However, ourgoal in this work was to match or surpass the performanceof Watson with a system requiring far less computationalresources. The proposed approaches achieved a comparableresult to the ground truth labels. For the PDF approachwith the confidence factor of higher than 3, the accuracyreached 97%. The best results achieved by ML modelingwas 88.2% through the multilayer perceptron. In this article,we utilized 345 adjectives in order to train accurate modelsfor personality analysis. 345 adjectives were outcome of ourexperiments based on the available text. We are planningon collecting more text data and include other words suchas verbs, nouns, and pronouns in our experiments. Fur-thermore, by increasing the samples more adjectives canbe added to the list. Although 345 input features are notideal for training a model, in this work we targeted a low-complexity solution for running the proposed method onsmartphones rather than an alternative for IBM Watson thatneeds a supercomputer to run.
In this work, commonsense decisions were predictedbased upon an individual’s personality trait. For perform-ing commonsense reasoning, we proposed an approachutilizing ML techniques. We collected data from 345 peo-ple through Amazon Mechanical Turk and Facebook. Twosets of questions were provided. The first set containedthe personality analysis questions which were utilized asinput features for the commonsense reasoning problem.The second set contained the commonsense questions. Eachcommonsense question was modeled separately. In this con-text, commonsense questions had a number of reasonableand justifiable answers. Each answer was a class label for
12
training a model. Based on the evaluations, Random ForestClassifier achieved 82.3% accuracy in predicting the answersto some commonsense questions. We note that common-sense questions, which may target different political point-of-views, ethical priorities, marketing practices, and otherpersonal choices, preferences, and habits, may have a num-ber of reasonable and justifiable answers instead of a uniquefactual answer.
In this study, we performed experiments to generateconvincing evidences that the proposed low-complexity ap-proach is in fact functional. However, more complex andcomprehensive experiments are needed to assess large scaleadoption of these technologies. We focused on a compu-tationally simple method that can be easily implementedas a mobile application. Our goal in this study was not todevelop an alternative approach to the IBM Watson tool,which uses a supercomputer to run. In the future, we planto utilize a comprehensive set of word categories (e.g., verbs,nouns) throughout the text samples and use those datafor design and development of our computational models.Moreover, in this study, we utilized one personality trait inour experiments. A future study will involve collecting datafrom various sources to incorporate all five personality traitsin our experiments.
There were few sources of error in designing the al-gorithms. The number of subjects participated in the datacollection was limited so that we could not cover a largenumber of personality traits at a large scale. To train com-monsense reasoning models, we collected data throughAmazon mechanical Turk; however, we were able to provideonly a nominal amount of monetary incentive to the partic-ipants. This may have potentially impacted the quality ofthe collected data. Moreover, the number of subjects takingthe survey was limited, and the survey results may havebeen affected by many factors such as personal biases. Inour ongoing research, we are currently designing a datacollection framework, which prompts the user with ques-tions throughout the day, asking them to express their dailyactivities such as their experiences, their dietary intake, andinteractions. The inputs to this system are in form of spokenlanguage. The speech is then converted to text and utilizedfor personality analysis. For the commonsense reasoninganalysis, we plan on recruiting more participants to trainhighly accurate models for commonsense reasoning.
7 DISCUSSION AND FUTURE DIRECTION
In this work IBM Watson Personality Insight tool, whichauthors modeled by two low-complexity approaches, onlyprovides 5 FFM scores instead of providing probable re-sponses to a standard FFM questionnaire. However, thereare 550 distinct cases when a standard FFM questionnaireconsisting of 50 questions and 5 possible answers per ques-tion is filled, thus the questionnaire conveys more informa-tion than only 5 FFM percentiles. Questions in a standardFFM questionnaire, are expected to be independent andorthogonal to each other meaning that each question tar-gets a unique dimension of human personality and closelymimic Eigenvectors in an N-dimensional space. Authorsbuilt the commonsense reasoning scheme directly based onresponses to the questionnaire.
As stated, personalizing the algorithms and servicesin most efficient way has been the focus of many re-cent researches in human-centered applications [47], [70]–[72]. Therefore, assessing the personality of humans wouldbe a helpful and necessary step. This work is the firststep towards integrating personality analysis and common-sense reasoning with all the limitations that research inacademia has Nevertheless. However, it has achieved itsobjective to establish the connection between personalitytrait and common-sense reasoning. In this work, we havenot achieved the ultimate goal to directly extract common-sense reasoning from text. To achieve this ultimate goal, oneshould exceed IBM Watson and need to have access to areally large number of text samples. Although, authors didnot have access to the required resources; they highlightedthe missing parts and will pursue these in their futuredirection.
Although this study clearly showed that commonsensereasoning is related to personality trait, the authors do/didnot have the means to fill the FFM questionnaire directlyfrom a text sample. Authors believe FFM scores, that areeasier to use, convey less information than the question-naire and therefore, IBM Watson as well as our methodcannot rigorously revive the loss of information and fill thequestionnaire using FFM percentiles. This is an interestingtopic that can be explored in the future and requires a largenumber of text samples and responses to questionnaires tofacilitate predicting responses to questions based on textfeatures.
Although applications need many words to learn per-sonalities, it does not need to happen instantaneously. Theusers do not need to write an essay every time. The finalproduct can use peoples voice, text, and other modalitiesand collect data overtime. As more users adopt the systemand add more data to the applications, the performanceimproves. In our ongoing work, we are collecting data fromusers in real-time on daily basis by prompting them withspecific questions throughout the day. For example, howwas your day? What did you eat? Tell me about a movieyou watched, a book you read, or a conversation you had.Then this data is utilized for personality assessment andcommonsense reasoning. We expect that the amount of datacollected over time will be an order of magnitude greaterthan what we have right now. Huge amounts of data will beavailable to improve the performance dramatically.
In this work, the questionnaire is filled by a small num-ber of users with different personality traits, rather thana large number of people in the society. The connectionbetween a commonsense question and a specific feature inthe personality trait is established using machine learning.In general, people in the society will not need to fill aquestionnaire if their personality can be estimated by othermeans including their digital footprint.
REFERENCES
[1] H. Isselhorst, “The human factor, forcasting investor behavior,”Behavioural Finance Solutions.
[2] J. Krueger, “How illusory is the voter’s illusion?” Proceedings of ThePsychology of Voting and Election Campaigns Conference, Oct 2006.
[3] P. Foot, “The problem of abortion and the doctrine of doubleeffect,” 1967.
13
[4] J. J. Thomson, “The trolley problem,” The Yale Law Journal, vol. 94,no. 6, pp. 1395–1415, 1985.
[5] N. J. Goodall, “Can you program ethics into a self-driving car?”IEEE Spectrum, vol. 53, no. 6, pp. 28–58, 2016.
[6] B. W. Roberts, K. E. Walton, and W. Viechtbauer, “Patterns ofmean-level change in personality traits across the life course: ameta-analysis of longitudinal studies.” Psychological bulletin, vol.132, no. 1, p. 1, 2006.
[7] K. M. DeNeve and H. Cooper, “The happy personality: A meta-analysis of 137 personality traits and subjective well-being.” Psy-chological bulletin, vol. 124, no. 2, p. 197, 1998.
[8] L. R. Goldberg, “The structure of phenotypic personality traits.”American psychologist, vol. 48, no. 1, p. 26, 1993.
[9] T. Correa, A. W. Hinsley, and H. G. De Zuniga, “Who interacts onthe web?: The intersection of users personality and social mediause,” Computers in Human Behavior, vol. 26, no. 2, pp. 247–253,2010.
[10] D. J. Hughes, M. Rowe, M. Batey, and A. Lee, “A tale of two sites:Twitter vs. facebook and the personality predictors of social mediausage,” Computers in Human Behavior, vol. 28, no. 2, pp. 561–569,2012.
[11] A. Buchheim and E. Mergenthaler, “The relationship among at-tachment representation, emotion-abstraction patterns, and narra-tive style: A computer-based text analysis of the adult attachmentinterview,” Psychotherapy Research, vol. 10, no. 4, pp. 390–407, 2000.
[12] G. C. Gleser, L. A. Gottschalk, and W. John, “The relationship ofsex and intelligence to choice of words: A normative study ofverbal behavior,” Journal of Clinical Psychology, vol. 15, no. 2, pp.182–191, 1959.
[13] R. Brown, Words and Things: An Introduction to Language: with aNew Preface by the Autor. Free Press, 1968.
[14] D. S. Berry, J. W. Pennebaker, J. S. Mueller, and W. S. Hiller, “Lin-guistic bases of social perception,” Personality and Social PsychologyBulletin, vol. 23, no. 5, pp. 526–537, 1997.
[15] F. Mairesse, M. Walker et al., “Words mark the nerds: Compu-tational models of personality recognition through language,” inProceedings of the Annual Meeting of the Cognitive Science Society,vol. 28, no. 28, 2006.
[16] J. W. Pennebaker, M. R. Mehl, and K. G. Niederhoffer, “Psycho-logical aspects of natural language use: Our words, our selves,”Annual review of psychology, vol. 54, no. 1, pp. 547–577, 2003.
[17] D. Antonioni, “Relationship between the big five personalityfactors and conflict management styles,” International journal ofconflict management, vol. 9, no. 4, pp. 336–355, 1998.
[18] J. W. Pennebaker and L. A. King, “Linguistic styles: Languageuse as an individual difference.” Journal of personality and socialpsychology, vol. 77, no. 6, p. 1296, 1999.
[19] A. Beauducel, D. Liepmann, J. Felfe, and W. Nettelnstroth, “Theimpact of different measurement models for fluid and crystallizedintelligence on the correlation with personality traits,” EuropeanJournal of Psychological Assessment, vol. 23, no. 2, pp. 71–78, 2007.
[20] J. M. Digman, “Personality structure: Emergence of the five-factormodel,” Annual review of psychology, vol. 41, no. 1, pp. 417–440,1990.
[21] B. E. de Raad and M. E. Perugini, Big five assessment. Hogrefe &Huber Publishers, 2002.
[22] A. A. J. Hendriks, “The construction of the five-factor personalityinventory (ffpi),” Ph.D. dissertation, Rijksuniversiteit GroningenGroningen, 1997.
[23] C. G. DeYoung, L. C. Quilty, J. B. Peterson, and J. R. Gray,“Openness to experience, intellect, and cognitive ability,” Journalof personality assessment, vol. 96, no. 1, pp. 46–52, 2014.
[24] R. R. McCrae and O. P. John, “An introduction to the five-factormodel and its applications,” Journal of personality, vol. 60, no. 2, pp.175–215, 1992.
[25] M. Zeidner and G. Matthews, “Intelligence and personality.” 2000.[26] C. G. DeYoung, L. C. Quilty, and J. B. Peterson, “Between facets
and domains: 10 aspects of the big five.” Journal of personality andsocial psychology, vol. 93, no. 5, p. 880, 2007.
[27] Y. Amichai-Hamburger and G. Vinitzky, “Social network use andpersonality,” Computers in human behavior, vol. 26, no. 6, pp. 1289–1295, 2010.
[28] C. Ross, E. S. Orr, M. Sisic, J. M. Arseneault, M. G. Simmering, andR. R. Orr, “Personality and motivations associated with facebookuse,” Computers in human behavior, vol. 25, no. 2, pp. 578–586, 2009.
[29] T. A. Judge, J. B. Rodell, R. L. Klinger, L. S. Simon, and E. R.Crawford, “Hierarchical representations of the five-factor model
of personality in predicting job performance: integrating threeorganizing frameworks with two theoretical perspectives.” Journalof Applied Psychology, vol. 98, no. 6, p. 875, 2013.
[30] F. Drasgow and J. B. Olson-Buchanan, Innovations in computerizedassessment. Psychology Press, 1999.
[31] A. Cooper and K. Petrides, “A psychometric analysis of the traitemotional intelligence questionnaire–short form (teique–sf) usingitem response theory,” Journal of personality assessment, vol. 92,no. 5, pp. 449–457, 2010.
[32] G. Shahar, E. F. Gallagher, S. J. Blatt, G. P. Kuperminc, and B. J.Leadbeater, “An interactive-synergetic approach to the assessmentof personality vulnerability to depression: Illustration using theadolescent version of the depressive experiences questionnaire,”Journal of Clinical Psychology, vol. 60, no. 6, pp. 605–625, 2004.
[33] S. Argamon, S. Dhawle, M. Koppel, and J. Pennebaker, “Lexicalpredictors of personality type,” 2005.
[34] R. P. Hart, “Redeveloping diction: theoretical considerations,”Progress in communication sciences, pp. 43–60, 2001.
[35] C. K. Chung and J. W. Pennebaker, “Linguistic inquiry and wordcount (liwc): pronounced luke,... and other useful facts,” in Appliednatural language processing: Identification, investigation and resolution.IGI Global, 2012, pp. 206–229.
[36] J. W. P. M. E. Francis and R. J. Booth, “Linguistic inquiry and wordcount,” 1993.
[37] J. McAuley and A. Yang, “Addressing complex and subjectiveproduct-related queries with customer reviews,” in Proceedings ofthe 25th International Conference on World Wide Web. InternationalWorld Wide Web Conferences Steering Committee, 2016, pp. 625–635.
[38] R. Lambiotte and M. Kosinski, “Tracking the digital footprints ofpersonality,” Proceedings of the IEEE, vol. 102, no. 12, pp. 1934–1939,2014.
[39] D. Lazer, A. S. Pentland, L. Adamic, S. Aral, A. L. Barabasi,D. Brewer, N. Christakis, N. Contractor, J. Fowler, M. Gutmannet al., “Life in the network: the coming age of computational socialscience,” Science (New York, NY), vol. 323, no. 5915, p. 721, 2009.
[40] A. Pentland and T. Heibeck, Honest signals: how they shape our world.MIT press, 2010.
[41] J.-P. Onnela, J. Saramaki, J. Hyvonen, G. Szabo, D. Lazer, K. Kaski,J. Kertesz, and A.-L. Barabasi, “Structure and tie strengths in mo-bile communication networks,” Proceedings of the national academyof sciences, vol. 104, no. 18, pp. 7332–7336, 2007.
[42] N. Eagle, A. S. Pentland, and D. Lazer, “Inferring friendshipnetwork structure by using mobile phone data,” Proceedings of thenational academy of sciences, vol. 106, no. 36, pp. 15 274–15 278, 2009.
[43] D. J. Watts, “A twenty-first century science,” Nature, vol. 445, no.7127, p. 489, 2007.
[44] J. Teevan, “How people recall, recognize, and reuse search results,”ACM Transactions on Information Systems (TOIS), vol. 26, no. 4,p. 19, 2008.
[45] J. Golbeck, C. Robles, and K. Turner, “Predicting personality withsocial media,” in CHI’11 extended abstracts on human factors incomputing systems. ACM, 2011, pp. 253–262.
[46] D. Quercia, M. Kosinski, D. Stillwell, and J. Crowcroft, “Ourtwitter profiles, our selves: Predicting personality with twitter,” in2011 IEEE third international conference on privacy, security, risk andtrust and 2011 IEEE third international conference on social computing.IEEE, 2011, pp. 180–185.
[47] B. Ferwerda, M. Schedl, and M. Tkalcic, “Predicting personalitytraits with instagram pictures,” in Proceedings of the 3rd Workshopon Emotions and Personality in Personalized Systems 2015. ACM,2015, pp. 7–10.
[48] S. J. Russell and P. Norvig, Artificial intelligence: a modern approach.Malaysia; Pearson Education Limited,, 2016.
[49] H. Liu and P. Singh, “Conceptneta practical commonsense reason-ing tool-kit,” BT technology journal, vol. 22, no. 4, pp. 211–226, 2004.
[50] ——, “Commonsense reasoning in and over natural language,” inInternational Conference on Knowledge-Based and Intelligent Informa-tion and Engineering Systems. Springer, 2004, pp. 293–306.
[51] J. F. Sowa et al., Knowledge representation: logical, philosophical,and computational foundations. Brooks/Cole Pacific Grove, 2000,vol. 13.
[52] G. A. Miller, “Wordnet: a lexical database for english,” Communi-cations of the ACM, vol. 38, no. 11, pp. 39–41, 1995.
[53] P. Singh, T. Lin, E. T. Mueller, G. Lim, T. Perkins, and W. L.Zhu, “Open mind common sense: Knowledge acquisition fromthe general public,” in OTM Confederated International Conferences”
14
On the Move to Meaningful Internet Systems”. Springer, 2002, pp.1223–1237.
[54] E. T. Mueller, “A calendar with common sense,” in Proceedings ofthe 5th international conference on Intelligent user interfaces. ACM,2000, pp. 198–201.
[55] T. Buchanan, J. A. Johnson, and L. R. Goldberg, “Implementing afive-factor personality inventory for use on the internet,” EuropeanJournal of Psychological Assessment, vol. 21, no. 2, pp. 115–127, 2005.
[56] R. Humphrey, Stream of consciousness in the modern novel. Univ ofCalifornia Press, 1958.
[57] A. Verma, G. Dasgupta, T. K. Nayak, P. De, and R. Kothari, “Serverworkload analysis for power minimization using consolidation,”in Proceedings of the 2009 conference on USENIX Annual technicalconference. USENIX Association, 2009, pp. 28–28.
[58] S. Bird and E. Loper, “Nltk: the natural language toolkit,” inProceedings of the ACL 2004 on Interactive poster and demonstrationsessions. Association for Computational Linguistics, 2004, p. 31.
[59] D. Basak, S. Pal, and D. C. Patranabis, “Support vector regression,”Neural Information Processing-Letters and Reviews, vol. 11, no. 10, pp.203–224, 2007.
[60] M. H. Kutner, C. Nachtsheim, and J. Neter, Applied linear regressionmodels. McGraw-Hill/Irwin, 2004.
[61] A. Liaw, M. Wiener et al., “Classification and regression by ran-domforest,” R news, vol. 2, no. 3, pp. 18–22, 2002.
[62] M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf,“Support vector machines,” IEEE Intelligent Systems and their appli-cations, vol. 13, no. 4, pp. 18–28, 1998.
[63] S. R. Safavian and D. Landgrebe, “A survey of decision treeclassifier methodology,” IEEE transactions on systems, man, andcybernetics, vol. 21, no. 3, pp. 660–674, 1991.
[64] L. E. Peterson, “K-nearest neighbor,” Scholarpedia, vol. 4, no. 2, p.1883, 2009.
[65] Y. Freund and R. E. Schapire, “Large margin classification usingthe perceptron algorithm,” Machine learning, vol. 37, no. 3, pp. 277–296, 1999.
[66] M. W. Gardner and S. Dorling, “Artificial neural networks (themultilayer perceptron)a review of applications in the atmosphericsciences,” Atmospheric environment, vol. 32, no. 14-15, pp. 2627–2636, 1998.
[67] M. Pal, “Random forest classifier for remote sensing classifica-tion,” International Journal of Remote Sensing, vol. 26, no. 1, pp.217–222, 2005.
[68] I. E. Allen and C. A. Seaman, “Likert scales and data analyses,”Quality progress, vol. 40, no. 7, p. 64, 2007.
[69] A. McClintock, “Family feuds: Gender, nationalism and the fam-ily,” Feminist review, vol. 44, no. 1, pp. 61–80, 1993.
[70] Z. Esna Ashari and H. Ghasemzadeh, “Mindful active learning,”in 2019 IJCAI conference (International Joint Conference on ArtificialIntelligence), Aug 2019.
[71] M. Pedram, M. Rofouei, F. Fraternali, Z. Esna Ashari, andH. Ghasemzadeh, “Resource-efficient computing in wearable sys-tems,” in 2019 IEEE International Conference on Smart Computing(SmartComp), Jun 2019.
[72] Y. Ma, Z. Esna Ashari, M. Pedram, N. Amini, D. Tarquinio,K. Nouri-Mahdavi, M. Pourhomayoun, R. D. Catena, andH. Ghasemzadeh, “Cyclepro: A robust framework for domain-agnostic gait cycle detection,” IEEE Sensors Journal, vol. 19, no. 10,pp. 3751–3762, 2019.
15
APPENDIX ASURVEY PERSONALITY QUESTIONS
I see myself as someone ... Very Inaccu-rate
ModeratelyInaccurate
NeitherInaccuratenor Accurate
ModeratelyAccurate
VeryAccurate
who is generally trusting � � � � �Don’t like to draw attention to myself � � � � �
Believe in the importance of art � � � � �tends to find fault in others � � � � �
pays attention to details � � � � �Shirks duties � � � � �
who is generally trusting � � � � �who is reserved � � � � �
has a good word for everyone � � � � �insults people � � � � �
feels comfortable around people � � � � �dislikes self � � � � �
has a little to say � � � � �has frequent mood swings � � � � �
often feels blue � � � � �does not like art � � � � �
cuts others to pieces � � � � �is the life of the party � � � � �
believes that others have good intentions � � � � �avoids philosophical discussions � � � � �
makes friends easily � � � � �gets chores done rightaway � � � � �accepts people as they are � � � � �
relaxed, handles stress well � � � � �does a thorough job � � � � �
does just enough work to get by � � � � �enjoys going to art museums � � � � �
often down in the dumps � � � � �always prepared � � � � �
has few artistic interests � � � � �not easily bothered by things � � � � �
knows how to captivate people � � � � �is not interested in abstract ideas � � � � �makes plans and sticks to them � � � � �
doesn’t see things as they are � � � � �is outgoing and sociable � � � � �
doesn’t talk a lot � � � � �gets back at others � � � � �
tends to vote for the liberal candidates � � � � �has an active imagination � � � � �
seldoms feel blue � � � � �carries out the plans � � � � �gets nervous easily � � � � �
feels comfortable with self � � � � �panics easily � � � � �
keeps in the background � � � � �finds it difficult to get down to work � � � � �
tends to vote for conservative candidates � � � � �skilled in handling social situations � � � � �
wastes time � � � � �tends to be lazy � � � � �respects others � � � � �
16
TABLE 8
17
APPENDIX BCOMMONSENSE SURVEY AND RESULTS
Question RFC DT KNN SVM-Linear SVM-Poly SVM-rbf Perceptron MLP
What are your gender/color priori-
ties for the US commander in chief?84% 70% 83% 77% 77% 83% 83% 72%
President Trumps travel ban is 82% 72% 78% 76% 77% 79% 70% 71%
When you’re looking for work,
name a way you might find out
about the job opening
78% 66% 75% 79% 75% 79% 58% 70%
An engineer who lost her father
in childhood in a shooting accident
hates guns and blames gunmakers
for her loss.However, the only job of-
fer that she has received is from a big
gunmaker company. If she declines
the job offer (refusing to make guns
that can kill other kids’ dads), her
retired mother must sell her house to
pay off her school loan.She should ...
78% 66% 75% 79% 75% 79% 58% 70%
Name something your body tells
you it’s time to do.73% 59% 70% 70% 70% 72% 47% 64%
What is your opinion about creat-
ing united health care insurance by
slightly increasing the sales/income
tax?
73% 59% 70% 70% 70% 72% 47% 64%
You share values with 70% 59% 54% 64% 59% 66% 47% 57%
Name something that people like to
do in their backyard70% 59% 54% 64% 59% 66% 47% 57%
Name something that’s embarrass-
ing to fall asleep while doing.69% 61% 65% 65% 65% 69% 63% 60%
A patient who has agreed to do-
nate his vital organs is in a coma
and lives only with the assistance of
life-support machines,but will pass
away in the next few days anyway.
There are a few patients who need
the organs withing a few hours to
survive. Which one do you agree
with:
67% 54% 64% 65% 62% 67% 62% 68%
Name a reason people change their
names66% 54% 65% 60% 61% 66% 59% 61%
Why did Hillary Clinton lose the
presidential election?65% 59% 57% 62% 61% 65% 58% 58%
If a scientific theory does not agree
with your religious belief:65% 55% 60% 61% 65% 65% 58% 56%
Tell me the most important habit a
mom should teach her son.63% 54% 56% 64% 62% 60% 55% 60%
Name something that people do to
their faces62% 56% 57% 62% 59% 65% 58% 62%
If you were arrested and could make
only one call, whom would you call?62% 50% 56% 59% 58% 62% 59% 50%
Name something you’ve stayed up
all night worrying about.61% 55% 56% 55% 54% 59% 53% 52%
Is this a big deal if the fired missile
from Yemen to Saudi Arabia was in
fact made in Iran?
60% 54% 57% 59% 59% 61% 44% 58%
18
A US-based solar panel manufactur-ing company can no longer competewith fully-automated companies inAsia that manufacture solar panels50% cheaper. Hundreds of workerswill lose their jobs in the US if thecompany is closed. The companyshould ...
60% 55% 56% 58% 60% 63% 60% 55%
Name something that your boss
could tell you that would come as
a big shock.
60% 55% 54% 55% 55% 55% 52% 52%
In a smartphone, the main issue for
you is59% 55% 57% 59% 60% 59% 50% 56%
Name something people do when
riding in the back of a taxicab59% 53% 52% 57% 55% 57% 57% 43%
If money is not an issue, you would
like to buy59% 55% 51% 50% 54% 57% 54% 50%
Name something specific you might
remember about one of your teach-
ers from high school.
58% 52% 53% 59% 59% 59% 53% 53%
Name a reason a young man might
join the army57% 52% 51% 54% 56% 57% 50% 55%
Name a reason you might ask the
manager of a hotel for another room.57% 53% 55% 62% 60% 60% 58% 57%
You are tired of a driver who is tail-
gating you for over an hour, what
would you do?
57% 54% 56% 56% 56% 58% 58% 55%
Name a way you can tell that your
child is lying57% 52% 53% 45% 48% 54% 53% 50%
If you have ever been fined for any
driving violations, it has most likely
been for ...
57% 48% 57% 56% 58% 57% 52% 59%
The politician who wins the election
is usually the person with the most
what?
57% 52% 56% 53% 54% 56% 48% 49%
Name something that you’d like todo all day without feeling any guilt. 56% 51% 54% 53% 55% 58% 56% 51%
Which food do you think should bechosen as the ”National Food” ofAmerica?
56% 49% 56% 55% 57% 56% 54% 52%
Name something most people cleanat least once a day. 55% 55% 54% 55% 52% 54% 50% 52%
Name a reason why people some-times need to take a day off from thework.
54% 51% 53% 52% 51% 58% 54% 54%
What’s the hardest part of your bud-get to cut out? 54% 39% 52% 43% 47% 54% 33% 54%
Name a tip people give you on how
to get rich.53% 52% 52% 49% 48% 56% 49% 52%
Where would you look to find out
how old someone is?52% 53% 51% 55% 54% 55% 49% 53%
Name a technique people use to re-
member things.51% 50% 48% 48% 50% 51% 53% 52%
Name something that’s important to
parents when house hunting?50% 39% 42% 49% 46% 50% 46% 38%
Tell me a reason why a man would
want to have a son.50% 48% 49% 51% 49% 52% 49% 48%
Name a reason you might make
your spouse get out of bed at the
middle of the night
48% 43% 37% 44% 42% 40% 39% 39%
19
Many jobs will be lost in the next
few decades due to smart robots and
unemployment will thus increase.
Therefore,
48% 36% 39% 40% 41% 48% 39% 37%
When you pick a password for your
phone, you pick a password that is.48% 31% 41% 40% 40% 48% 37% 42%
On a sunny day, in a freeway withthe maximum speed of 60 miles perhour if there is no police officer, youwill typically drive with a speed ...
50% 51% 52% 50% 52% 50% 48% 57%
If money is not an issue, you willbuy a driver-less car 47% 43% 39% 42% 44% 48% 45% 42%
What should the US do if welearn that the false nuclear alarm inHawaii was real but the North Ko-rean missile exploded before reach-ing Hawaii:
47% 39% 44% 40% 41% 48% 35% 41%
If you are not in a hurry, which routewould you choose to your destina-tion:
45% 37% 34% 39% 40% 47% 40% 35%
True democracy can be implementedusing smartphone voting, as peoplecan vote every day on important is-sues and there is no need to have acostly congress or senate anymore.The people will finally be in charge!
43% 35% 36% 37% 36% 41% 38% 38%
You really like a luxury car thatcosts $70,000 but you can only afford$40,000.
43% 34% 37% 33% 35% 43% 43% 38%
Name something you hope younever find in your house. 40% 24% 34% 30% 32% 39% 24% 29%
Name a subject that sons discusswith their fathers rather than withtheir mothers.
40% 37% 39% 43% 39% 45% 30% 38%
Besides being fast, name somethingpeople like about fast food. 39% 35% 35% 35% 36% 40% 36% 35%
Name something one family mem-
ber might steal from another.37% 38% 30% 32% 32% 37% 34% 33%
A study at the University of Wash-
ington showed that by increasing the
minimum wage, unemployment rate
will increase, therefore
37% 29% 36% 33% 34% 37% 29% 34%
Name a reason you might pull your
car over to the side of the road.37% 36% 36% 30% 30% 37% 26% 31%
Name something that would be
smart to know how to ask for in a
foreign country.
37% 35% 33% 36% 32% 34% 36% 32%
Name something you do while driv-
ing, even though you know it’s
wrong
36% 32% 35% 34% 34% 40% 31% 33%
As micro sized robots will be de-
veloped soon,people can be filmed
anytime and anywhere without their
knowledge, which will destroy their
privacy.
35% 26% 34% 36% 36% 37% 32% 31%
Name something people do to guard
against wrinkles34% 27% 28% 31% 31% 34% 32% 36%
Name a reason why someone might
go a whole day without eating.33% 36% 34% 36% 31% 34% 34% 34%
Name a Doctor’s order people never
take seriously.32% 25% 27% 31% 33% 32% 29% 30%
We asked 100 married women... If
your husband asked you for a di-
vorce Sunday night, what’s the first
thing you’d do Monday morning?
30% 30% 27% 28% 26% 29% 31% 31%
20
Name a question a person who’s
afraid of flight might ask a flight
attendant.
30% 24% 26% 22% 23% 32% 30% 32%
If time flies when you’re having fun,
name a place it crawls29% 24% 30% 32% 30% 30% 26% 29%
Name something specific everyone
complains about21% 18% 22% 23% 18% 25% 16% 27%
TABLE 9
Que
stio
nA
nsw
ers
Nam
ea
subj
ectt
hats
ons
disc
uss
wit
hth
eir
fath
ers
rath
erth
anw
ith
thei
rm
othe
rs.
•G
irls
•Sp
orts
•Be
com
ing
am
an•
Bird
san
dbe
esN
ame
ate
chni
que
peop
leus
eto
rem
embe
rth
ings
.•
Thi
nkin
gab
outi
t•
Rep
eati
ngit
•W
ord
asso
ciat
ion
•St
ring
onfin
ger
•W
riti
ngit
dow
nN
ame
som
ethi
ngpe
ople
doto
guar
dag
ains
twri
nkle
s•
Boto
x•
Stay
outo
fsu
n•
Iron
clot
hes
•Su
nscr
een
•Fa
cial
crea
m•
Wea
rm
akeu
pN
ame
som
ethi
ngpe
ople
like
todo
inth
eir
back
yard
•G
ames
•Ba
rbec
ue•
Cat
chso
me
rays
•G
arde
n/M
owgr
ass
•Sw
imN
ame
som
ethi
ngpe
ople
dow
hen
ridi
ngin
the
back
ofa
taxi
cab
•Ta
lkto
the
cabb
y•
Look
outt
hew
indo
w•
Talk
once
ll•
Giv
edi
rect
ions
•R
ead
•W
atch
met
er•
Slee
pN
ame
aqu
esti
ona
pers
onw
ho’s
afra
idof
fligh
tmig
htas
ka
fligh
tatt
enda
nt.
•W
illw
ecr
ash?
•Yo
uce
rtifi
ed?
•Is
the
pilo
tgoo
d?•
How
long
isth
efli
ght?
•H
ow’s
the
wea
ther
•Is
the
plan
esa
fe?
We
aske
d10
0m
arri
edw
omen
...If
your
husb
and
aske
dyo
ufo
ra
divo
rce
Sund
ayni
ght,
•C
lean
outt
heba
nkw
hat’s
the
first
thin
gyo
u’d
doM
onda
ym
orni
ng?
•K
ick
him
out
•Le
ave
the
hous
e•
Ask
him
why
•Se
ea
law
yer/
File
•C
hang
eth
elo
cks
Whe
nyo
u’re
look
ing
for
wor
k,na
me
aw
ayyo
um
ight
find
outa
bout
the
job
open
ning
•N
ewsp
aper
•Fr
iend
•In
tern
et•
Une
mpl
oym
ento
ffice
21
Nam
eso
met
hing
you
dow
hile
driv
ing,
even
thou
ghyo
ukn
owit
’sw
rong
•Sp
eed
•Sm
oke
ciga
rett
e•
Not
use
the
seat
belt
s•
Eat/
Dri
nk•
Don
’tus
eth
esi
gnal
s•
Talk
onth
eph
one
Nam
ea
doct
or’s
orde
rpe
ople
neve
rta
kese
riou
sly
•St
opsm
okin
g•
Get
exer
cise
•Pl
enty
ofre
st•
Wei
ght/
Eatr
ight
•M
edic
ine/
vita
min
sW
here
wou
ldyo
ulo
okto
find
outh
owol
dso
meo
neis
?•
Lice
nse/
wal
let
•Bi
rth
cert
ifica
te•
Thei
rha
nds
•Th
eir
eyes
•A
skfa
mily
mem
bers
•A
skth
emN
ame
som
ethi
ngth
atw
ould
besm
artt
okn
owho
wto
ask
for
ina
fore
ign
coun
try
•Th
eem
bass
y•
Food
•D
irec
tion
s•
Bath
room
•H
elp
Nam
eso
met
hing
that
wou
ldbe
smar
tto
know
how
toas
kfo
rin
afo
reig
nco
untr
y•
The
emba
ssy
•Fo
od•
Dir
ecti
ons
•Ba
thro
om•
Hel
pN
ame
som
ethi
ngth
at’s
impo
rtan
tto
pare
nts
whe
nho
use
hunt
ing?
•Lo
cati
on•
Back
yard
•Sa
fety
•Sc
hool
•En
ough
room
sBe
side
sbe
ing
fast
,nam
eso
met
hing
peop
lelik
eab
outf
astf
ood.
•D
on’t
cook
/cle
an•
Tast
e/fla
vor
•V
arie
ty•
Its
hot
•Pr
ice/
Che
ap•
Easy
/C
onve
nien
tIf
you
wer
ear
rest
edan
dco
uld
mak
eon
lyon
eca
ll,w
hom
wou
ldyo
uca
ll?•
bail
bond
sman
•Si
blin
g•
frie
nd•
part
ner/
spou
se•
pare
nt•
law
yer
22
Ifyo
urce
llph
one
ring
sw
hile
you’
reat
chur
ch,w
ho’d
bett
erbe
calli
ng•
My
sibl
ing
•M
yki
ds•
The
lord
•T
hedo
ctor
•M
ysp
ouse
•M
ybo
ss•
My
mot
her
Tell
me
the
mos
tim
port
anth
abit
am
omsh
ould
teac
hhe
rso
n•
Res
pect
for
wom
en•
Pick
upaf
ter
self
•To
ilete
tiqu
ette
•H
ones
ty•
Man
ner
•C
lean
lines
sN
ame
som
ethi
ngth
atpe
ople
doto
thei
rfa
ces
•H
ave
faci
al•
Puto
nm
akeu
p•
Shav
e•
Get
boto
x•
Was
hN
ame
are
ason
ayo
ung
man
mig
htjo
inth
ear
my
•D
raft
ed•
Car
eer
•Ed
ucat
ion
•Se
rve
his
coun
try
•Be
allh
eca
nbe
Nam
ea
reas
onpe
ople
chan
geth
eir
nam
es•
Div
orce
s•
Dis
like
thei
rre
alna
me
•G
etm
arri
ed•
Hid
eN
ame
som
ethi
ngyo
u’d
like
todo
alld
ayw
itho
utfe
elin
gan
ygu
ilt•
Shop
ping
•W
atch
ing
TV
/M
ovie
s•
Slee
ping
/Rel
axin
g•
Rea
ding
•Ea
ting
Nam
ea
way
you
can
tell
that
your
child
isly
ing.
•Lo
okin
the
eye
•St
utte
ring
•Fi
dget
•Tu
rnre
d•
Expr
essi
on•
No
eye
cont
act
Nam
ea
reas
onw
hype
ople
som
etim
esne
edto
take
ada
yof
ffro
mth
ew
ork
•St
ress
relie
f•
Tire
d•
Han
gove
r•
Sick
23
Nam
ea
reas
onyo
um
ight
mak
eyo
ursp
ouse
geto
utof
bed
atth
em
iddl
eof
the
nigh
t•
Get
agl
ass
ofw
ater
•C
heck
the
nois
e•
Snor
ing
•Fe
elsi
ck•
Att
end
toth
eba
byN
ame
som
ethi
ngyo
urbo
dyte
llsyo
uit
’sti
me
todo
•Ea
t•
Slee
p•
Wak
eup
•G
oto
bath
room
Nam
eso
met
hing
that
your
boss
coul
dte
llyo
uth
atw
ould
com
eas
abi
gsh
ock.
•Pr
omot
ion/
rais
e•
Buis
ness
clos
ing
•Bo
ssis
gay
•Yo
udo
bad
wor
k•
You’
refir
edIf
tim
efli
esw
hen
you’
reha
ving
fun,
nam
ea
plac
eit
craw
ls•
Jail
•In
the
traf
fic•
DM
V•
Wor
k•
Chu
rch
•Sc
hool
Wha
t’sth
eha
rdes
tpar
tofy
our
budg
etto
cuto
ut?
•Sh
oppi
ng/
clot
hes
•En
tert
ainm
ent
•R
ent/
Mor
tgag
e•
Food
/Eat
ing
out
Whi
chfo
oddo
you
thin
ksh
ould
bech
osen
asth
e”N
atio
nalF
ood”
ofA
mer
ica?
•Fr
ies
•Pi
zza
•Fr
ied
chic
ken
•H
otdo
g•
Ice
crea
m•
Turk
ey•
Burg
erN
ame
som
ethi
ngsp
ecifi
cev
eryo
neco
mpl
ains
abou
t•
Taxe
s•
Traf
fic•
Wea
ther
•M
oney
•W
ork
Nam
eso
met
hing
that
’sem
barr
assi
ngto
fall
asle
epw
hile
doin
g•
Mak
ing
love
•Ea
ting
•D
rink
ing
•Ta
lkin
g/ph
one
•D
rivi
ngN
ame
som
ethi
ngsp
ecifi
cyo
um
ight
rem
embe
rab
outo
neof
your
teac
hers
from
high
scho
ol.•
Nic
e/go
odte
ache
r•
Goo
dlo
okin
g•
Way
they
dres
sed
•T
heir
glas
ses
•M
ean/
yelle
d
24
Nam
eso
met
hing
mos
tpeo
ple
clea
nat
leas
tonc
ea
day.
•H
ands
•Ba
thro
om•
Dis
hes/
Sink
•Fa
ce•
Teet
h•
Body
Nam
eso
met
hing
one
fam
ilym
embe
rm
ight
stea
lfro
man
othe
r.•
Food
•M
oney
•Lo
ver
•C
loth
es/S
hoes
•Je
wle
ryN
ame
som
ethi
ngsp
ecifi
cev
eryo
neco
mpl
ains
abou
tNam
ea
reas
onyo
um
ight
ask
the
man
ager
ofa
hote
lfor
anot
her
room
.•
Bigg
erbe
d/ro
om•
The
rmos
tatb
roke
n•
No
TV
/Cab
le•
Bed/
room
dirt
y•
Noi
sy•
Nee
dno
nsm
okin
gN
ame
ati
ppe
ople
give
you
onho
wto
getr
ich.
•W
ork
hard
•Sa
vem
oney
•G
ambl
e/lo
tter
y•
Inve
stN
ame
are
ason
why
som
eone
mig
htgo
aw
hole
day
wit
hout
eati
ng.
•Fa
stin
g/di
etin
g•
Forg
ot/b
usy
•D
epre
ssed
/str
ess
•Si
ckN
ame
are
ason
you
mig
htpu
llyo
uca
rov
erto
the
side
ofth
ero
ad.
•Fl
atti
re•
Phon
e•
Polic
e•
Out
ofga
s•
Am
bula
nce/
Fire
Nam
eso
met
hing
you
hope
you
neve
rfin
din
your
hous
e.•
Snak
e•
Inse
ct•
Dea
dbo
dy•
Rod
ent
•Bu
rgla
r•
Spou
sech
eati
ngTh
epo
litic
ian
who
win
sth
eel
ecti
onis
usua
llyth
epe
rson
wit
hth
em
ostw
hat?
•Vo
te•
Mon
ey•
Cha
rism
aTe
llm
ea
reas
onw
hya
man
wou
ldw
antt
oha
vea
son.
•To
play
spor
ts•
Tobe
like
him
•To
carr
yon
nam
e•
Tohe
lpw
hen
old
•To
love
25
Nam
eso
met
hing
that
you’
vest
ayed
upal
lnig
htw
orry
ing
abou
t.•
Kid
s/fa
mily
•Sc
hool
/Tes
t•
Hea
lth
issu
es•
Mon
ey/
Bills
•Jo
b/in
terv
iew
An
engi
neer
who
lost
her
fath
erin
child
hood
ina
shoo
ting
acci
dent
and
blam
esgu
nmak
ers
for
her
loss
.•
Acc
ept,
look
for
othe
rjo
bsH
owev
er,t
heon
lyjo
bof
fer
that
she
has
rece
ived
isfr
oma
gunm
aker
com
pany
.•
Dec
line,
ask
her
mot
her
tose
llth
eho
use
topa
yof
fthe
loan
Ifsh
ede
clin
esth
ejo
bof
fer
(ref
usin
gto
mak
egu
nsth
atca
nki
llot
her
kids
’dad
s),
•D
elin
e,lo
okfo
ran
othe
rjo
b,pa
ylo
anw
ith
crei
tcar
dhe
rre
tire
dm
othe
rm
usts
ellh
erho
use
topa
yof
fhe
rsc
hool
loan
.She
shou
ld•
Dec
line,
look
for
anot
hejo
b,do
notp
ayof
fthe
loan
AU
S-ba
sed
sola
rpa
nelm
anuf
actu
ring
com
pany
can
nolo
nger
com
pete
•Lo
bby
and
add
tarr
ifs
toim
port
edso
lar
pane
lsw
ith
fully
-aut
omat
edco
mpa
nies
inA
sia
that
man
ufac
ture
sola
rpa
nels
50%
chea
per.
•sp
read
fals
ein
fore
gard
ing
child
abus
ela
bor
inA
sia
Hun
dred
sof
wor
kers
will
lose
thei
rjo
bsin
the
US
ifth
eco
mpa
nyis
clos
ed.T
heco
mpa
nysh
ould
•H
ire
low
wag
ein
tern
atio
nalw
orke
rs•
Becl
osed
and
wor
kers
mov
eto
bett
erjo
bsIn
asm
artp
hone
,the
mai
nis
sue
for
you
is•
Spee
d•
Mem
ory
size
•D
ispl
ayqu
alit
yan
dsi
ze•
Batt
ery’
slif
espa
nTh
roug
ha
smal
linc
reas
ein
sale
s/in
com
eta
x,a
univ
ersa
lhea
lthc
are
insu
ranc
epl
an•
No
Uni
vers
alhe
alth
care
isn’
tgoo
dfo
rhi
ghqu
alit
ym
edic
alse
rvic
esca
nbe
crea
ted
inth
eU
S.D
oyo
ulik
eth
isid
ea?
•Ye
s,th
isis
agr
eati
dea!
•N
oun
ives
alhe
alth
care
isgo
od,b
utth
eta
xes
are
alre
ady
high
•N
o,U
nive
salh
ealt
hcar
eis
notg
ood
relig
ous
peop
lesh
ould
n’tp
ayto
fund
abor
tion
med
ical
bills
for
adr
unk
driv
erYo
ure
ally
like
alu
xury
car
that
cost
s$7
0,00
0,bu
tyou
can
only
affo
rd$4
0,00
0.•
you
leas
eca
rsun
till
you
can
buy
the
luxu
ryca
r•
You
buy
the
best
new
car
that
you
can
affo
rdfo
r$4
0,00
0•
You
just
buy
the
car
you
need
,for
less
than
$40,
000
•Yo
ubu
ya
used
mod
elof
the
luxu
ryca
rfo
r$4
0,00
0.W
hen
you
pick
apa
ssw
ord
for
your
phon
e,yo
upi
ckon
eth
aris
•Ea
syto
rem
embe
ran
dke
epit
unch
ange
d•
Diffi
cult
tore
mem
ber
and
keep
itun
chan
ged
•D
ifficu
ltto
rem
embe
ran
dch
ange
itre
gula
rly
•Ea
syto
rem
embe
ran
dyo
uch
ange
itre
gula
rly
You
shar
eva
lues
wit
h•
Dem
ocra
tic
part
y•
Rep
ublic
anpa
rty
•In
depe
nden
t/G
reen
part
y•
Non
eIf
mon
eyis
nota
nis
sue,
you
wou
ldlik
eto
buy
•Sa
msu
ngsm
artp
hone
s•
App
leip
hone
•H
TC
smar
tpho
nes
•LG
smar
tpho
nes
Ifno
tin
ahu
rry,
whi
chro
utyo
uch
oose
?•
wit
hle
astt
raffi
c•
Shor
test
/fas
test
•m
osts
ceni
c•
best
/saf
est
26
Adr
iver
ista
ilgat
ing
you
for
over
anho
ur,w
hatw
ould
you
do?
•ig
nore
itan
dre
mai
npa
tien
t•
Spee
dup
and
leav
eit
behi
nd•
slow
dow
nan
dan
noy
the
driv
er•
Pull
over
and
leti
tpas
sW
hydi
dH
illar
yC
linto
nlo
seth
epr
esid
enti
alel
ecti
on?
•R
ussi
ancy
bera
ttac
ksan
dfa
kene
ws
•U
Sel
ecto
rals
yste
m•
Vote
rsdi
dn’t
like
her
•FB
Ian
noun
cem
entr
egar
din
new
emai
lson
ala
ptop
Wha
tsho
uld
the
US
doif
we
lear
nth
atth
efa
lse
nucl
ear
alar
min
Haw
aiiw
asre
al•
Do
noth
ing
asno
harm
was
caus
edbu
tthe
Nor
thK
orea
nm
issi
leex
plod
edbe
fore
reac
hing
Haw
aii:
•D
etro
yN
orth
Kor
eaby
anu
clea
rat
tack
•A
ssas
sina
teN
orth
Kor
ean
lead
er•
Add
mor
esa
ncti
ons
Wha
tare
your
gend
er/c
olor
prio
riti
esfo
rth
eU
Sco
mm
ande
rin
chie
f?•
Col
oran
gend
erdo
notm
atte
r•
Col
orm
atte
rsbu
tgen
der
does
not
•G
ende
rm
atte
rsbu
tcol
ordo
esno
t•
Gen
der
and
colo
rm
atte
rIf
mon
eyis
nota
nis
sue,
you
will
buy
adr
iver
-les
sca
r•
Rig
htaw
ay•
Onl
yw
hen
25%
ofca
rsar
edr
iver
-les
s.•
Onl
yw
hen
50%
ofca
rsar
edr
iver
-les
s•
Onl
yw
hen
you
have
noot
her
opti
onPr
esid
entT
rum
pstr
avel
ban
is•
Gre
atan
dim
prov
esth
esa
fety
ofle
gala
ndpa
trio
tic
US
citi
zens
•N
otgo
od,b
ecau
seit
isdi
scri
min
ator
y.•
Not
good
,sin
ceno
tall
Mus
limpe
ople
inth
eba
nned
coun
trie
ssh
ould
bepu
nish
edbe
caus
eof
the
Sept
.11
atta
ck•
Goo
d,bu
tsho
uld
incl
ude
mor
eco
untr
ies
Isth
isa
big
deal
ifth
efir
edm
issi
lefr
omYe
men
toSa
udiA
rabi
a•
No,
mis
sile
sco
uld
have
been
sent
befo
reth
esa
ncti
ons
was
infa
ctm
ade
inIr
an?
•N
o,Sa
udiA
rabi
aus
esU
Sbu
iltm
issi
les
•Ye
s,it
show
svi
olat
ion
insa
ncti
ons
•Ye
s,it
show
sIr
anhe
lped
Yem
eto
atta
ckSa
udiA
rabi
aIf
asc
ient
ific
theo
rydo
esno
tagr
eew
ith
your
relig
ious
belie
f•
You
deny
the
scie
ntifi
cth
eory
,as
scie
nce
isno
tflaw
less
.•
You
dono
tpra
ctic
ean
yre
ligio
n.•
You
conv
ertt
oa
new
relig
ion
that
agre
esw
ith
that
theo
ry.
•Yo
upr
acti
ceyo
urre
ligio
nbu
tliv
ew
ith
the
diff
eren
ces.
Ast
udy
atth
eU
nive
rsit
yof
Was
hing
ton
show
edth
atby
incr
easi
ngth
em
inim
umw
age
•T
hest
udy
isfla
wed
,une
mpl
oym
entr
ate
will
incr
ease
,the
refo
re•
The
min
imum
wag
esh
ould
befu
rthe
rin
crea
sed
tore
duce
unem
ploy
emen
t•
The
min
imum
wag
esh
ould
bere
duce
d•
The
resh
ould
bein
com
eta
xto
supp
ortu
nem
ploy
edpe
ople
.M
any
jobs
will
belo
stin
the
next
few
deca
des
due
tosm
artr
obot
s•
No
idea
!an
dun
empl
oym
entw
illth
usin
crea
se.T
here
fore
,•
Cap
ital
ism
will
colla
pse,
sinc
eth
em
iddl
ecl
ass
will
vani
sh•
Smar
trob
ots
shou
ldbe
bann
edto
save
jobs
.•
The
gove
rnm
ents
houl
dpr
ovid
ea
“liv
ing
wag
e”in
com
efo
ral
lune
mpl
oyed
wor
kers
.
27
As
mic
rosi
zed
robo
tsw
illbe
deve
lope
dso
on,p
eopl
eca
nbe
film
edan
ytim
e•
Peop
lew
illle
arn
topr
otec
tthe
irpr
ivac
y.an
dan
ywhe
rew
itho
utth
eir
know
ledg
e,w
hich
will
dest
roy
thei
rpr
ivac
y•
Peop
lew
illst
opm
icro
-siz
edro
bots
.•
The
mea
ning
ofpr
ivac
yw
illch
ange
•Pe
ople
will
tole
rate
the
loss
ofpr
ivac
y.Tr
uede
moc
racy
can
beim
plem
ente
dus
ing
smar
tpho
nevo
ting
,•
Ther
ew
illbe
nore
aldi
ffer
ence
.Dem
ocra
cyis
nott
hean
swer
.as
peop
leca
nvo
teev
ery
day
onim
port
anti
ssue
s,•
You
pref
era
com
bina
tion
ofsm
artp
hone
voti
ngan
dtr
adit
iona
lrep
rese
ntat
ives
and
ther
eis
none
edto
have
aco
stly
cong
ress
orse
nate
anym
ore.
•Yo
upr
efer
trad
itio
nalr
epre
sent
ativ
esth
atyo
uca
ntr
usta
ndvo
tes
The
peop
lew
illfin
ally
bein
char
ge!
•Th
isis
agr
eati
dea.
Why
shou
ldw
epa
yto
keep
abr
oken
syst
em?