Words

Words

JOURNAL OF Journal of Research in Personality 36 (2002) 615–621 RESEARCH IN PERSONALITY www.academicpress.com Brief report Words Carla J. Groom *...

72KB Sizes 3 Downloads 282 Views

JOURNAL OF

Journal of Research in Personality 36 (2002) 615–621

RESEARCH IN PERSONALITY www.academicpress.com

Brief report

Words Carla J. Groom

*

and James W. Pennebaker

1

Department of Psychology, The University of Texas at Austin, Austin, TX 78712, USA

Abstract The words people use in their writing and speaking can serve as markers of personality. Computerized text analysis programs based on word counts represent a methodological strategy for obtaining linguistic information objectively and rapidly from diverse text samples. Several studies are summarized that point to particles or function words (e.g., pronouns, articles, and prepositions) as central determinants of linguistic styles. Particles exhibit solid psychometric properties and are related to measures of depression, social concerns, physical health, and, to a modest degree, traditional personality indices. Possible reasons for the psychological significance of particles are offered, and the challenges facing the linguistic approach to personality are considered. Ó 2002 Elsevier Science (USA). All rights reserved.

1. Introduction Words are the building blocks of meaning, social interaction, and language itself. The ways people speak say a lot about who they are. The combination of language, accent, and dialect can identify them as a child born to Italian parents in New York, a man who grew up in rural Tennessee but moved to Washington DC, or a working-class British woman from LondonÕs East End. What is less obvious is that there may be at least as much information about individuals contained in the ways in which they select *

Corresponding author. E-mail addresses: [email protected] (C.J. Groom), [email protected] (J.W. Pennebaker). 1 Also Corresponding author. 0092-6566/02/$ - see front matter Ó 2002 Elsevier Science (USA). All rights reserved. PII: S 0 0 9 2 - 6 5 6 6 ( 0 2 ) 0 0 5 1 2 - 3

616

Brief report / Journal of Research in Personality 36 (2002) 615–621

and assemble words as with their dialect or other non-verbal or paralinguistic mannerisms.

2. Language as a marker of personality and individual differences Freud believed that a personÕs everyday utterances, especially the errors that came to be called ‘‘Freudian slips,’’ could reveal details of the unconscious. McClelland also believed language had a hidden layer of meaning. He scrutinized narrative responses to projective tests such as the Thematic Apperception Test (TAT) for themes reflecting individual difference dimensions such as Need for Achievement (e.g., Atkinson & McClelland, 1948). More recently, linguistic style has been used to identify the anonymous author of a bestselling book about Bill Clinton (Foster, 1996) and to track the personality fluctuations of Rudolph Giuliani during his years as mayor of New York City (Pennebaker & Lay, 2002). It appears that word usage patterns are sufficiently specific to determine not only the type of person who produced a particular text sample, but also the unique identity of the author, much like a fingerprint or DNA sample. In all three cases, language serves as a proxy marker of individual difference variables. The clues may be in the content of speech (explicitly or implicitly) or in the ways that the people habitually package what they say in terms of pronouns, sentence length, and so on. This approach contrasts with the two ways in which language is usually treated in psychology: (a) a shared system that is used differently only as a result of trying to communicate different ideas, or (b) an instrument for extracting introspective information from participants. Using language as a marker of personality has several potential advantages. First, language samples are easy to gather. Transcriptions of interviews, pages from journals, essays written for class, and responses to open-ended questionnaire items are just a few of the many options. Participants do not need to come into the lab, nor even have to be alive at the time the study is conducted (see for example the study of suicidal poets by Stirman and Pennebaker, 2001). Second, language use is spontaneous and natural, avoiding the problems of introspective self-report. Third, language provides a common dependent measure for diverse studies, making it much easier to meaningfully synthesize their insights and extract a ‘‘big picture.’’ Until recently, one major disadvantage of a linguistic approach to the study of individual differences has been the difficulty of devising a reliable and practical means of analyzing text samples. Many researchersÕ instinctive approach to textual data has often been to assume that the complexity of the system that produced the text in the first place must at least approximate any system seeking to decode that text. With the current level of technology, this means employing human raters to perform qualitative coding. Such coding is both subjective and time-consuming. An alternative strategy is

Brief report / Journal of Research in Personality 36 (2002) 615–621

617

to step back from the context-bound complexity of language and explore whether the simple surface features can have useful predictive properties. One way of doing this is via a computerized system that simply counts the proportions of words in the text that fall into a number of pre-defined categories. Although several word-count programs have been developed over the years (e.g., Martindale, 1990; Stone, Dunphy, Smith, & Ogilvie, 1966), we focus on Linguistic Inquiry and Word Count (LIWC; Pennebaker, Francis, & Booth, 2001) as a particular methodological strategy for identifying individual differences in language content and style.

3. How LIWC works The 2001 version of LIWC codes text files along over 70 dimensions. Most of these dimensions are categories of words, reported as proportions of the total words used in the text sample under scrutiny. Some of the categories are simple language composition features, such as pronouns, prepositions, articles, and so on. The rest of the categories tap word meaning, falling under the broad headings of psychological processes (e.g., positive emotions, negative emotions, and cognitive words), relativity in time and space (e.g., verb tense, spatial and movement references), and personal concerns of the speaker or writer (e.g., work, leisure, and metaphysical concerns). Many of the categories are arranged hierarchically. For example, all negative emotion words are also counted as instances of emotion words in general (a type of psychological process reference). LIWC recognizes more than 2200 words (about 80% of the words people normally use in non-technical speech and writing), and can accept additional user-defined dictionaries. Human judges determined which words qualified for which category. For example, judges determined that the word cried was a good example of the category sadness, and by hierarchical implication, negative emotion, and emotion in general, as well as past tense words. In addition to the word categories, LIWC also returns other surface features of the text including mean sentence length and proportion of non-fluencies such as ‘‘um.’’ In several validation studies, undergraduatesÕ essays on emotional and non-emotional topics were coded by both the program and human raters on a selection of the LIWC dimensions, and found to correlate at least .5 on the majority of dimensions. The LIWC approach to text analysis contrasts with ‘‘bottom-up,’’ inductive strategies such as Latent Semantic Analysis (LSA, Landauer, Foltz, & Laham, 1998). The latter determines which words belong together not by asking judges whether they come from the same semantic or grammatical category, but rather by observing the statistical patterns of covariation between words in large text corpora, analogous to factor analysis.

618

Brief report / Journal of Research in Personality 36 (2002) 615–621

4. Evaluating the psychometric properties of language use: Content versus style For linguistic features to be a useful marker in personality research, they must display consistency across time and context. Pennebaker and King (1999) explored the reliability of the LIWC dimensions using text samples drawn from three very different populations: (a) the diaries of inpatients at an addiction clinic, completed as part of treatment, (b) essays on diverse topics written by health psychology summer school students as part of their course requirement, (c) journal abstracts written by members of the Society of Experimental Social Psychology over the course of their careers. The mean alpha across 72 LIWC dimensions was .59, suggesting that individuals show quite reliable patterns of word usage even when the topic changed dramatically, and as in the case of the SESP sample, the text samples were produced over a period of several years. Of central importance has been the finding that the reliability of language is generally much higher for language composition dimensions than for more traditional content dimensions. That is, over time, people are more consistent in how they talk than in what they say. Perhaps the best markers of talking or writing style are pronouns, prepositions, articles, conjunctions, and auxiliary verbs. These types of words comprise a broad class of words variously referred to as function words or particles. Particles form the linguistic ‘‘glue’’ that hold content words together. Consider, for example, how two people might describe a glass of water on a table. Person A: ‘‘ThereÕs a glass of water that is approximately half full resting in the left center portion of the 4-foot table.’’ Person B: ‘‘I see that someone left her glass of water on the table. I wonder why she didnÕt finish it?’’ It is immediately apparent that these two speakers differ tremendously in how they define their worlds. Person A is much more detached, using more articles and prepositions; Person B uses more pronouns and more focused on human interactions. Particles are intriguing for several reasons. They are processed differently in the brain than most nouns and regular verbs (Miller, 1995). Particles exhibit far greater reliability across speech and writing samples than any other classes of words. For example, in the case of inpatientsÕ diaries, 69% of the LIWC word categories had alphas of at least .60. For particle-relevant categories taken alone, the figure jumped to 83% compared with 58% for current concerns (Pennebaker & King, 1999). The stability of word usage patterns also transfers to spoken language. Recently, we have developed a digital tape-recording system (the Electronically Activated Recorder or EAR) that captures 30-s speech samples once every 12 min as people went about their daily lives (Mehl & Pennebaker, 2002). Comparing language samples on two days separated by a 4-week interval, the average test–retest reliability for the linguistic categories of

Brief report / Journal of Research in Personality 36 (2002) 615–621

619

LIWC was .41, compared to .24 for psychological processes and .29 for relativity. There was also considerable variability within these groups. Positive and negative emotion words showed test–retest reliabilities of .51 and .50, respectively, whereas insight words for example were correlated only .10. Most function word categories hovered around the .35 level. The highest reliabilities were found for the LIWC categories of swear words (.86), non-fluencies (.62), and filler words (.59), suggesting that spoken language may contain even more important linguistic markers of personality compared to written language which tends to omit all three categories. In terms of predictive validity, particles have been shown to relate to sources of individual differences such as age and to personality constructs such as self-monitoring (for a review, see Pennebaker, Mehl, & Niederhoffer, in press). It is ironic that the parts of speech that have no intrinsic meaning say the most about us. Although counter-intuitive, this discovery actually makes sense from two perspectives. First, particles have no pre-defined referent but do carry important meaning about the social structure of the world and the nature of the speaker–audience relationship. For example, you could respond to questions about your family using the pronoun they, we, he, she, or even it. All are grammatically permissible, but the options connote very different qualities of familial relationship. Similarly, to select the article the rather than a implies a greater extent of shared reality between the speaker and audience. In this sense, it is less surprising that particle use might be related to important personality variables. Second, the very fact that there is often room for individual choice in the decision to use a particular particle can help account for the greater reliability of this linguistic marker. Just as Mischel (1968) argued that behavior in general was highly dependent on contextual demands, so the use of particular lower-frequency words (e.g., taxes) is obligatory in certain situations and non-sensical in most others. Although people do exert influence on the situations they find themselves in and the parts of those situations that they choose to attend to, people are likely to use these content words in shared ways. Indeed, language could not serve its primary communicative function if everyone fashioned their own referents for nouns, verbs, and adjectives. Allport (1961) distinguished between adaptive and stylistic aspects of behavior. Adaptive behavior is what people do, and corresponds to the situationally-constrained action that Mischel discusses. In contrast, stylistic aspects of behavior are the unintentional, automatic, and expressive parts of action. It is in those aspects of behavior that do not have direct and obvious implications for the task at hand that personal habit becomes the primary basis for decision-making. The LIWC data suggest that this distinction applies to linguistic behaviors, and the stylistic aspects will be found in particles. It is also logical that it took a program like LIWC to identify these

620

Brief report / Journal of Research in Personality 36 (2002) 615–621

markers, because the attention of human judges will be drawn in a Strooplike fashion to the primary communicative intent of words. Scrutinizing the way that we communicate may be the original implicit measure. With the help of programs such as LIWC, it is often easier to collect sufficient data to determine linguistic style than it is to administer a TAT, lengthy questionnaire, or computerized Implicit Association Test. However, in order for the full potential of this technique to be realized, a number of challenges must be overcome. Demonstrations of the empirical reliability of linguistic style have not yet been accompanied by a strong theory of language use. A proxy marker can function without a theoretical construct, as in the case of the Rod-and-Frame and Embedded Figures tasks in the field-dependence literature, but its use is thereby limited. Given the particular difficulty of using intuition and informal observation in the realm of linguistic style, which is ‘‘invisible to the naked eye,’’ it is difficult to interpret the implications of patterns of correlations. It will also prove helpful to combine word-count approaches with more inductive strategies such as LSA, which can help suggest refinements of the category structure of word-count programs. Until linguistic style is better understood, we will suffer from an embarrassment of riches. Language samples are all around us, and a better understanding of peopleÕs particle preferences would help to identify which of the virtually infinite number of feasible projects will be most likely to help us to understand how and why people differ from each other.

Acknowledgments Preparation of this manuscript was funded by a grant from the National Institutes of Health, MH52391.

References Allport, G. W. (1961). Pattern and growth in personality. New York: Holt Rinehart & Winston. Atkinson, J. W., & McClelland, D. C. (1948). The effect of different intensities of the hunger drive on thematic apperception. Journal of Experimental Psychology, 38, 643–658. Foster, D. (1996). Primary culprit: Who is anonymous? New York Magazine, 29, 50–57. Freud, S. (1901). Psychopathology of everyday life. New York: Basic Books. Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 25, 259–284. Martindale, C. (1990). A clockwork muse: The predictability of artistic change. New York, NY: Basic Books. Mehl, M.R., & Pennebaker, J.W. (2002). The sounds of social life: A psychometric analysis of studentsÕ daily social environments and natural conversations. Manuscript submitted for publication. Miller, G. (1995). The science of words. New York: Scientific American Library.

Brief report / Journal of Research in Personality 36 (2002) 615–621

621

Mischel, W. (1968). Personality and assessment. New York: Wiley. Pennebaker, J. W., Francis, M. E., & Booth, R. J. (2001). Linguistic inquiry and word count: LIWC 2001. Mahwah, NJ: Erlbaum Publishers. Pennebaker, J. W., & King, L. (1999). Linguistic styles: Language use as an individual difference. Journal of Personality and Social Psychology, 77, 1296–1312. Pennebaker, J. W., & Lay, T. C. (2002). Language use and personality during crises: Analyses of Mayor Rudolph GiulianiÕs press conferences. Journal of Research in Personality, 36, 271– 282. Pennebaker, J.W., Mehl, M.R., & Niederhoffer, K.G. (in press). Psychological aspects of natural language use: Our words, our selves. Annual Review of Psychology. Stirman, S. W., & Pennebaker, J. (2001). Word use in the poetry of suicidal and non-suicidal poets. Psychosomatic Medicine, 63, 517–522. Stone, P. J., Dunphy, D. C., Smith, M. S., & Ogilvie, D. M. (1966). The general inquirer: A computer approach to content analysis. Cambridge, Mass: MIT Press.