Common Sense Knowledge Based Personality Recognition from Text

Common Sense Knowledge Based Personality Recognition from Text Soujanya Poria,1 Alexandar Gelbukh,2,3 Basant Agarwal,4 Erik Cambria,1 Newton Howard 5 1

National University of Singapore, Singapore CIC, Instituto Politecnico Nacional, 07738 DF, Mexico 3 Institute for Modern Linguistic Research, “Sholokhov” Moscow State University for Humanities, Moscow, Russia 4 Malaviya National Institute of Technology, Jaipur 302017, India 5 Massachusetts Institute of Technology, USA 2

[email protected], www.gelbukh.com, [email protected], [email protected], [email protected]

Abstract. Past works on personality detection has shown that psycho-linguistic features, frequency based analysis at lexical level, emotive words and other lexical clues such as number of first person or second person words carry major role to identify personality associated with the text. In this work, we propose a new architecture for the same task using common sense knowledge with associated sentiment polarity and affective labels. To extract the common sense knowledge with sentiment polarity scores and affective labels we used Senticnet which is one of the most useful resources for opinion mining and sentiment analysis. In particular, we combined common sense knowledge based features with phycho-linguistic features and frequency based features and later the features were employed in supervised classifiers. We designed five SMO based supervised classifiers for five personality traits. We observe that the use of common sense knowledge with affective and sentiment information enhances the accuracy of the existing frameworks which use only psycho-linguistic features and frequency based analysis at lexical level. KEYWORDS: personality detection, common sense knowledge, affective and sentiment information

1

Introduction

The existence of various personality types and its connection with the different patterns of human behavior has been discussed since the times of Aristotle [21]. The computer era has made it possible to access and analyze large amounts of text samples in order to automatically identify personality types of authors and predict potential reactions and behaviors. The discipline of the Computational Psychology [10] has experienced a tremendous boost having roots in artificial intelligence, traditional psychology and natural language processing.

The rapid development of Web 2.0, the appearance of Social Media tools, and the interest to Social network analysis brought a necessity of modeling personality of the main agent of online interactions. Recent studies show that the connection between personality and user behavior online preserves [9]. Earlier researchers have also found correlations between personality and success at work, in personal relationships and general feeling of happiness in life [16]. The computer-based Personality Recognition (PT) discipline studies the approaches for constructing personality models of Social Web participants, their evaluation and application for the needs of electronic social services. Personality Recognition from Text (a sub-disciplinary of PT) focuses on the analysis of textual samples. Various researchers found correlations between linguistic features and personality characteristics (first-person singular pronouns correlate with depression levels [24], anger and swearing words correlate with [20]). The further development of the discipline is beneficial for many activities that are performed by means of online facilities on a daily basis (customer support, recommendation of services and products, etc.). Recruiters of the HR department analyze hundreds of job applications working hard to map them to the required characteristics the future stuff should have [18]. At the same time the developers of the e-commerce resources are constantly improving the personification algorithms to help the customers obtain products and services that match the needs more precisely and present the information in a more appealing way to increase sales [8]. All these tasks will eventually involve a crucial step of implicit (mental) or explicit (through a user profile) modeling of the user personality. In our present work we extracted common sense knowledge [2] available in text and further using sentic computing [7] we enhance the accuracy of the PRT system. We show that how personality is inferred by common sense knowledge concepts used by a person and sentic computing uses the affective information and sentiment information of these concepts along with psycholinguistic information from LIWC as the features to train the personality classifier. The paper is organized as follows. In Section 2 we give an overview of the related work done in the Personality Recognition from Text field. In Section 3 we provide a description of the proposed algorithm followed by its evaluation in Section 4.We conclude with the discussion of the results and future work.

2

Background

2.1 Personality Estimation The definition of the personality has been among one of the vague and philosophical questions. Modern trait theory is trying to model personality through fixing of a number of classification dimensions (usually following a lexical approach) and construction of the questionnaire to measure them [21]. Researchers use various schemes for personality modeling such as 16PF [12], EPQ-R [15], MBTI [11].Apart from Myers-Briggs classification [22] one of the most

widely exploited schemes for Personality Recognition from Text is the Big Five (BF) [19]. It shows consistency across age and gender, and its validity remains the same when using different tests and languages [16]. The model provides the following five descriptive dimensions abbreviated as OCEAN: • • • • •

Openness to experience (tendency to non-conventional, abstract, symbolic thinking vs preference of non-ambiguous, familiar and non-complex things) Conscientiousness (tendency to hold to long-term plans vs impulsive and spontaneous behavior) Extraversion (active participation in the world around vs concentration on one’s own feelings) Agreeableness (eagerness to cooperate and help vs self-interest) Neuroticism (tendency to experience negative feelings and being overemotional vs emotional stability and calmness)

2.2 Related Work The general procedure for Personality Recognition from Text (PRT) involves collecting the dataset labeled with personality scores gathered through questionnaires, selection of particular linguistic features, construction of the recognition algorithm and its evaluation over a gold standard. Pioneers in PRT showed that individual words usage can reveal personality. Even a small set of extracted features based on bi-grams and tri-grams has a correlation with particular traits [23] Features based on linguistic (LIWC) and psycholinguistic (MRC) categories also reflect personality types [20, 27, 28]. In [24], authors used lexical categories from Linguistic Inquiry and Word Count (LIWC) to identify the impact of linguistic features on personality. Their study revealed that the fewer use of articles and frequent use of positive emotion words actually support agreeability but neuroticism actually supported by the frequent use of negative emotion words and first person pronouns. In [13], authors adapted 22 features exploited by authors in [20] to construct the PRT algorithm over the Italian FriendFeed dataset [13] with 1065 user posts. Their system did not require a dataset to be annotated as they were using previously published correlations between the features and personality traits. For each of those 22 features the mean, standard deviation, minimum, and maximum values were calculated based on a sample of 500 posts. The overall score for a particular feature was calculated in the following way: if the frequency of the feature was higher than the previously estimated mean value plus standard deviation and: 1. 2.

the feature for in a particular sentence correlates positively with the specific trait then the score of the trait was incremented. the feature for in a particular sentence correlates negatively with the specific trait then the score of the trait was decremented.

Then the resulting score was substituted with “y” (if it is positive), “n” (if it is negative) or “o” (if it is zero). Finally, the majority score per each trait was treated as a personality model of the particular user. It is interesting to note that people who produced the largest amount of posts were extravert, insecure, friendly, not very

precise and unimaginative while the average user was extravert, insecure, agreeable, organized and unimaginative. Some researchers answer specific questions about personality while concentrating only at a subset of the Big Five traits. Tomlinson et al. [25] studied the Conscientiousness trait to detect goal, motivation, and the way the author perceives control over the described situations. They performed the analysis of event structures of textual user status updates in a Facebook dataset. The features under consideration were event-based verbs graded by their objectivity and specificity (calculated using WordNet1). Less objectivity and greater specificity are suggested to have connection with more control and stronger goal orientation. Also two thematic roles or relations (agent and patient) were taken into account (annotated using Propbank2 corpus). The accuracy of predicting the score of the Conscientiousness trait being above or below the median (3.5) was 58.13%. Golbeck et al. [16] concentrated on the analysis of Facebook profiles through the processing of 161 statistics of 167 users including personal information (name, birthday, gender, etc.), list of interests (music, movies, etc.), language features of status updates and “About Me” section, and internal statistics (userID, time of the last profile update, etc.). They found that the extracted features had the largest number of correlations with the Conscientiousness trait. The authors concluded that conscious people use fewer words that describe perceptive information (something they see or feel) and tend to discuss other people more. In [26], authors used various emotion lexicons like NRC hash tag emotion lexicon and NRC emotion lexicon for the personality detection and found key improvement in the accuracy of the PRT system. In particular, this work of Mohammad et al. motivated us to use common sense knowledge and sentic computing to infer personality associated with the text.

3

Resources

To detect personality associated with the texts, information related to the language and the properties of individual words of concepts was used. Specifically, we used the following lexical resources. As the aim of this research is to improve the accuracy of the personality framework using emotional features carried by the common sense knowledge exist in the text, we use senticnet as a sentiment polarity dataset, emosenticnet as an emotion lexicon and emosenticspace and conceptnet as common sense knowledge base. In the other hand, LIWC and MRC were used as a psycholinguistic lexicon to extract linguistic features. The SenticNet dataset As an a priori polarity lexicon of concepts, we used the SenticNet 2.0 [1], a lexical resource that contains more than 14,000 concepts along with their polarity scores in the range from –1.0 to +1.0. Among these concepts, 7,600 are multiword concepts. SenticNet 2.0 contains all WordNet Affect concepts as well.

1 2

http://wordnet.princeton.edu/ http://verbs.colorado.edu/~mpalmer/projects/ace.html

The first 20 SenticNet concepts in the lexicographic order along with the corresponding polarities are shown in Table 1. Table 1 A sample of SenticNet data a lot a lot sex a way of Abandon Abase Abash abashed abashment Abhor abhorrence

+0.970 +0.981 +0.303 –0.858 –0.145 –0.130 –0.135 –0.118 –0.376 –0.376

Abhorrent able read able run able use abominably abominate abomination Abortion Abroad Absolute

–0.396 +0.964 +0.960 +0.941 –0.396 –0.376 –0.376 –0.116 +0.960 +0.495

3.1 The ConceptNet The ConceptNet [2] represents the information from the Open Mind corpus as a directed graph, in which the nodes are concepts and the labeled edges are commonsense assertions that interconnect them. For example, given the two concepts person and cook, an assertion between them is CapableOf, i.e. a person is capable of cooking; see Figure 1 [2].

Fig. 1. Labelling facial images in the sequence as neutral or carrying a specific emotion

3.2 The EmoSenticNet The Emosenticnet [3] contains about 5,700 common-sense knowledge concepts, including those concepts that exist in Wordnet Affect list, along with their affective labels in the set {anger, joy, disgust, sadness, surprise, fear}. 3.3 EmoSenticSpace In order to build a suitable knowledge base for emotive reasoning, we applied the blending technique to ConceptNet and EmoSenticNet. Blending is a technique that performs inference over multiple sources of data simultaneously, taking advantage of the overlap between them [4]. Basically, it linearly combines two sparse matrices into a single matrix, in which the information between the two initial sources is shared. Before doing the blending, we represented EmoSenticNet as a directed graph similarly to ConceptNet. For example, the concept birthday party is assigned an emotion joy. We took them as two nodes, and added the assertion HasProperty on the edge directed from the node birthday party to the node joy. Then, we converted the graphs to sparse matrices in order to blend them. After blending the two matrices, we performed the Truncated Singular Value Decomposition (TSVD) on the resulted matrix to discard those components representing relatively small variations in the data. We discarded all of them keeping only 100 components of the blended matrix to obtain a good approximation of the original matrix. The number 100 was selected empirically: the original matrix could be best approximated using 100 components. 3.4 LIWC (Linguistic Inquiry and Word Count) LIWC3 is a text analysis tool that counts and sorts words according to the psychological and linguistic categories defined in the program dictionaries [24]. It processes the text word by word to establish the category of each of them and calculate the overall percentage of words in the discovered categories. Appendix A shows the list of LIWC categories and examples of words [24]. 3.5 MRC (Medical Research Council) MRC 4 database of psycholinguistic categories is an online service (since version 1) [14] and a machine usable dictionary (since version 2) that can be freely utilized for the purposes of natural language processing and artificial intelligence tasks. Appendix B shows the full list of MRC categories and the explanation of each of them [17].

3 4

http://www.liwc.net http://websites.psychology.uwa.edu.au/school/MRCDatabase/uwa_mrc.htm

4

Algorithm

We use essays dataset by [24] which contains 2400 essays labelled manually with personality scores for five different personality traits. Later, the data was tuned by Fabio Celli who manually converted the regression scores into class labels of five different traits. In our case, we extracted several features from the text using LIWC (Appendix A), MRC (Appendix B) and combine them with the common sense knowledge based features extracted by sentic computing techniques. Later, we employed these features into five different classifiers (each for five traits) to build the model for personality prediction. 4.1 LIWC Features Below we show the features extracted from text using LIWC dictionary, along with a definition or examples. Total 81 features were extracted related to the frequency of word count, number of words which have different emotions according to Ekman's model, the number of verbs in the future tense etc. Linguistic processes Word count Words per sentence Dictionary words Words having more than 6 letters First-person singular First-person plural Second person Third person singular Third person plural Indefinite pronouns Articles Common verbs Auxiliary verbs Past tense Present tense Future tense Adverbs Prepositions Conjunctions Negations Quantifiers Numbers Swear words

Percentage of all words captured by the program Percentage of all the words longer than 6 letters I, me, mine We, us, our You, your She, her, him They, their, they’d It, it’s, those A, an, the Walk, went, see Am, will, have Went, ran, had Is, does, hear Will, gonna Very, really, quickly To, with, above And, but, whereas No, not, never Few, many, much Second, thousand Damn, piss, fuck

Psychological processes Social processes Family Friends Humans

Daughter, husband Buddy, friend Adult, baby, boy

Affective processes Positive emotion Negative emotion Anxiety Anger Sadness

Love, nice, sweet Hurt, ugly, nasty Worried, nervous Hate, kill, annoyed Crying, grief, sad

Cognitive processes Insight Causation Discrepancy Tentative Certainty Inhibition Inclusive Exclusive

Think, know Because, effect, hence Should, would, could Maybe, perhaps Always, never Block, constrain And, with, include But, without

Perceptual processes See Hear Feel

View, saw, seen Listen, hearing Feels, touch

Biological processes Body Health Sexual Ingestion

Cheek, hands, spit Clinic, flu, pill Horny, love, incest Dish, eat, pizza

Relativity Motion Space Time

Arrive, car, go Down, in, thin End, until, season

Personal concerns Work Achievement Leisure Home Money Religion Death

Job, majors, Xerox Earn, hero, win Cook, chat, movie Apartment, kitchen Audit, cash, owe Altar, church, mosque Bury, coffin, kill

Spoken categories Assent Nonfluencies Fillers

Agree, OK, yes Er, hm, umm Blah, Imean, yaknow

4.2 MRC Features Below we present the features extracted using MRC. These are mainly linguistic features, such as the number of syllables and phonemes in the word. NLET NPHON NSYL K-F-FREQ K-F-NCATS K-F-NSAMP T-L-FREQ BROWN-FREQ FAM CONC IMAG MEANC MEANP AOA TQ2 WTYPE PDWTYPE ALPHSYL STATUS VAR CAP IRREG WORD PHON DPHON STRESS

Number of letters in the word Number of phonemes in the word Number of syllables in the word Kucera and Francis written frequency Kucera and Francis number of categories Kucera and Francis number of samples Thorndike-Lorge frequency Brown verbal frequency Familiarity Concreteness Imagery Mean Colorado Meaningfulness Mean Paivio Meaningfulness Age of Acquisition Type (for example, shows whether the word is a derivational variant of another word or ends in letter “R” that is not pronounced) Part of Speech (10 categories) PD Part of Speech (only 4 categories: noun, verb, adjective and other) Shows whether the word is an abbreviation or a suffix, or a prefix, or is hyphenated, or is a multi-word phrasal unit or none of these Status Variant Phoneme Written Capitalised Irregular Plural the actual word Phonetic Transcription Edited Phonetic Transcription Stress pattern

4.3 Sentic Based Emotional Features In our present research we show how emotional clues can help to detect personality from the text. We used emotional features in the personality engine is to find out the role of emotional features to detect personality from text. Below we first discuss the major difficulties associated with the detection of emotion from text and then discuss

the features to grasp emotion from text. All of these emotional features described below were used as the features of personality detection engine. Identifying emotions in text is a challenging task, because of ambiguity of words in the text, complexity of meaning and interplay of various factors such as irony, politeness, writing style, as well as variability of language from person to person and from culture to culture. In this work, we followed the sentic computing paradigm developed by Cambria and his collaborators, which considers the text as expressing both semantics and sentics [6]. We used a novel approach for identifying the emotions in the text by extracting the following key features using our new resource, EmoSenticnetSpace, described in Section 3.3. Later, we use these following features (which are recognized as emotional features) for the personality detection classifier. 4.4 Bag of concepts For each concept in the text, we obtained a 100-dimensional feature vector from the EmoSenticSpace. Then we aggregated the individual concept vectors into one document vector by coordinate-wise summation: N

xi = ∑ xij , i =1

where xi is the i-th coordinate of the document’s feature vector, xij is the i-th coordinate of its j-th concept’s vector, and N is the number of concepts in the document. We have also experimented with averaging instead of summation:

xi =

1 N

N

∑x , i =1

ij

but contrary to our expectation and in contrast to our past experience with Facebook data, summation gave better results than averaging. 4.5 Sentic feature The polarity scores of each concept extracted from the text were obtained from the SenticNet and summed up to produce one scalar feature. 4.6 Negation As we mentioned earlier, negations [5] can change the meaning of a statement. We followed the approach of [5] to identify the negation and reverse the polarity of the sentic feature corresponding to the concept that followed the negation marker.

5

Evaluation Results and discussions

The primary goal of the proposed approach is to investigate the impact of common sense knowledge with affective and sentiment information in personality recognition. To evaluate the proposed methods, feature vector is constructed using LIWC features, MRC features, Sentic based emotional features, and Sentic features, further these features are used to build the learning model using Sequential Minimal Optimization (SMO) classifier. We trained five different classifiers, one for each trait. Performance evaluation is performed using 10 fold cross validation. All the results in for all these classifiers for every trait are reported in Table 2. For the trait 1, openness the F-score is 0.662 which is highest among the other traits it shows that it is easiest to identity the openness trait in the text over other traits. Similarly, Agreeableness trait gives the minimum F-score of 0.615 as shown in Table 2. It shows the most difficult trait to identify among all other traits. For other traits, our model produces the F-score of 0.633, 0.634 and 0.637 respectively for Conscientiousness, Extraversion and Neuroticism trait. Table 2. Results for five traits with SVM classifier Trait Openness Conscientiousness Extraversion Agreeableness Neuroticism

Precision 0.662 0.634 0.636 0.622 0.637

Recall 0.662 0.634 0.636 0.622 0.637

F-score 0.661 0.633 0.634 0.615 0.637

Proposed method performs much better than previously reported state-of-art methods on the same dataset as shown in Table 2. Mohammad et al. (2012) [26] investigated with various feature sets and achieved the best accuracy by incorporating the emotion features. However, their reported best accuracy is quite low as compared to our proposed method as shown in Table 3. In [20], authors extracted various linguistic features which were able to incorporate the syntactic and semantic information. Our method could give better performance as compared to previous methods due to incorporation of more important information for personality detection in form of common sense knowledge with affective and sentiment information. Table 3 performance comparison with state-of-art methods

[26] [20] Proposed method

Extraversion 0.546 0.549 0.634

Neuroticism 0.557 0.573 0.637

Agreeableness 0.540 0.557 0.615

Conscientiouness 0.564 0.552 0.633

Openness 0.604 0.621 0.661

6

Conclusion

The task of personality recognition from the text has been very important. In this paper, a new approach is proposed for this task that is based on incorporating the sentiment, affective and common sense knowledge from the text using resources viz. SenticNet, ConceptNet, EmoSenticNet and EmoSenticSpace. In the proposed approach, we combined common sense knowledge with phycho-linguistic features to get the feature vector. Further, this feature vector is used by five SMO based supervised classifier for five personality traits. Experimental results show the effectiveness of the proposed approach. The main reason for this observation is the use of common sense knowledge with affective and sentiment information unlike other state-of-art approaches those were based on mostly on psycho-linguistic features and frequency based analysis at lexical level. Acknowledgments. The second author recognizes the support from the Instituto Politécnico Nacional grants SIP-IPN 20131702 and 20131441, CONACYT grant 50206-H, and CONACYT-DST India grant 122030 Answer Validation through Textual Entailment. We thank Prof. Dr. Anton Zimmerling (MSUH) for the valuable discussions during the development of this work.

References 1. Cambria, E., Havasi, C. and Hussain, A. “SenticNet 2: A semantic and affective resource for opinion mining and sentiment analysis.” Proceedings of FLAIRS, Marco Island, pp. 202–207 (2012). 2. Havasi, C., Speer, R., Alonso, J.: ConceptNet 3: A Flexible, Multilingual Semantic Network for Common Sense Knowledge. In: RANLP (2007) 3. Poria, S.; Gelbukh, A.; Hussain, A.; Das, D.; Bandyopadhyay, S., “Enhanced SenticNet with Affective Labels for Concept-based Opinion Mining,” Intelligent Systems, IEEE, vol. 28, no. 2, pp. 31–38, March-April 2013; doi: 10.1109/MIS.2013.4 (2013). 4. Havasi, C., Speer, R., Pustejovsky, J.: Automatically Suggesting Semantic Structure for a Generative Lexicon Ontology. In: Generative Lexicon (2009) 5. Lapponi, E., Read, J. and Ovrelid, L. “Representing and resolving negation for sentiment analysis,” in ICDM SENTIRE, Brussels, pp. 687–692 (2012). 6. Cambria, E. and Hussain, A. Sentic Computing: Techniques, Tools, and Applications. Dordrecht, Netherlands: Springer, (2012). 7. E. Cambria, R. Speer, C. Havasi, and A. Hussain. SenticNet: A publicly available semantic resource for opinion mining. In: AAAI CSK, pp. 14-18, Arlington (2010) 8. Alghamdi, A., Aldabbas, H., Alshehri, M., & Nusir, M. (2012). Adopting User-Centered Development Approach For Arabic E-Commerce Websites. International Journal of Web & Semantic Technology (IJWesT). 9. Amichai-Hamburger, Y., & Vinitzky, G. (2010). Social network use and personality. Computers in Human Behavior, 26(6), pages 1289–1295. 10. Boden, M., & Mellor, D. H. (1984). What Is Computational Psychology? Proceedings of the Aristotelian Society, Supplementary Volumes, 58, 37–53.

11. Briggs Myers, I., McCaulley, M. H., Quenk, N., & Hammer, A. (1998). MBTI Handbook: A Guide to the development and use of the Myers-Briggs Type Indicator (3rd ed.). Consulting Psychologists Press. 12. Cattell, H. E. P., & Mead, A. D. (2008). The Sixteen Personality Factor Questionnaire (16PF) : SAGE Knowledge. The SAGE handbook of personality theory and assessment. Retrieved May 27, 2013, from http://people.wku.edu/richard.miller/520 16PF Cattell and Mead.pdf 13. Celli, F. (2012). Unsupervised Personality Recognition for Social Network Sites. ICDS 2012 The sixth international conference on digital society, (c), 59–62 14. Coltheart, M. (1981). The MRC Psycholinguistic Database. Quarterly Journal of Experimental Psychology, 33A(4), 497–505. 15. Eysenck, H. J., & Eysenck, S. B. . (1991). Manual for the EPQ-R. San Diego, CA: EdITS. 16. Golbeck, J., Robles, C., & Turner, K. (2011). Predicting personality with social media. In CHI Extended Abstracts. in CHI Extended Abstracts, 253–262. 17. Wilson, M. (1987). MRC Psycholinguistic Database: Machine Usable Dictionary, Version 2.00. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.52.6928 18. Isaacson, K., & Peacey, S. (2012). Human resources and social media. 19. John, O. P., Naumann, L. P., & Soto, C. J. (2008). Paradigm shift to the integrative Big Five trait taxonomy: History, measurement, and conceptual issues. Handbook of Personality Theory and Research, 114–158. 20. Mairesse, F., Walker, M. A., Mehl, M. R., & Moore, R. K. (2007). Using Linguistic Cues for the Automatic Recognition of Personality in Conversation and Text. Journal of Artificial Intelligence Research, 30(1), 457–500. 21. Matthews, G., Deary, I. J., & Whiteman, M. C. (2009). Personality Traits (Cambridge., pp. 23–26). Cambridge, UK. 22. Noecker, J., Ryan, M., & Juola, P. (2013). Psychological profiling through textual analysis. Literary and Linguistic Computing. 23. Oberlander, J., & Nowson, S. (2006). Whose thumb is it anyway? Classifying author personality from weblog text. Computational Linguistics, 627–634. 24. Tausczik, Y. R., & Pennebaker, J. W. (2009). The Psychological Meaning of Words: LIWC and Computerized Text Analysis Methods. Journal of Language and Social Psychology, 29(1), 24–54. 25. Tomlinson, M. T., Hinote, D., & Bracewell, D. B. (2013). Predicting Conscientiousness through Semantic Analysis of Facebook Posts. In Proceedings of WCPR13, Workshop on Computational Personality Recognition at ICWSM13 (7th International AAAI Conference on Weblogs and Social Media). 26. Mohammad, S.M., Kiritchenko S. (2012). Using Nuances of Emotion to Identify Personality. 27. Sidorov, G., Castro-Sánchez, N.A. (2006). Automatic emotional personality description using linguistic data. Research in computing science, vol. 20, 2006, pp. 89–94. 28. Sidorov, G., Castro-Sánchez, N.A. (2006). System for linguistically-based evaluation of psychological profiles. In: Dialogue-2006, Russia, pp. 464–467.