Automatic Personality Perception: Prediction of Trait Attribution Based on Prosodic Features Extended Abstract

Transcription

1 215 International Conference on Affective Computing and Intelligent Interaction (ACII) Automatic Personality Perception: Prediction of Trait Attribution Based on Prosodic Features Extended Abstract Gelareh Mohammadi Alessandro Vinciarelli Department of Neuroscience University of Geneva Switzerland School of Computing Science and Institute of Neuroscience and Psychology University of Glasgow,UK Abstract This paper proposes a prosody based approach for Automatic Personality Perception. Social psychology has shown that whenever we listen to a voice for the first time, we spontaneously and unconsciously attribute personality traits to the speaker. The attribution process is not necessarily accurate, but it is important because it shapes our behavior towards others. The experiments of this work are performed over a corpus of 64 speech samples (322 individuals in total) assessed in terms of speaker s personality traits by 11 judges. The results show that it is possible to predict some of the personality traits with accuracy higher than 7%. The effect of different prosodic features has also been analyzed and compared with findings in the psychological literature. Keywords Personality traits; prosody; Big Five; social signal processing; automatic personality perception I. I NTRODUCTION One of the characteristics that make humans social is that they make spontaneous - i.e., effortless, habitual and unintentional [1]) - inferences about a wide range of socially relevant characteristics of others. The process takes place outside conscious awareness and leads to the development of impressions, especially in the earliest stages of an interaction [2]. The phenomenon is pervasive and it takes place not only in face-to-face encounters, but also when people see others in pictures [3] and videos [4] or listen to them in speech recordings [5]. This paper is the summary of a previous publication [6] and studies one aspect of this phenomenon, namely the spontaneous attribution of personality traits to unacquainted speakers that people hear for the first time. In particular, the paper proposes an Automatic Personality Perception (APP) approach based on prosody, i.e. an approach that maps prosodic features into personality traits attributed by listeners. The reason behind the choice of prosody is that there is a large body of literature showing that prosodic features affect the perception of personality traits [7]. Furthermore, nonverbal behavioural cues (e.g., prosody, facial expressions, gestures, postures, etc.) have been shown to be reliable markers of social and psychological phenomena in domains like Affective Computing [8] and Social Signal Processing [9]. At the moment the /15/$ IEEE original work was written [6], only a few papers on APP were available in the literature (see [1] for an extensive survey). The main contributions of the original work were as follows: It was the first work to measure quantitatively the effect of individual prosodic features on APP effectiveness; It was the first work using personality assessments as features for predicting other, independent personal characteristics; The number of subjects used in the original work was three times larger than in any previous APP experiment presented in the literature; The original work considered nonverbal speech features previously neglected in both computing and psychological literature. On the long term, APP can be considered as a contribution to the efforts being done towards bridging the social intelligence gap between people and machines [9]. However, some early applications of APP have already been explored like the generation of synthetic voices eliciting desired social perceptions (see, e.g., [11]), the use of personality assessments in recommender systems [12], the interaction between humans and robots [13], or the indexing of multimedia data in terms of social and emotional user perceptions [14]. In this respect, the development of APP technologies is expected to be beneficial for several computing areas. II. P ERSONALITY AND ITS A SSESSMENT Personality is the latent construct that accounts for individuals characteristic patterns of thought, emotion, and behavior together with the psychological mechanisms - hidden or not - behind those patterns [15]. The literature proposes several personality models (see [1] for a short survey) and the most commonly applied ones are based on traits, i.e., on broad dimensions that capture the most salient characteristics of an individual [16]. According to these models, personalities can be represented as vectors where each component measures how well the behavior of a person actually matches the tendencies associated to a given trait. Such a representation is particularly suitable for computer processing. 484

2 TABLE I: THE BFI-1 QUESTIONNAIRE USED IN THE EXPER- IMENTS OF THIS WORK. ID Question 1 This person is reserved 2 This person is generally trusting 3 This person tends to be lazy 4 This person is relaxed, handles stress well 5 This person has few artistic interests 6 This person is outgoing, sociable 7 This person tends to find fault with others 8 This person does a thorough job 9 This person gets nervous easily 1 This person has an active imagination In particular, this work adopts the Big Five (BF) Traits, five broad dimensions that parsimoniously and comprehensively describe most phenotypic individual differences [17]. The BFs have been identified by applying factor analysis to the 18, words that describe personality in English [18]. While being so many, these personality descriptors tend to group into five major clusters corresponding to the BFs: Extraversion: Active, Assertive, Energetic, etc. Agreeableness: Appreciative, Kind, Generous, etc. Conscientiousness: Efficient, Organized, Planful, etc. Neuroticism: Anxious, Self-pitying, Tense, etc. Openness: Artistic, Curious, Imaginative, etc. In the BF framework, assessing a personality means to find five scores that account for how well the adjectives associated to each trait describe the behavior of an individual. This is typically done using psychometric questionnaires where items describing observable behavior are associated to Likert scales (see Table I). This work has used the Big Five Inventory 1 (BFI-1) of Table I, a short version of the full BFI [19]. The 1 items of the BFI-1 are those that better correlate with the results of the full BFI questionnaire. Each item is associated to a 5 points Likert scale (from Strongly Disagree to Strongly Agree ). The main advantage of BFI-1 is that it takes less than one minute to be completed. III. THE APPROACH The APP approach proposed in this work includes three main steps: extraction of low-level prosodic features from the speech signal, estimation of statistical measures accounting for long-term prosodic properties of speech, and finally mapping of statistical features into attributed traits. A. Low-level Feature Extraction The low-level features adopted in this work account for the most important aspects of prosody (pitch, energy and tempo). For this reason, they have been extensively investigated not only in speech processing, but also in psychological fields interested in the interplay between speech and personality (see [1]). In particular, the low-level features adopted in the experiments are pitch (number of vibrations per second produced by the vocal cords, the main acoustic correlate of tone and intonation), first two formants (resonant frequencies of the vocal tract), energy of the speech signal (how loud or soft people speak), and length of voiced and unvoiced segments (an indirect measure of speaking rate). The features are extracted from 4 ms long analysis windows that span the entire duration of each speech sample and start at regular time steps of 1 ms. In this way, two consecutive windows overlap for 3 ms. This process converts a speech sample into a sequence of six-dimensional feature vectors f i (one vector per window) where each component f (j) i (j = 1,..., 6) corresponds to one of the low-level features above. After having been extracted, the features are normalized to standard scores using mean and variance as estimated over the training set. B. Estimation of Statistical Features The low-level features account for the short-term aspects of speech. Hence, the representation of long-term properties requires the use of four statisticals, namely minimum, maximum, mean and relative entropy of the differences between low-level feature values extracted from consecutive analysis windows. The reason for using minimum and maximum is that together they measure the dynamic range of a given feature. The mean provides an indication of what are the most frequent values and, finally, the entropy represents the predictability of a given prosodic characteristic. In particular, the higher the entropy, the more it is difficult to predict the next value of a low-level feature given the current one (the approach is inspired by [2]). The estimate of the entropy is obtained as follows: first all differences f (j) i = f (j) i f (j) i 1 between consecutive values of low-level feature j are collected. This allows one to build the set Y (j) = {y (j) 1, y(j) 2,..., y(j) } of all possible difference Y (j) that can be observed in a speech sample. At this values f (j) i point, it is possible to estimate the entropy: H( f (j) i ) = Y (j) k=1 p(y(j) k ) log p(y(j) k ), (1) log( Y (j) ) where p(y (j) k with the fraction of times the value y (j) ) is the probability of f (j) k i = y (j) k (estimated is actually observed) and Y (j) is the cardinality of Y (j) (number of elements in Y (j) ). The term log Y (j) works as a normalization factor; The upper bound (H = 1) is reached when the distribution is uniform (maximum uncertainty). When the entropy is higher, it means there is higher uncertainty and the feature is less predictable. Given that the low-level features are extracted every 1 ms from a 4 ms window, a record of 1 seconds (the length of the samples used in the experiments) results into 996 feature vectors. This is sufficient to obtain reliable statistics and avoid the effect of possible outliers. Since the low-level features are six and the statisticals are four, the total number of statistical features is 24. The vector of the statisticals is used to represent each sample of the dataset. C. Recognition The recognition step predicts the personality traits that people attribute to the speakers of the dataset using the /15/$ IEEE 485

3 statisticals extracted at the previous step. In particular, this step predicts whether a speaker has been perceived to be below the median or not with respect to each of the Big Five Traits. The task is performed using a logistic regression: p(c x) = exp( D i=1 θ ix i θ ) 1 + exp( D i=1 θ (2) ix i θ ) where D is the dimension of the feature vectors and the θ i are the parameters of the model. If p(c x).5 the feature vector x is assigned to class C. The model is trained by maximizing the entropy over the training set with the Limited Memory Broyden-Fletcher-Goldfarb-Shanno (BFGS) method [21] 1. In this work, the main advantage of the logistic regression is that the parameters θ i provide a measure of how individual features influence the outcome of the classification. This can be considered an indirect indication of the features driving the perception of the judges that have assessed the speech samples (see Section IV-A). IV. EXPERIMENTS AND RESULTS This section reports on dataset, experiments and results obtained in this work. A. The Data The experiments of this work are based on the publicly available SSPNet Speaker Personality Corpus 2, a collection of 64 speech clips involving 322 individuals. The samples have been extracted randomly from a set of 96 news bulletins broadcast by Radio Suisse Romande, the French speaking Swiss national broadcast service, during February 25. The clips are emotionally neutral and they do not contain words that might be easily understood by non-french speakers (e.g., names of places or well known people). This is to restrict the personality assessment, performed by non-french speakers, to be influenced mainly by nonverbal vocal behavior. Overall, the speakers can be grouped into professional (37 samples), the journalists that work for Radio Suisse Romande, or nonprofessional (333 samples), people that happen to talk on the radio because of a specific news. The personality assessments have been done by 11 judges who did not speak or understand French. The assessors listened to all clips and then completed the BFI-1 questionnaire over an online platform. Once the assessment of a clip was completed, there was no possibility to go back or to edit the answers. The final personality score for each clip is the average of the scores assigned by each of the judges individually. B. Perceived Personality as a Predictor The personality assessments of the 64 clips can be thought of as vectors in a five-dimensional space. Figure 1 shows the projection of these vectors onto the first two Principal Components [22] (roughly 7 percent of the total variance). The symbol is different for professional (37 cases) and nonprofessional speakers (333 cases). While overlapping, the two 1 weinman/code/index.shtml 2 sspnet.eu/213/1/sspnet-speaker-personality-corpus/ PC Personality Patterns 4 professional non professional PC1 Fig. 1: Personality patterns.the coordinates of each point are the projections of a personality assessment over the first two Principal Components (P C1 and P C2) Professional vs Non Professional Ext. Agr. Con. Neu. Ope. Trait Fig. 2: Categorization. The upper plot shows the weights of the traits used as features to distinguish between professional and nonprofessional speakers. The lower plot shows how the accuracy changes when using an increasing number of traits ordered by their weight (the error bars account for the 95% confidence interval). categories appear to cover different parts of the personality space. This seems to confirm that the assessments are not random, but capture meaningful differences between speakers. As a further confirmation, a logistic regression was trained to distinguish between professional and non-professional speakers using the personality scores as features. Figure 2 shows the θ coefficients of the model and the lower plot corresponds to the model performance when using only the K traits with highest weights (K = 1,...,5). Conscientiousness and Extraversion are the two traits that influence most the classification outcome. In particular, the performance using the /15/$ IEEE 486

4 TABLE II: LOGISTIC REGRESSION ACCURACY AS A FUNCTION OF THE AGREEMENT BETWEEN ASSESSORS (INCLUDING 95% CONFIDENCE INTERVAL). THE NUMBER IN PARENTHESIS IS THE PERCENTAGE OF THE CORPUS FOR WHICH AT LEAST N JUDGES AGREE ON THE SAME LABEL FOR A GIVEN TRAIT. THE B COLUMNS SHOW THE BASELINE PERFORMANCE, NAMELY THE ACCURACY OF A SYSTEM THAT ALWAYS GIVES AS OUTPUT THE CLASS WITH THE HIGHEST A-PRIORI PROBABILITY. Trait n 6 B n 7 B n 8 B n 9 B Extraversion 71.4 ± 3.5 (1.) 5..5 ± 3.8 (77.6) ± 4.1 (57.3) ± 4.6 (36.1) 53. Agreeableness 58.8 ± 3.8 (1.) ± 4.6 (67.2) ± 5.7 (4.9). 63. ± 8.7 (18.6) 59. Conscientiousness 72.5 ± 3.5 (1.). 79. ± 3.8 (67.) ± 4.8 (38.1) ± 7. (14.5) 62. Neuroticism 66.1 ± 3.7 (1.) ± 4.3 (68.4) ± 5.5 (38.9) ± 7.4 (2.8) 52. Openness 58.6 ± 3.8 (1.) ± 4.9 (.1). 7.6 ± 7.4 (22.8) ± 12.4 (6.2) 67. TABLE III: CONFUSION MATRICES: THE ROWS CORRESPOND TO THE ACTUAL LABEL OF A THE SAMPLES, WHILE THE COLUMNS CORRESPOND TO THE LABEL ASSIGNED BY THE APPROACH. THE ELEMENT (I, J) OF EACH MATRIX IS THE PERCENTAGE OF SAMPLES BELONGING TO CLASS I THAT HAVE BEEN ASSIGNED TO CLASS J. Logistic Regression Ext. Agr. Con. Neu. Ope. Class Low High Low High Low High Low High Low High Low High Conscientiousness score alone is around 7. Adding the other traits does not lead to statistically significant improvements. This is not surprising given that the characteristics of the trait ( organized, knowledgeable, thorough, reliable, etc. [17]) are probably those that journalists tend to convey when they talk on the radio. C. Prosody Based Personality Perception The 11 judges fill the BFI-1 for each clip. Thus, for each of the judges, it is possible to estimate the median along every trait. This makes it possible to verify whether a given sample is below median or not along a given trait for all the judges. When the majority of the judges perceives the speaker to be below median, the sample is assigned to class Low, otherwise the sample is assigned to class High. As the judges are 11, there is always a majority of at least n = 6 judges. It is possible to restrict the experiments to samples for which the size n of the majority group is larger. However, the larger n the smaller the fraction of data that can be used. Table II reports the accuracies (percentage of times the logistic regression assigns a sample to the right class). The percentage in parenthesis is the fraction of data for which at least n judges agree. The results are compared with a baseline that corresponds to the performance obtained when predicting always the class with the highest a-priori probability. The improvement with respect to such a baseline is always statistically significant except in the case of Openness. The first column of the table corresponds to n 6, and the experiments are performed over the entire corpus. Extraversion and Conscientiousness tend to be recognized better than the other traits. This is not surprising for Extraversion because it has been shown that people perceive easily this trait [23]. The high accuracy for Conscientiousness is peculiar of this work and is consistent with the results of Section IV-B showing that the trait discriminates between the two categories of speakers represented in the corpus. Table III reports the confusion matrices for all traits; the recognition rates are roughly similar for both classes except for Openness where the accuracy is not significantly above the performance of the baseline. The results show that the accuracy increases when there is more consensus across judges, meaning that there is a more coherent and consistent relationship between measurable speech characteristics and attributed traits. This suggests that variability across judges is probably one of the main difficulties in APP. Variability is an inherent aspect of personality perception because it is not possible to say whether a judge is right or wrong. The judges are expected to rate the subjects according to their impressions and not according to a supposed groundtruth. However, it is still possible to find the features that influence most the outcome of the classification. One of the main reasons to use the Logistic Regression is that the parameter vector θ provides indications about the features impact on the classification results. Indirectly, this suggests what are the features that possibly drive the perception of the judges. Figure 3 shows the θ coefficients of the model for each trait separately and how the accuracy changes by using only the first K highest ranking features where K = 1,2,...,2. Four features have been eliminated because they are highly correlated with one of the retained features and this would make the value of the θ i coefficients unreliable. For Extraversion, the feature that influences most the outcome of the classification is the pitch entropy. This is in line with previous findings of the psychological literature showing that people with higher pitch variability tend to be perceived as more extravert [24]. The means of the first two formants play an important role as well. This might be due to possible gender effects (female subjects tend to have higher formants) or the /15/$ IEEE 487

5 Extraversion Agreeableness Conscientiousness Neuroticism Openness 1 F Measure F Measure Ext. Agr. Con. Neu. Ope. Traits Fig. 3: For each trait, the upper chart shows the θ coefficients associated to the different features. For each cue (e.g., pitch), there are four statistics, namely mean, minimum, maximum and entropy. The lower plot of each trait shows the accuracy achieved when using only the N top ranking features (in terms of absolute values θ of the coefficients). The error bars correspond to the 95% confidence interval. The last plot shows the F-measures obtained when using all features. All plots correspond to n /15/$ IEEE 488

6 influence of the content (formants account for the phonemes uttered by a speaker). However, this latter effect should not be particularly important because the judges do not understand the language of the speech samples. The last important feature, at least according to the θ coefficient value, is the mean length of unvoiced segments. This suggests that people making long pauses and/or speaking slower tend to be perceived as less extravert, in line with previous findings in psychology [25]. In the case of Conscientiousness, entropy plays a major role when it comes to pitch, first formant and energy, thus confirming previous findings of the psychological literature showing that variability tends to be interpreted as higher Conscientiousness [26]. of the first formant and minimum of the second one seem to suggest that there is a gender effect known as benevolent stereotype [23]: Women tend to be perceived as higher in Extraversion, but lower in Conscientiousness. However, it could be the effect of the words being uttered as well. of unvoiced segments length also appears to be effective negatively, meaning that people using too many pauses are perceived as less competent. The psychological literature does not offer terms of comparison for the other three traits. However, the experiments of this work can suggest some cues that influence their perception. The mean of the formants appears to have an effect in both Agreeableness and Neuroticism. This suggests that speakers with higher formants tend to be perceived as less agreeable and more neurotic. In the case of Openness, the indications are probably unreliable because the classifier does not improve to a statistically significant extent the performance of the baseline. The main reason is probably that this trait is difficult to be perceived in the particular setting of the experiments. V. CONCLUSIONS This work has proposed an approach for Automatic Personality Perception [1], the automatic prediction of the personality traits that people attribute to others. At the moment the original work was written [6], the corpus used in the experiments was the largest of its kind (see Section IV-A). Furthermore this work has proposed one of the first attempts to use pattern recognition techniques to confirm previous findings of the psychological literature while, possibly, introducing new ones. The actual task addressed in the experiments is to predict whether a speaker is perceived to be below median or not with respect to each of the Big-Five Traits. The performances range between 6% and 72% depending on the particular trait. Extraversion and Conscientiousness tend to be predicted better, in line with the observations of personality psychology [23]. The performance for the latter trait is particularly high if compared with other works published in the computing literature [1]. The probable reason is that the SSPNet Speaker Personality Corpus tends to make the trait available [27], i.e., easy to access for the judges. In fact, the speakers are people talking on the radio and, presumably, trying to appear as Conscientious as possible, especially in the case of professional speakers (see Section IV-A). Since the experiments focus on the attribution of personality traits at zero acquaintance (speakers and judges do not know one another and have never been in contact), the agreement between judges tends to be low [28], [29], but still statistically significant. This has to be considered a normal characteristic of these experiments because attributed traits are not right or wrong, but simply account for the impressions of the judges. This is probably one of the main difficulties in APP and the only way to avoid the problem is to point on scenarios where the judges are more likely to agree with one another. Possible directions for future work include the prediction of personal characteristics based on automatically attributed traits [27], [3], the modeling of dimensional personality representations, or the inclusion of cultural effects. In all cases, the main obstacle is probably the collection of corpora large enough to allow progress. The reason is that the collection of multiple judgments, necessary to obtain reliable personality assessments [31], is expensive and time-consuming. On the long-term, the main reason to develop technologies capable to deal with human personality is that this latter accounts for most individual differences and observable behaviors. In this respect, computers capable to deal with personality traits will be capable to deal with people as well. In other words, technologies like APP can contribute to the development of socially intelligent machines that interact with their users like people interact with others in human-human exchanges. ACKNOWLEDGMENT The research that has led to this work has been supported in part by the European Communitys Seventh Framework Programme (FP7/27-213), under grant agreement no (SSPNet), in part by the Swiss National Science Foundation via the National Centre of Competence in Research IM2 (Information Multimodal Information management) and Indo- Swiss Joint Research Project CCPP (Cross-Cultural Personality Perception). The authors wish to thank Marcello Mortillaro for his help on the psychological aspects of this work and Olivier Bornet for the technical support. When the work was done, Gelareh Mohammadi and Alessandro Vinciarelli were with the Idiap Research Insitute ( REFERENCES [1] J. S. Uleman, L. S. Newman, and G. B. Moskowitz, People as flexible interpreters: Evidence and issues from spontaneous trait inference, in Advances in Experimental Social Psychology, M. P. Zanna, Ed., 1996, vol. 28, pp [2] J. S. Uleman, S. A. Saribay, and C. M. Gonzalez, Spontaneous inferences, implicit impressions, and implicit theories, Annual Reviews of Psychology, vol. 59, pp , 28. [3] C. Olivola and A. Todorov, Elected in 1 milliseconds: Appearancebased trait inferences and voting, Journal of Nonverbal Behavior, vol. 34, no. 2, pp , 21. [4] B. Reeves and C. Nass, The media equation: How people treat computers, television, and new media like real people and places. New York (USA): Cambridge University Press, [5] C. Nass and S. Brave, Wired for speech: How voice activates and advances the Human-Computer relationship. The MIT Press, 25. [6] G. Mohammadi and A. Vinciarelli, Automatic personality perception: Prediction of trait attribution based on prosodic features, IEEE Transactions on Affective Computing, vol. 3, no. 3, pp , /15/$ IEEE 489

7 [7] K. R. Scherer, Personality inference from voice quality: The loud voice of extroversion, European Journal of Social Psychology, vol. 8, pp , [8] R. Picard, Affective computing. The MIT Press, 2. [9] A. Vinciarelli, M. Pantic, and H. Bourlard, Social Signal Processing: Survey of an emerging domain, Image and Vision Computing Journal, vol. 27, no. 12, pp , 29. [1] A. Vinciarelli and G. Mohammadi, A survey of personality computing. 214, vol. 5, no. 3, pp [11] C. Nass and K. Min Lee, Does computer-synthesized speech manifest personality? experimental tests of recognition, similarity-attraction and consistency-attraction, Journal of Experimental Psychology: Applied, vol. 7, no. 3, pp , 21. [12] M. Tkalcic, T. Tasic, and J. Kosir, Emotive and personality parameters in multimedia recommender systems, in Proceedings of the IEEE International Conference on Affective Computing and Intelligent Interaction, 29. [13] A. Tapus and M. Mataric, Socially assistive robots: The link between personality, empathy, physiological signals, and task performance, in Proceedings of AAAI Spring Symposium, 28. [14] M. Pantic and A. Vinciarelli, Implicit human-centered tagging, IEEE Signal Processing Magazine, vol. 26, no. 6, pp , 29. [15] D. Funder, Personality, Annual Reviews of Psychology, vol. 52, pp , 21. [16] G. Matthews, I. Deary, and M. Whiteman, Personality Traits. Cambridge University Press, 29. [17] G. Saucier and L. Goldberg, The language of personality: Lexical perspectives on the five-factor model, in The Five-Factor Model of Personality, J. Wiggins, Ed., [18] G. Matthews, I. Deary, and M. Whiteman, Personality traits. Cambridge University Press, 23. [19] B. Rammstedt and O. John, Measuring personality in one minute or less: A 1-item short version of the big five inventory in english and german, Journal of Research in Personality, vol. 41, no. 1, pp , 27. [2] C. Song, Z. Qu, N. Blumm, and A. Barabási, Limits of predictability in human mobility, Science, vol. 327, no. 5968, pp , 21. [21] D. Liu and J. Nocedal, On the limited memory BFGS method for large scale optimization, Mathematical programming, vol. 45, no. 1, pp , [22] C. Bishop, Pattern recognition and machine learning. Springer New York, 26. [23] C. Judd, L. James-Hawkins, V. Yzerbyt, and Y. Kashima, Fundamental dimensions of social judgment: Unrdestanding the relations bteween judgments of competence and warmth, Journal of Personality and Social Psychology, vol. 89, no. 6, pp , 25. [24] G. B. Ray, Vocally cued personality prototypes: An implicit personality theory approach, Journal of Communication Monographs, vol. 53, no. 3, pp , [25] A. W. Siegman and B. Pope, Effects of question specificity and anxiety producing messages on verbal fluency in the initial interview, Journal of Personality and Social Psychology, vol. 2, pp , 19. [26] S. Ketrow, Attributes of a telemarketer s voice persuasiveness, Journal of Direct Marketing, vol. 4, no. 3, pp. 8 21, 199. [27] A. Wright, Current directions in personality science and the potential for advances through computing, IEEE Transactions on Affective Computing, vol. 5, no. 3, pp , 214. [28] D. Kenny, L. Albright, T. Malloy, and D. Kashy, Consensus in interpersonal perception: acquaintance and the Big Five. Psychological Bulletin, vol. 116, no. 2, pp , [29] D. Kenny, PERSON: A general model of interpersonal perception, Personality and Social Psychology Review, vol. 8, no. 3, pp. 2 28, 24. [3] D. Ozer and V. Benet-Martinez, Personality and the prediction of consequential outcomes, Annual Reviews of Psychology, vol. 57, pp , 26. [31] R. Rosenthal, Conducting judgment studies: Some methodological issues, in The new handbook of methods in nonverbal behavior research, J. Harrigan, R. Rosenthal, and K. Scherer, Eds. Oxford University Press, 25, pp /15/$ IEEE 49