The language of character strengths: Predicting morally valued traits on social media.

OBJECTIVE: Social media is increasingly being used to study psychological constructs. This study is the first to use Twitter language to investigate the 24 Values in Action Inventory of Character Strengths, which have been shown to predict important life domains such as well-being.
METHOD: We use both a top-down closed-vocabulary (Linguistic Inquiry and Word Count) and a data-driven open-vocabulary (Differential Language Analysis) approach to analyze 3,937,768 tweets from 4,423 participants (64.3% female), who answered a 240-item survey on character strengths.
RESULTS: We present the language profiles of (a) a global positivity factor accounting for 36% of the variances in the strengths, and (b) each of the 24 individual strengths, for which we find largely face-valid language associations. Machine learning models trained on language data to predict character strengths reach out-of-sample prediction accuracies comparable to previous work on personality (rmedian = 0.28, ranging from 0.13 to 0.51).
CONCLUSIONS: The findings suggest that Twitter can be used to characterize and predict character strengths. This technique could be used to measure the character strengths of large populations unobtrusively and cost-effectively.

INTRODUCTION

Studies have shown that social media (e.g., Facebook and Twitter) provides substantial quantities of autobiographical language and linguistic behavior that are related to users' psychological characteristics (e.g., Eichstaedt et al., 2015; Kosinski, Stillwell, & Graepel, 2013). Twitter averages about 330 million monthly active users (Statista, 2018), with about 500 million daily tweets (Aslam, 2018). People use social media to discuss thoughts, opinions, feelings, and the activities and relationships that constitute their everyday lives (Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Kosinski, et al., 2013). For these reasons, social media platforms are rapidly gaining recognition as research tools for the social sciences (Ruths & Pfeffer, 2014). Among studies which show that social media can be used to generate insights and predictions concerning psychological constructs, personality traits have received considerable attention in recent years (e.g., Dewey, 2015; Park et al., 2015; Qiu, Lin, Ramsay, & Yang, 2012; Schwartz, Eichstaedt, Kern, et al., 2013; Youyou, Kosinski, & Stillwell, 2015).

Research using social media to study psychological traits has so far focused primarily on the Five‐Factor Model of personality (FFM or the Big 5, see Azucar, Marengo, & Settanni, 2018, for an overview). For example, results reveal that extraverts are more likely to mention social words (e.g., party, Schwartz, Eichstaedt, Kern, et al., 2013), are more prone to use social media (e.g., Blackwell, Leaman, Tramposch, Osborne, & Liss, 2017), connect with more friends on social media (Kosinski, Bachrach, Kohli, Stillwell, & Graepel, 2014), and tend to have more Twitter followers (Quercia, Kosinski, Stillwell, & Crowcroft, 2011) than introverts. Individuals who are high in Openness are more likely to use words related to creativity and imagination (e.g., art and dream, Schwartz, Eichstaedt, Kern, et al., 2013), tend to have larger networks (Quercia, Lambiotte, Stillwell, Kosinski, & Crowcroft, 2012), express more “likes,” have more status updates, and engage in more group activities on social media (Bachrach, Kosinski, Graepel, Kohli, & Stillwell, 2012) than individuals low in the trait. Individuals with high Neuroticism use more negative words in their posts (Schwartz, Eichstaedt, Kern, et al., 2013), are more prone to use social media as a safe place for self‐presentation (Seidman, 2013), have fewer Twitter followers (Quercia et al., 2011), and are more likely to be addicted to the Internet (Blackwell et al., 2017) than low Neuroticism individuals. Agreeable individuals are relatively likely to express positive emotions in their posts (Schwartz, Eichstaedt, Kern, et al., 2013) and to display positive emotions in their profile pictures (Liu, Preotiuc‐Pietro, Samani, Moghaddam, & Ungar, 2016). Individuals with high Conscientiousness appear to be cautious in their online self‐presentation; they tend to post fewer pictures (Amichai‐Hamburger & Vinitzky, 2010), join fewer groups, and use “likes” less frequently on social media (Kosinski et al., 2014) than low Conscientiousness individuals. Their tweets tend to be more clicked, replied, and retweeted (Quercia et al., 2011). In addition to looking at links between social media behavior and the Big 5, recent studies also have explored how social media can be used as a tool to predict the Big 5. For example, Park et al. (2015) provide evidence that language‐based assessments (Facebook language) agree with self‐reports and informant reports of personality. Surprisingly, Facebook Likes are more accurate than peer‐ratings of personality, that is, those made by participants' Facebook friends (Youyou et al., 2015). Another study has shown that Twitter profiles (e.g., followers, following, and listed counts) can accurately predict users' Big 5 traits with a root‐mean‐squared error below 0.88 on a 1–5 scale (Quercia et al., 2011).

Despite growing interest in research on social media and the Big 5, much less is known about other models of personality. We focus on the morally valued traits, which have been mostly neglected within personality psychology for a long time. This neglect can be dated to the time of Gordon Allport (1897–1967), who claimed in the 1930s that character is merely “personality evaluated, and personality is character devalued” (Allport, 1937, p. 52). Peterson and Seligman (2004) can be seen as a milestone in reviving interest in morally valued traits as a distinct topic of research by proposing the Values‐in‐Action (VIA) classification of character strengths. Character strengths are a family of morally valued traits that have emerged across cultures and throughout history as important for contributing to a fulfilling life (Peterson & Seligman, 2004). Character strengths are associated with the good life, or positive life outcomes: studies have shown links between character strengths and positive emotions (e.g., Güsewell & Ruch, 2012), academic achievement (e.g., Weber & Ruch, 2012), healthy behaviors (e.g., Proyer, Gander, Wellenzohn, & Ruch, 2013), mindfulness (Pang & Ruch, 2019a), life satisfaction, and multi‐dimensional well‐being (Wagner, Gander, Proyer, & Ruch, 2019), and orientation to happiness (e.g., Buschor, Proyer, & Ruch, 2013; Peterson, Ruch, Beermann, Park, & Seligman, 2007; Ruch, Huber, Beermann, & Proyer, 2007). Beyond that, character strength interventions have been shown to improve well‐being and reduce depressive symptoms and stress (e.g., Gander, Proyer, Ruch, & Wyss, 2013; Proctor et al., 2011; Proyer, Gander, Wellenzohn, & Ruch, 2015; Proyer, Ruch, & Buschor, 2013, Pang & Ruch, 2019b). Supplemental Table S1 outlines the framework of the VIA classification, including an overview of the 24 character strengths.

Although there are both conceptual and empirical overlaps between character strengths and the Big 5, such as Agreeableness with kindness and Conscientiousness with perseverance (Macdonald, Bore, & Munro, 2008; Peterson & Seligman, 2004), recent studies also have identified substantial distinctiveness between the two models of personality traits. Park and Peterson (2006) found correlations between VIA strengths and the FFM variables no greater than 0.50 in a group of adolescents. Noftle, Schnitker, and Robins (2011) revealed that the percentage of variance in character strengths explained by the Big 5 domains ranges from 14% (spirituality) to 46% (persistence) with a mean percentage of 33% across the 24 strengths. McGrath, Hall‐Simmonds, and Goldberg (2017) demonstrated that spirituality is the least effectively represented by the FFM facet measures (less than 20% of explained variance) and in only three cases (creativity, forgiveness, and perseverance) does their best single predictor account for as much as half the variance in the strength scale. In addition to the direct relationship between character strengths and the Big 5, there is evidence for the incremental validity of the former over the latter from predicting self‐reports of well‐being (Johnsen, 2014; Noftle et al., 2011), helping behaviors (Lefevor & Fowers, 2016), and other behavioral criteria (e.g., friendliness and erudition; McGrath et al., 2017). Therefore, looking at the language of the VIA character strengths on social media in addition to the current findings for the Big 5 would allow us to capture more nuanced individual differences and provide a richer understanding of character strengths.

The 24 strength scales are positively intercorrelated, raising the question of an underlying global factor. For example, Ruch et al. (2010) discovered comparable intercorrelations among the scales in both self‐ (median r = 0.36) and peer‐reports (median r =0.38), and McGrath (2014) reported a mean intercorrelation of derived factors of 0.39. The first unrotated principal component (FUPC) alone typically explains about 40% of the variance (McGrath, 2015). This is why Ng, Cao, Marsh, Tay, and Seligman (2017), when identifying the factor structure of the scales, chose to apply a bi‐factor model with a separate global positivity factor (GPF) capturing dispositional tendencies toward well‐being (rather than a methodological artifact).

Thus, we expect that substantial overlap among the language correlates of the 24 character strengths would make it difficult to determine patterns distinctive to each of the 24 strengths. For this reason, we examine the language insights of this GPF separately from the language insights of the 24 individual strengths. We postulate that (a) the GPF, namely the FUPC, will capture increased use of positivity‐related words associated with higher scores on character strengths overall (Ng et al., 2017); and (b) each character strength will yield specific language insights when the other 23 strengths, as well as age and gender, are controlled for.1 Our goal is to identify a unique linguistic profile for each of the 24 character strengths and to provide insights regarding the cognitive, affective, and behavioral concomitants of these morally valued traits.

THE PRESENT STUDY

The primary goal of this study is to use Twitter language to illuminate the expression of the 24 character strengths. We use both a dictionary‐based approach (Linguistic Inquiry and Word Count [LIWC] 2015; Pennebaker, Booth, Boyd, & Francis, 2015) and a data‐driven open‐vocabulary method (Differential Language Analysis [DLA], Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013). We hypothesize that significantly associated words and topics will yield nuanced linguistic cues for the GPF and each character strength. A supplementary goal of the present study is to predict user‐level character strengths from Twitter language models, which eventually could serve as a cost‐effective and scalable way to assess character strengths. A prediction tool will be useful because the reliable and valid measures of the character strengths, including the original 240‐item (Peterson, Park, & Seligman, 2005) and the revised 120‐item measures (McGrath, 2017), are quite long and mostly self‐report measures.

MATERIALS AND METHODS

Participants and procedure

From an initial pool of 17,636 self‐selected volunteers, 4,423 participants ultimately were analyzed in the current study (see Supplemental Figure S1 for the participant flow and the selection criteria2). The initial self‐selected volunteers registered on the Authentic Happiness site (http://www.authentichappiness.sas.upenn.edu) hosted by the Positive Psychology Centre at the University of Pennsylvania and completed the Values in Action Inventory (VIA) of Strengths using their personal devices. Upon registration, participants had the option to provide their Twitter handle for research purposes, after reading and agreeing to an informed consent statement.

The final sample consisted of 4,423 participants (64.3% female) ranging from 18 to 65 years in age (M = 32.3, SD = 12.5). The participants were well‐educated: 0.9% of them had less than a high school degree (n = 39); 38.9% of them were high school graduates or some college course work (n = 1,722); and 60.2% of them had a bachelor's degree or more (n = 2,262). The sample covered a variety of occupations, including students (29.3%), professionals (13.2%), clerks (8.2%), chief executives (6.1%), manual laborers (5.3%), artists and actors (3.8%), homemakers (0.4%), and people who were retired/unemployed/invalid (3.8%). Around one‐third of them did not report their occupations (29.8%). The majority of the participants came from the United States (n = 2,783; 62.9%). The rest of the participants came from the United Kingdom (n = 383; 8.7%), Canada (n = 280; 6.3%), Australia (n = 256; 5.8%), and other countries (<2%).

We used the Twitter Application Programming Interface (API) to query up to the most recent 3,200 tweets from each volunteer.3 This resulted in 3,937,768 status updates. Respondents were not paid for participating but were provided with an automatically generated summary of their character strengths. All procedures were approved by the University of Pennsylvania Institutional Review Board (protocol #816091).

Character strengths measure

The character strengths of the participants were measured by the Values in Action Inventory of Strengths (VIA‐IS; Peterson et al., 2005). It is a self‐report questionnaire consisting of 240 items, which measures the 24 character strengths of the VIA classification (10 items for each). A sample item is “I never quit a task before it is done (perseverance).” The reliability of the 24 scales was adequate to high with Cronbach's alpha ranging from 0.74 (prudence/honesty) to 0.90 (spirituality) with a median value of 0.80 (see Table 1).

Table 1

Descriptive statistics of the 24 character strengths and their correlations with the outcome variables

	α	M	SD	Skewness	Kurtosis	L GPF	r residual
Gratitude	0.83	4.03	0.58	−0.57	0.16	0.74	0.58
Zest	0.82	3.72	0.64	−0.49	0.10	0.73	0.56
Leadership	0.75	3.94	0.50	−0.42	0.41	0.73	0.63
Hope	0.84	3.78	0.67	−0.68	0.37	0.73	0.58
Perspective	0.78	3.92	0.51	−0.34	−0.02	0.72	0.62
Social intelligence	0.76	3.87	0.53	−0.52	0.30	0.67	0.68
Honesty	0.74	4.03	0.48	−0.39	0.17	0.66	0.66
Kindness	0.78	4.03	0.52	−0.51	0.24	0.65	0.66
Fairness	0.76	4.13	0.47	−0.56	0.47	0.65	0.62
Teamwork	0.75	3.86	0.53	−0.58	0.59	0.64	0.66
Curiosity	0.79	4.08	0.51	−0.44	−0.02	0.62	0.64
Bravery	0.82	3.77	0.60	−0.39	−0.06	0.62	0.68
Perseverance	0.87	3.68	0.66	−0.45	−0.10	0.61	0.66
Love	0.75	3.95	0.57	−0.63	0.33	0.60	0.76
Self‐regulation	0.76	3.34	0.65	−0.20	−0.28	0.57	0.73
Judgment	0.78	4.08	0.47	−0.37	0.02	0.56	0.68
Spirituality	0.90	3.43	0.95	−0.23	−0.83	0.52	0.79
Forgiveness	0.86	3.74	0.66	−0.58	0.37	0.51	0.78
Creativity	0.88	3.85	0.66	−0.48	−0.04	0.48	0.76
Humor	0.84	3.98	0.59	−0.68	0.55	0.48	0.78
Prudence	0.74	3.52	0.57	−0.14	−0.28	0.47	0.71
APP beauty	0.85	3.84	0.69	−0.64	0.24	0.46	0.76
Love of learning	0.83	3.89	0.63	−0.49	−0.38	0.34	0.74
Modesty	0.80	3.42	0.65	−0.39	−0.10	0.32	0.77
Median	0.80					0.61	0.68
	Component
	1	2	3	4	5	6	7
Eigenvalues	8.57	2.27	1.75	1.55	1.24	0.96	0.81
% of variance	35.69	9.46	7.29	6.47	5.18	3.99	3.35

N = 4,423. α = Cronbach's alpha; M = mean; SD = standard deviation; L GPF = loading on the first unroated principal component, namely the global positivity factor; r residual = Pearson's correlation with the residual of each character strength partialled out the other 23 strengths. The display order of the strengths was sorted by the loadings of the FUPC, all L GPF ≥ 0.60 were bold. APP beauty = appreciation of beauty and excellence. All correlations are significant at p < 0.001, two‐tailed.

Outcome variables

Two sets of outcome variables were defined in the present study. First, the FUPC of the 24 character strengths represented the GPF. Second, to derive distinctive linguistic insights for each character strength, we controlled for the influence of the other 23 character strengths using the residual of the character strength from a regression analysis with the specific character strength as a criterion and the remaining others as predictors. As shown in previous studies, age and gender impacted language use (e.g., Kern et al., 2014, Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013), and thus we controlled for these demographics in all analyses by including them as covariates in our regression models.

Linguistic analyses

Closed‐vocabulary

First, using our Python‐based open‐source code base, the Differential Language Analysis Toolkit (DLATK; Schwartz et al., 2017), we extracted 73 dictionaries (“categories”) provided by LIWC2015 (Pennebaker, Boyd, Jordan, & Blackburn, 2015). Dictionaries included psychological (e.g., positive emotion), life domain (e.g., family and home), and syntactic categories (e.g., pronouns). We also extracted the relative frequency of each dictionary (i.e., the total number of time a word written by the user matches a word in a given dictionary, divided by the user's total number of words). We explored the most positive and negative correlations of the relative frequency of the LIWC categories with the GPF score. In addition, we explored the most positively correlated LIWC categories of each character strength. Our reason for using LIWC was to examine the correlates of the 24 strengths in a variety of domains. Moreover, using LIWC categories had the advantage that results found here could be more easily compared with the existing literature as LIWC has been widely used in psychology research.

Open‐vocabulary

Second, again using DLATK, we performed DLA (Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013) to identify the most distinguishing language features for our outcomes. We split (“tokenized”) the tweets into words, punctuation, emoticons (tokenization; Potts, 2011), and we extracted phrases consisting of two or three consecutive words (called 1–3 grams in the present study, for details of the methodology see Kern et al., 2016; Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Kosinski, et al., 2013; Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013). We kept only those two‐ and three‐word phrases with high pointwise mutual information (PMI = 5; Church & Hanks, 1990; Lin, 1998), a ratio of the probability of observing the phrase to the probability of observing the constituent words independently. This procedure yielded 11,901 language variables for each user, encoding the use of tokens and phrases. We correlated the GPF and the 24 residuals of the character strengths against all the one‐ to three‐word phrases we extracted from their tweets and shortlisted the most strongly associated words/phrases. As this is an exploratory technique, we utilized the Benjamini–Hochberg procedure (Benjamini & Hochberg, 1995) to correct for multiple comparisons and control the false discovery rate (FDR) over correlation tests for 11,901 language features. We selected only Benjamini–Hochberg significant 1–3 grams and topics.

Third, we used a set of 2,000 previously created topics (Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Kosinski, et al., 2013), clusters of semantically related words derived through Latent Dirichlet Allocation (LDA), a clustering algorithm akin to factor analysis but appropriate for the statistical distributions of words (Blei, Ng, & Jordan, 2003). For each user, we extracted the relative use of these 2,000 topics.

Predictive model based on language

We trained and evaluated a ridge regression model (Hoerl & Kennard, 1970) to predict the users' 24 character strengths (the original scale scores) using the 2,000 topics as predictors, using age, and gender as covariates. The statistical model could be summarized as follows:

Y , referred to the scores of each users' character strengths with i representing the user index (ranges from 1 to 4,423) and t representing the strength index (ranges from 1 to 24). X referred to the probability of a subject's use of each LDA topic with p representing the topic index (from 1 to 2,000) and C referred to the intercept for strength t. Ridge regression models were trained and evaluated using 10‐fold cross‐validation (CV). In this procedure, the 4,423 users were split into 10 groups. For each group G, a ridge regression model was trained on the other 9 groups and then used to predict scores for G. For each group, we tested an array of ridge regularization parameters and reported the predictions corresponding to the model with the best performance on G. In this way, we ultimately obtained out‐of‐sample predictions for all 4,423 users. The predictions were out‐of‐sample in the sense that the model was trained only on the training set, although the ridge parameter ultimately was tuned on the prediction set. The accuracy of the predictive model was assessed by the Pearson's r coefficients (correlation between the user's character strengths score and their out‐of‐sample predicted values) and as the mean absolute error (MAE, the absolute difference between the user's character strengths score and their predicted values, in units of the 1–5 original scale). As a baseline, we used age and gender of the users to predict each character strength. After conducting the Fisher's (1915) z transformation, a t test was used to compare the two correlation coefficients.

RESULTS

For preliminary analyses, we computed descriptive statistics for all 24 character strengths, the loadings of each character strength on the FUPC and the Pearson's r correlation of each character strength with the other 23 strengths partialled out. Additionally, we included the eigenvalues of the first seven components as well as their explained variance. The results are displayed in Table 1.

The language of the GPF

As shown in Table 1, the GPF explained 35.7% of the variance in our sample. Almost all 24 character strengths loaded well on the FUPC with loadings ranging from 0.32 (modesty) to 0.74 (gratitude), and most loadings were close to the median of 0.61. Both closed (LIWC) and open (DLA) vocabulary analyses revealed that the strongest positive correlations with the GPF score were words suggestive of social affiliation, positive emotions, and first person plural pronouns (e.g., love and our, β = 0.13 to.15, p < 0.001). The strongest negative language correlations with the GPF score were common adverbs, negative emotions, and words related to differentiation and tentativeness (e.g., also, bad, but, and would, β = −0.17 to −0.16, p < 0.001). Table 2 shows the top 10 most positively and negatively correlated LIWC categories (as well as the most frequent words within these categories). In addition, a high GPF score was associated with words that indicated a positive life attitude, such as blessed, patience, moments, and passion, whereas a low GPF score associated with hedging words such as actually, probably, supposed, and apparently.

Table 2

The top correlations between linguistic categories and the global positivity factor

LIWC categories	Representative words driving the LIWC correlation	β	95% CI
			Lower	Upper
Affiliation	we, love, our, Twitter, us	0.15	0.12	0.18
Positive emotion	love, good, great, thanks, happy	0.14	0.11	0.17
1st pp plural	we, our, us, let's, we're	0.13	0.10	0.16
2nd pp	you, your, u, you're, yourself	0.12	0.09	0.15
Achievement	best, work, first, better, trying	0.10	0.07	0.13
Power	up, best, over, help, down	0.09	0.06	0.12
Family	family, mom, baby, dad, bro	0.08	0.05	0.11
Female references	her, she, girl, mom, she's	0.06	0.03	0.09
Reward	get, good, great, best, got	0.06	0.03	0.09
Religion	god, hell, soul, holy, pray	0.06	0.03	0.09
Auxiliary verbs	is, be, are, have, I'm	−0.10	−0.13	−0.07
Nonfluencies	well, oh, ugh, ah, huh	−0.10	−0.13	−0.07
Conjunctions	and, so, but, when, how	−0.11	−0.14	−0.08
1st pp singular	I, my, me, I'm, I've, I'll	−0.11	−0.14	−0.08
Sexual	fuck, gay, sex, sexy, dick	−0.12	−0.15	−0.09
Past focus	was, got, been, had, did	−0.13	−0.16	−0.10
Anger	hate, fuck, hell, stupid, mad	−0.15	−0.18	−0.12
Tentative	if, or, some, hope, most	−0.16	−0.18	−0.13
Differentiation	not, but, if, or, really	−0.17	−0.19	−0.14
Common adverbs	just, so, when, how, about	−0.17	−0.20	−0.15

pp = personal pronouns; 95% CI = 95% confidence interval; β = standardized linear regression coefficients adjusted for gender and age. All results were a significant (p < 0.001).

In a similar pattern, the LDA topics that correlated most positively with the GPF score revolved around social connections (e.g., family and friends, β = 0.17), religiousness (e.g., god and lord, β = 0.16), a sense of gratitude (e.g., blessed and thankful, β = 0.16), faith and optimism (strengths and overcome, β = 0.15), an attitude of living in the present (life and cherish, β = 0.14), and positive emotions (e.g., happiness and joy, β = 0.14). By contrast, the topics that correlated most negatively with GPF revolved around negative emotions (bad, β = 0.16), negations (e.g., wasn't and isn't, β = 0.16), a more past‐oriented cognitive style (e.g., thought, forgot, and realized, β = 0.14) and hedging (e.g., supposed and apparently, β = 0.14). Figure 1 shows the 100 most distinctive words and phrases as well as the LDA topics for the GPF.

Figure 1

The language of the global positivity factor (GPF). Words, phrases, and topics that most highly correlated with the GPF: positive (top) versus negative (bottom) correlations. Words and phrases are in the center, with the size of the words indicates the strengths of the correlation and color indicates relative frequency of the usage. Topics, represented as the most 15 prevalent words, place on the left and right sides with size indicating the prevalence of the words in the topic and the colors are random (see Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013). β = standardized linear regression coefficients adjusted for gender and age. Underscore (_) connect words of multiword phrases. All shown are Benjamini–Hochberg significant at FDR 0.05 [Color figure can be viewed at https://www.wileyonlinelibrary.com]

The language of the 24 character strengths

As discussed above, the 24 character strengths substantially co‐vary. To derive the distinctive language insights for each specific character strength, we partialled out the influence of the other 23 character strengths, using the residual as the outcome variable in our linguistic analysis. The residual of the character strengths correlated significantly with the scale scores of the strengths, with median correlation of r = 0.68 (ranging from 0.56 for zest to 0.79 for spirituality, see Table 1). The results of the linguistic analysis for each character strengths are summarized in Figure 2. Additionally, to give a better sense of the context in which the most correlated words/phrases/topics appeared, we present random selections of tweets featuring these items (see Supplemental Table S2).

Figure 2

The Linguistic Analysis of the 24 Character Strengths, including LIWC, Relative Frequency of Single Words and Phrases, and Topics. PP = personal pronouns. β = standardized linear regression coefficients adjusted for gender and age. Only the top 5 significant correlations with LIWC categories were displayed (*p < 0.05, *p < 0.01, *p < 0.001; empty means that no correlations were significant at p < 0.05). The words and phrases that most highly correlated with each character strength were shown in the middle, with the size of the words indicates the strengths of the correlation and color indicates relative frequency of the usage. Topics, represented as the most 15 prevalent words, place on the right sides with size indicating the prevalence of the words in the topic and the colors are random (see Schwartz, Eichstaedt, Dziurzynski, Kern, Blanco, Ramones, et al., 2013). Underscore (_) connect words of multiword phrases. The most correlated three LDA topics were selected while filtering duplicated topics (All shown are Benjamini–Hochberg significant at FDR 0.05, empty means that no correlations were significant at FDR 0.05)

As shown in Figure 2, the DLA results of creativity showed significant associations with words indicative of people who work in creative professions such as technology (e.g., Facebook and technology) and creative work (e.g., creative, design, and artist). The top three LDA topics also indicated creative professions (Facebook and hacked; computer and program; and art, design, and museum).

The language of curiosity suggested an interest in exploring new experiences. The top‐correlated LIWC categories were space and relativity (e.g., in, on, and at; indicators of being in new/different places), leisure (e.g., twitter, fun, and play; indicators of exploring) and ingestion (e.g., eat, water, and sweet; indicators of trying new food/drinks). Similarly, the most correlated words were related to space and relativity (e.g., on); traveling or other cultures (e.g., France and Korean); and leisure and activities (e.g., festival and park). The top‐correlated topics which also referenced travel destinations (Paris and London), leisure (lake and boat), and activities (festival and film).

The most representative words of judgment/critical thinking referenced thinking (e.g., know and think), consideration (e.g., bad and appropriate), and differentiation (e.g., not and but), which were essential for thinking things through (e.g., stop and bad) and examining from all sides (e.g., not and don't). The highly correlated topics likewise revolved around opinions, objective statements, and judgmental comments.

Love of learning was associated with syntactic categories that mark more complex language use (e.g., use of articles, the and a), topics concerning school (e.g., school, books, and read), insights (e.g., know and think), and inquisitive language (e.g., why and how). This is consistent with the top‐correlated topics like opinions, books, and reading, as well as political discussions.

The language of perspective was in line with a view toward life that emphasizes what matters most. Top‐related words and phrases included important, makes me so, so much, life, and statements, similar to the only cluster of topics significantly associated with perspective (e.g., important, life, things, and realize).

Bravery seemed to be associated with references to aggression (e.g., hate, kill, and fuck), masculinity (e.g., he, father, and man), freedom (e.g., fight), rights (e.g., rights, racist, and woman), and politics (e.g., vote and political). The top‐related topics included swear words, freedom and rights, and politics.

The most representative words of perseverance revolved around work (e.g., work), school (e.g., graduation and congrats), achievement (e.g., best and first), and suggested a future‐oriented mindset (e.g., when and new), in line with individuals who tended to complete the tasks they set out to accomplish. The highly correlated topics likewise revolved around study (e.g., English and history), graduation (e.g., congrats and graduation), and achievement (e.g., grades and final).

The most representative words of individuals high in the honest/authentic strength included self‐reference (e.g., I and my) and revelations of personal distress (e.g., sleep and head) as well as potentially reduced self‐control, indicated by greater use of swear words (e.g., fuck and hell). These also appeared similarly in the top‐related topics, such as sleep, gotta, and tired.

The most representative words of zest showed positive emotions (e.g., love and good), excitement (e.g., great, passion, and forward), and energetic pursuit of life and work (e.g., work and school). Individuals high in this strength also mentioned more social connections (e.g., we and our) and words related to achievement (e.g., best and work). The highly correlated topics likewise revolved around weekend (e.g., weekend and holiday), positive emotions (e.g., great and awesome), and future‐orientation (e.g., forward and hope).

The words most associated with love referred to relationships (e.g., my boyfriend and he loves) and gratitude (e.g., thank and thanks so much). Individuals high in this trait also seemed to value close relationships with others (e.g., wish I could and guardian [angels], Supplemental Table S2), and care about sharing (e.g., care about and recourses).

Individuals high in kindness expressed support of others (well, hugs, and sorry to hear), were interested or engaged in charity (e.g., raise [money], Supplemental Table S2), and appeared to value close relationships (e.g., best friends and washing [for others], Supplemental Table S2).

The most representative words of individuals high in social intelligence tended to be informal (e.g., u, :) and lol), with positive emotional content (e.g., love and good), and social language (e.g., catch up and conversation). The top‐related LDA topics showed similar patterns, revolving around social events (e.g., night, town, and carnival) and positive emotions (e.g., love, hugs, and xoxo).

The most representative words of teamwork reflected achievement. The top‐correlated LIWC categories were reward (e.g., good and great, β = 0.04, p < 0.05), achievement (e.g., best and first, β = 0.04, p < 0.05), and religion (e.g., god and holy, β = 0.04, p < 0.05). In addition, words of support and encouragement (e.g., c'mon [come on] and congrats), future orientation (e.g., future), family (e.g., father), and work life (e.g., office) were highly correlated with the character strength of teamwork. However, we found no significant correlations between LDA topics and teamwork.

Fairness was associated with words related to self‐reference (e.g., I and my) and more frequent negations (e.g., but, no, and don't). The highly correlated topics likewise revolved around common adverbs (e.g., honestly and anymore), negations (e.g., don't and won't), and common verbs (e.g., talking and suppose).

Leadership as a character strength was associated with the language of affiliation (e.g., we and our) and activities and events commonly engaged in by leaders (e.g., challenging, workshops, presentation, and manage). The top‐correlated topics suggested further common behaviors of leaders, such as giving charity (e.g., donate and raise), attending events (e.g., event and ticket), and acting future‐oriented (e.g., hope and forward).

The most representative words of forgiveness were indicative of close relationships (e.g., marriage) and the process of apology (e.g., believed and appeal).

Similarly, modesty showed associations with races, which in this context appeared to be running competitions (see Supplemental Table S2 for more details). In addition, the words related to modesty also involved proclamations of effort (e.g., forward to, hustle, and catch up).

Prudence was associated with adverbs such as simultaneously, also, recently and apparently. Individuals high in prudence seemed to talk about clicking links (clicked) and also tended to use LDA topics previously shown to be characteristic of introverts (Schwartz, Eichstaedt, Kern, et al., 2013), such as suspenseful movies (e.g., sherlock and inception), anime (e.g., anime and manga), and common adverbs and verbs (e.g., apparently and found).

The most representative words associated with self‐regulation was suggestive of rigorous, healthy lifestyle and self‐discipline (e.g., life, health, gym, workouts, diet, and weights). The highly correlated topics likewise included workouts (e.g., gym and exercise), diet (e.g., eating and drinking), and losing weight (e.g., lose and pounds).

The most representative words of appreciation of beauty and excellence was suggestive of individuals who expressed themselves (e.g., I and my; indicators of expressing oneself) emotionally, liked to observe esthetic work attentively (e.g., see, look, art, and beautiful; indicators of observing attentively), and expressed intensity (e.g., fuck and hell; to address the intensity). These patterns also were revealed in the top‐correlated topics, which indicated emotional sensitivity (tears and cry), art (song and music), and positive emotions (beautiful and wonderful).

The most representative words of gratitude showed that individuals high in this strength were thankful (e.g., so grateful for and blessed) and experienced positive moods (e.g., impressive and amazing) in social contexts (e.g., congrats and happy birthday to). The most‐correlated LIWC categories were male (e.g., he and his) and social processes (e.g., you and love); both indicated objects/subjects of gratefulness. The highly correlated topics likewise included male references (e.g., dad and boyfriend), people (e.g., baby and girl), and social processes (e.g., family and friends).

Individuals high in hope were future‐orientated (e.g., a brand new) and optimistic (e.g., confident, id [I'd], and new products). The topics showed one significant result, namely abbreviation denoting one's mood (e.g., na and sa [concrete example: when you're ready come and get it la na na, see Supplemental Table S2 for more details]).

The language of humor showed that individuals high in this trait tended to talk about themselves frequently (e.g., I and me) and responded to jokes (e.g., jokes and funnier) and funny content (e.g., toilet and dumb). The highly correlated topics likewise included funny content (e.g., toilet, pee, smell and fart) and responses to jokes (e.g., hahaha and laughing).

The language associated with spirituality was indicative of individuals who practice their religion actively (e.g., god, church, praying, and lord), positive emotions (e.g., blessed), and were socially connected (e.g., family and mum). The highly correlated topics likewise revolved around religious themes such as god (god and prayers), gratitude (e.g., blessed and grateful), and religious events (e.g., service and church).

Predicting character strengths with language

As shown in Table 3, the predictive models performed comparably to other models predicting constructs like personality from behaviors,4 with models for love of learning and spirituality performing excellently (r reaching 0.51) and models for another six strengths (i.e., zest, appreciation of beauty and excellence, gratitude, curiosity, hope and self‐regulation) performing relatively well (r greater than 0.30). The strengths that were easiest to predict from Twitter language had also higher relative frequency of 1–3 grams and topics (as indicated by the color of the word clouds and topic clouds). The reason behind the difference in prediction values might be that certain character strengths were more manifest on social media platform, while other strengths were more hidden. For example, love of learning indicates a certain degree of Openness, which is linked to more social media activities (e.g., more “likes” and larger network) and also a tendency to post more content, whereas prudence indicates a degree of introversion which may correlate with less social media use.

Table 3

Prediction accuracies for the 24 character strengths

	Baseline		LDA topics
	r	MAE	r	MAE
Spirituality	0.12	0.50	0.51	0.67
Love of learning	0.24	0.39	0.51	0.44
Zest	0.03	0.53	0.36	0.47
APP beauty	0.12	0.52	0.36	0.51
Gratitude	−0.03	0.42	0.36	0.43
Curiosity	0.09	0.48	0.34	0.37
Hope	0.05	0.80	0.34	0.50
Self‐regulation	0.04	0.46	0.31	0.50
Creativity	0.17	0.54	0.29	0.51
Teamwork	0.10	0.45	0.29	0.40
Bravery	0.14	0.52	0.28	0.46
Perseverance	0.16	0.46	0.28	0.51
Kindness	0.14	0.41	0.27	0.40
Social intelligence	0.04	0.40	0.26	0.41
Humor	0.08	0.41	0.25	0.45
Perspective	0.12	0.37	0.24	0.39
Leadership	−0.04	0.52	0.24	0.39
Love	0.11	0.47	0.23	0.44
Honesty	0.03	0.38	0.22	0.37
Modesty	0.01	0.53	0.22	0.51
Forgiveness	0.09	0.38	0.20	0.51
Fairness	0.38	0.48	0.18	0.37
Judgment	0.21	0.52	0.16	0.37
Prudence	0.04	0.41	0.13	0.45

N = 4,423. r = Pearson's correlation coefficient. MAE = mean absolute error; APP beauty = appreciation of beauty and excellence. The display order of the strengths was sorted by the value of r of LDA topics, r ≥ 0.30 were bold. All targeted predictive model (prediction accuracy using LDA topics with age and gender controlled) were significant (p < 0.001) improvement over the baseline (prediction accuracy of just using age and gender) except for fairness and judgment.

DISCUSSION

The present study investigates language use associated with the 24 VIA character strengths, extending previous work on the language profiles of the Big 5 to morally valued traits (Schwartz, Eichstaedt, Kern, et al., 2013). We demonstrate that each of the 24 character strengths and a GPF are associated with distinctive language profiles and can be accurately predicted by social media language with fair accuracy.

The present study expands existing knowledge on the overlaps and distinctiveness of the Big 5 and the VIA character strengths. Consistent with what Peterson and Seligman (2004) point out, our results show that four out of the Big 5 traits have clear counterparts in the virtue domain (see Table 4). Comparing the results of Twitter language on the Big 5 (c.f. Schwartz, Eichstaedt, Kern, et al., 2013, Figure 6 and Figure S2) with our results reveals how the word clouds of the Big 5 differ or coincide the VIA character strengths. For instance, both Extraversion and zest correlate with words related to time for socialization (e.g., weekend) and positive emotion (e.g., great), but the language of zest additionally shows indications of enthusiasm (e.g., passion). In a similar vein, both Openness and appreciation of beauty are related to artistic work (e.g., music), yet the latter further emphasizes esthetic value (e.g., beautiful). This constitutes evidence that the Big 5 and VIA character strengths are complementary measures, with analysis of the VIA strengths contributing to a more nuanced understanding of individual differences.

Table 4

The overlaps of the social media language between the Big 5 and the 24 VIA character strengths

Big 5 (and the representative examples)	Suggested counterparts in VIA	Overlap of the social media language (1–3 grams and topics, compared to Schwartz, Eichstaedt, Kern, et al., 2013)
Neuroticism (worried, nervous, emotional)	None	–
Extroversion (sociable, fun‐loving, active)	Zest	Time for socialization such as night and weekend; positive emotions such as great and amazing
	Humor (playfulness)	Words related to jokes such as fun and wit
Openness (imaginative, creative, artistic)	Curiosity	Leisure such as festival and music
	Creativity	Creative work such as art and writing
	Appreciation of beauty	Artistic work such as music and poetry
Agreeableness (good‐natured, softhearted, sympathetic)	Kindness	Words that support others such as hugs and love you all
	Gratitude	Expressions of being thankful such as grateful and blessed; experiencing positive mood such as impressive and amazing; words related to close relationship such as family and friends
Conscientiousness (reliable, hardworking, punctual)	Self‐regulation	Words related to workout such as gym and workout
	Perseverance	Words related to study such as semesters and finals; words related to achievement such as work and successful
	Prudence	Instead of overlapping with Conscientiousness, prudence overlaps more with introversion, such as suspenseful movies (e.g., sherlock and inception) and anime (e.g., anime and manga)

The first two columns of the table were adapted from Peterson and Seligman (2004, Table 3.7, p. 69).

The words/phrases that are most positively associated with the GPF suggest positive emotionality, which captures a number of character strengths (e.g., beautiful, love, faith), and language associated with emotional (e.g., happy and passion) and social well‐being (family and friend) and accomplishment (e.g., success). This general pattern of results is largely consistent with previous computational linguistic analyses on religious affiliation (Yaden et al., 2017). We additionally observe language suggestive of mindfulness (i.e., focusing on the current moment, e.g., moment and breaths). This is consistent with previous studies on the association between character strengths and mindfulness (Pang & Ruch, 2019a) as well as life satisfaction (for an overview, see Bruna, Brabete, & Izquierdo, 2018). The linguistic cues in the present study provide support for the potential contribution and association of character strengths to both the hedonic (e.g., happiness or positive affect) and the eudemonic (e.g., a sense of meaning and purpose) aspects of well‐being (Wagner et al., 2019). In the pattern of negative association, we observe a tendency toward more cognition and differentiation but not the use of swear words, which mark disagreeableness and cognitive dysregulation.

These observations raise the question of what exactly the GPF is. Is it a method artifact that reflects social desirability or an indicator of positivity? Given the high loadings of all character strengths on the GPF (median loading 0.61, and higher than 0.70 for gratitude, zest, leadership, hope, and perspective), and its associated language profile suggestive of positive emotionality and well‐being, the GPF would suggest more of the latter (indicator of positivity) for the following reasons. First, gratitude, zest, hope, curiosity, and love loaded strongly (i.e., ≥0.60) on the GPF, which happen to be the five strengths most robustly correlating with well‐being across different samples (e.g., Buschor et al., 2013; Park, Peterson, & Seligman, 2004; Ruch et al., 2007, 2010; Shimai, Otake, Park, Peterson, & Seligman, 2006). The higher the loadings of each strength on the GPF, the higher its correlation value with life satisfaction (rank‐order correlation ranges 0.63 from to 0.81, computing correlations with Park et al., 2004, Table 3). Second, the overall level of virtuousness that has been ascribed to each character strength (gratitude [3.34], zest [2.72], love of learning [3.06], modesty [3.36]; Ruch & Proyer, 2015) does not seem to have a systematic impact on whether the character strength loads strongly (gratitude and zest) or weakly (love of learning and modesty) on the GPF. In sum, this suggests that the GPF may capture a sort of dispositional positivity—a trait‐like “meta positivity”—that constitutes emotional well‐being and emotional health rather than a methodological artifact. Nevertheless, it is possible that the GPF represents a prevalence of positivity‐related words because no reversed items are available in the VIA‐IS. This suggests that a “less fakable” balanced key version of the VIA‐IS (e.g., McGrath, 2017) may be worth developing and that researchers should additionally try to measure character strengths through peer ratings (Ruch et al., 2010) or structured interview (Peterson & Seligman, 2004). The theoretical implications of a general positivity factor observed in psychometric studies and computational linguistic analyses is worthy of further discussion and research.

The language insights help to reveal the everyday lives of individuals who are high in each particular strength. For example, it is easy to imagine that a person who scores high in love of learning may love books, read a lot, be interested in history and have the drive to study—and the language results support this prototypical view. We still see grateful and care in the language of love despite the fact that the positive emotionality of the GPF was partialled out, suggesting that these language cues are indicative of love over and above the GPF language profile.

If the goal is to have the full picture of each individual strength (i.e., rather than the distinctiveness of each strength), then an analysis of individual VIA‐IS scales should be undertaken. As noted, this leads to largely overlapping word clouds given the large shared overlapping variance in the GPF. We avoid this problem in the present study and provide a technique to address this challenge.5 For studies interested in what differentiates strengths, we recommend using this meta‐positivity GPF as a control variable and examining correlates of each strength above and beyond GPF.

Our prediction results are comparable to previous work on personality prediction and suggest that language‐based assessment of character strengths may one day serve as a cost‐effective and scalable alternate measurement system. Social media language (e.g., from Twitter) constitutes a new medium for assessing individual differences which allows insights into other life domains such as well‐being, job satisfaction, etc. For example, one interesting idea may be to use tweets from charismatic leaders (who probably are too busy to fill out the VIA‐IS) to predict their character strengths and thereby predict firm‐level outcomes (e.g., the cognitive strengths might be more related to improvement of revenue, while the justice strength might be more related to sustainable behaviors). All these predictions could contribute to an array of new research interests.

Limitations and implications

Despite the strong face‐validity of the language results, distinctive language insights for each individual strength need to be interpreted with caution because the influence of the other 23 strengths has been removed. For example, the language of bravery includes the connotation of being aggressive, likely because kindness and love are partialled out. This technique can provide a distillation of the unique qualities associated with any given strength but may provide a rather thin or caricatured view of each strength. As each of the VIA strengths contributes to generally positive life outcomes individually and on aggregate, the various strengths are often not only overlapping but mutually supportive and even constitutive of one another. Therefore, the VIA strengths can be viewed in a variety of different ways: as a sum total, as just a few factors, individually without controlling for the others, and individually while controlling for the influence of all of the others. We explore the last of these options because it had not been done previously and in order to provide a more granular and specific view of the linguistic correlates of the character strengths.

In addition, we acknowledge that our results are ultimately data rather than theory‐driven because we did not have a priori predictions about specific associations between character strengths and the linguistic markers. This approach is exploratory in the sense that it helps with hypothesis generation, in contrast to more traditional hypothesis testing often undertaken in psychology. The words presented in our clouds are the most highly correlated, yet our interpretations are subjective. Readers are welcome to agree or disagree and future research is welcomed to test the hypotheses we generate.

Our participants are mainly from an internet source (Authentic Happiness Website) and were not purposefully recruited, possibly resulting in a biased recruitment of people who are interested in positive psychology or who are curious about themselves. As shown in the participants flow chart (see Supplementary Figure S1), more than 3,000 volunteers gave invalid Twitter handles, which could indicate those who are honest were more likely to be included in our sample. These biases could affect the representativeness of the study. The lower amount of variance explained by the GPF compared to other studies (e.g., 41%, McGrath, 2015) might suggest greater homogeneity in our sample, which also could be seen from the characteristics of our sample (the majority of our participants were well‐educated, English‐speaking Americans). This means that the language features should be understood to describe our sample, not generalize widely. For this very reason, we do not encourage using our findings to estimate character strengths in new samples without prior replication of the findings. Finally, in this study, we predict inter‐individual differences in the strengths, not intra‐individual differences. Some applications of strengths focus on the “signature strengths,” that is, the 3–7 most highly developed strengths; further research is needed before user‐level signature strengths can be reliably predicted from Twitter language.

CONCLUSION

The current study demonstrates that social media can be used to further characterize and predict character strengths. The prediction results suggest that language‐based assessments of character strengths may well serve as a cost‐effective and scalable alternate measurement system. The consistent finding of a general “meta positivity” factor, in this study and in the literature, may suggest that research exploring differences across overlapping constructs (like character strengths) should adopt methods suitable to address this, such as partialling out the shared variance to foreground meaningful differences. The linguistic correlates associated with each character strength provide insights into the behavioral and social components of these morally valued traits, providing a rich set of hypotheses to explore in future research on character strengths.

CONFLICT OF INTERESTS

The authors declared no potential conflicts of interest with respect to the research, authorship, and publication of this article.

Click here for additional data file.