Personality Profiles in Online Job Advertisements: Applying a Weighted Closed Vocabulary Approach to Identify Generic Profiles and Clusters

Abstract

On the background of the supplementary applicant-job fit theory, this study tests the application of a weighted closed vocabulary approach as introduced by Ostendorf on a large corpus of online job advertisements (OJA). Therefore, a large dataset of online job advertisements (OJAs) was scraped from a German job portal (N = 151k), and a sample (n = 3,239) was selected based upon Porter’s Value Chain and corresponding job functions. The dataset was pre-processed with techniques of natural language processing, and the predefined words extracted from the OJAs were weighted with the average subject matter expert rating, as offered in the vocabulary of Ostendorf. The results reveal that, first, 0 to 12 words describing personality are used across OJAs, second, that HR, sales, and finance positions show values higher than sample mean (3.22), third, that HR positions show loadings across all dimensions of personality above the sample means (extroversion: 2.32, agreeableness: 3.98, conscientiousness: 4.28, emotional stability: 3.54, openness: 2.0) followed by sales. The job profiles being significantly parallel on equal levels are HR clerk and referee, as well as IT, R&D, and procurement. While conscientiousness is reported to be an important predictor of performance, our study showed only low to average importance in OJAs.

Introduction

Advancing vocabulary-based research in text-based computational measurement of personality (Koutsoumpis et al., 2022; Moreno et al., 2021) lies at the heart of language-based personality research (Boyd & Pennebaker, 2017). This is specifically important for the assessment of personality during personnel selection and assessment (Holtrop et al., 2022), targeting the person-environment fit (P-E fit; Kristof-Brown et al., 2005). Especially, the supplementary fit between an organisation and a person verifies that values, goals, culture, and personality do match with each other (Kristof, 1996).

Vocabulary-based research can generally be differentiated between the open- and closed-vocabulary approaches (Eichstaedt et al., 2021). The open vocabulary approach applies text mining techniques based on machine learning (Kern et al., 2014) and generates a computational word weighting (Qin et al., 2020). Results often suffer from poor construct validity in terms of substantial, structural, and external validity (Bleidorn & Hopwood, 2019), as well as psychometric issues (Tay et al., 2020). In the closed vocabulary approach, words are preselected by human raters, and their linkages to specific language categories or psycho-social constructs are defined (e.g. culture: Chapman et al., 2018; Pandey & Pandey, 2019; Ponizovskiy et al., 2020). Thus, these dictionaries show better construct validity but may suffer reliability issues, as even trained judges often do not agree, or judges get influenced by the text content they read over time (Tausczik & Pennebaker, 2010). These issues can be solved by increasing the number of raters, calculating inter-rater agreement scores, and defining an exclusion strategy.

The most common closed vocabulary (Pennebaker & King, 1999) is part of the proprietary software Linguistic Inquiry and Word Count (LIWC) (Pennebaker, 2025). It covers more than 10,000 words spread across more than one hundred categories (Boyd et al., 2022); some of them are language-specific (e.g. verbs or nouns), while others are construct-specific (e.g. positive or negative emotions). Interestingly, none of the categories explicitly addresses personality traits. As for culture (Ponizovskiy et al., 2020) and gender (Gaucher et al., 2011), specific closed vocabularies for the multidimensional construct of personality have been developed (e.g. Ostendorf, 1990, 1994). In contrast to other personality dictionaries (e.g. Holtrop et al., 2022), the one from Ostendorf does not only link words to one specific category, but it also offers a word weighting by subject matter experts (SME) (Ostendorf, 1990, 1994), highlighting the simultaneous importance of words for several categories at once. Interestingly, most research in text-based computational measurement of personality used the LIWC, but no empirical study could be identified that has applied the SME weighting approach.

In addition to the methodological question of whether to use (un-)weighted closed vocabularies in text-based personality research, the type of data analysed is also important. One of the first studies to implement LIWC analysed self-reported student essays (Pennebaker & King, 1999). Further studies addressed Facebook status messages or posts (Celli, 2013; Park et al., 2014) and, more generally, digital footprints on social media (e.g. Azucar et al., 2018). The major critique of these data types is that social media users have a specific personality structure (de Zuniga et al., 2017; Freberg et al., 2010); thus, the data is biased and does not vary significantly. Due to the call for more general corpora, recent research has focused on recruiting-related texts, such as answers to interview questions (Jayaratne & Jayatilleke, 2020), résumés, job descriptions (Pudasaini et al., 2022), job advertisements (Tokarz, 2019), and asynchronous video interviews (Holtrop et al., 2022).

Thus, the general purpose of this study is to test the application of a weighted closed vocabulary approach as introduced by Ostendorf (1990, 1994) on a large corpus of online job advertisements (OJA) and, thereby, to measure the personality being searched for in the job market. As the personality communicated in the OJAs is the organisation’ contribution in the search for the supplementary fit, we will briefly discuss the P-E fit theory, provide a short summary of the current research on text-based computational measurements of personality, and thereby link our research in more detail to P-E fit theory. During the discussion, we developed our research questions that will guide our empirical analysis.

Theory

First, we briefly discuss our research approach on the background theory of P-E fit to derive the use of OJAs as the text corpus. Second, we introduce the research about the Big 5 and their text-based computational measurement with NLP. In both sections we outline the resulting research questions.

From Person-Environment to a Supplementary Perceived A-J Fit

The P-E fit is defined as “the compatibility between an individual and a work environment” (Kristof-Brown et al., 2005, p. 281). It is described by a “complex and multidimensional concept” (Sekiguchi, 2004, p. 181) consisting of four facets: complementary and supplementary, as well as perceived and actual fit. Focusing on the essential conceptual aspects necessary for vocabulary-based research in text-based computational measurement of personality (Koutsoumpis et al., 2022; Moreno et al., 2021), the supplementary perceived fit plays a core role. Supplementary means that a person’s characteristic, such as personality, is similar to environmental requirements (Sekiguchi, 2004). The perceived fit indicates that the person has evaluated the fit on its own (Sekiguchi, 2004) without objective measurement.

One step closer from the concept of the supplementary perceived P-E fit to organisational research is when narrowing the environment to organisations (Kristof-Brown et al., 2023). Research mostly addresses the two aspects of either person-organisation fit (P-O) or person-job fit (P-J), while there are further aspects, such as Person-Group, Person-Supervisor, or Person-Industry fit (Vleugels et al., 2023). In correspondence with the P-E fit concept (Kristof, 1996), P-O fit addresses broader organisational attributes, such as organisational culture (Herold & Roedenbeck, 2025) or corporate image (Wei et al., 2016), and the P-J fit targets the different aspects of an individual, such as goals, needs, knowledge, skills, or abilities, and personality (Edwards, 1991; Werbel & Gilliland, 1999). Interestingly, personality can also be studied from a P-O fit perspective, addressing “the extent to which the individual’s personality is similar to the personality of (successful or prototypical) members” (Iyer et al., 2020, p. 144).

Further narrowing down the concept of the supplementary perceived P-J fit is the differentiation of a person into the organisational employee and the external applicant (Bretz et al., 1993; Kristof-Brown, 2000). While the organisational employee can acquire information about a job from various sources, these are limited for an external applicant. Relying on signalling theory (Guest et al., 2021), one of the applicants’ core sources are OJAs (Herold & Roedenbeck, 2025), which they use to draw assumptions about personality requirements (Koçak et al., 2022; Newman & Lyon, 2009) and their P-A fit. Ehrhart (2006) reported a significant interaction between personality and job characteristic beliefs when predicting job fit. And further evidence has shown a positive relationship between OJAs’ wording, including descriptors related to personality, and applicant personality levels (Stevens & Szmerekovsky, 2010). But the question is whether the information provided offers enough clues for the applicants to draw valid conclusions about personality requirements. Latest research on higher education OJAs showed that personality is not necessarily referenced even if employers are in search of particular personality traits (Antonie et al., 2024).

Thus, we can conclude that our research is rooted in the concept of the supplementary perceived A-J fit, focusing on OJAs as the applicants’ primary source of information, which they use to draw assumptions about personality requirements. Addressing the central question of whether the information provided in OJAs offers enough clues for the applicants, while enlarging the focus from higher education to the full labour market, we want to follow the methodological idea of a clustering approach of OJAs (Pejic-Bach et al., 2020). Thus, we derive the following detailed empirical research questions (ERQ) related to the central question of our study: Can a personality profile be identified across specific OJAs for a given job (ERQ1)? Can different personality profiles be identified across all OJAs, independent of the job (ERQ2)?

Big 5 and Its Text-Based Computational Measurement

There are different views on the definition of personality, such as Rauch and Frese (2007, p. 355), who define personality traits as “dispositions to exhibit a certain kind of response across various situations.”Wood (2012) characterises personality as the unique thoughts and actions that differentiate individuals. Van Lieshout (2000) defined traits as unchangeable over time and in different situations. McCrae and Costa (1994) wrote that traits reflect not only a person’s characteristics but also their selves; thus, personality traits are considered the primary source of interpersonal differentiation (O. P. John & Gosling, 2000).

The five-factor model is one of the most widely used frameworks for personality traits and is based on a lexical approach (Allport & Odbert, 1936). Five dimensions were identified: openness (O), conscientiousness (C), extraversion (E), agreeableness (A), and neuroticism (N). Openness refers to people who are creative and unconventional (Smith & Canger, 2004) or like new ideas and show intellectual curiosity (Grehan et al., 2011). Conscientiousness describes behavioural regularities, the efficiency of action (Hofmann & Jones, 2005), and the willingness to adhere to organisational rules or group norms (Smithikrai, 2008). Extraversion describes the ease of forming new relationships (Mushonga & Torrance, 2008) or the individual’s preference for seeking excitement and stimulation (Zhao & Seibert, 2006). Agreeableness refers to how positive interpersonal relations are maintained (Jensen-Campbell & Graziano, 2001) or how interpersonal interactions are designed in areas such as trust, altruism, or compliance (Patrick, 2011). Neuroticism describes people who are likely to experience stress and anxiety (Judge & Ilies, 2002), negative affect, and emotional self-absorption (Renn et al., 2011). Each dimension of the Big Five framework represents a summary of numerous varied and specified personality traits (John et al., 2008). McCrae and Costa (1994) argued that all dimensions of the Big Five framework are stable and do not change significantly over time for adults.

According to the latest meta-analysis (Koutsoumpis et al., 2022; Moreno et al., 2021), the text-based computational measurement of personality has been primarily driven by the LIWC methodology as a lexical analysis based on tokenisation introduced by Pennebaker and King (1999). More commonly, this approach falls under natural language processing (NLP), which includes text pre-processing, lexical analysis, syntactic parsing, and semantic analysis (Indurkhya & Damerau, 2010), whereas low-level tasks include sentence detection, tokenisation, and parsing (Nadkarni et al., 2011). LIWC uses a closed vocabulary with approximately 12,000 words spread across 110 categories (Boyd et al., 2022). The categories cover verbs, nouns, past or future, or several content-related domains, such as power, social, positive, or negative emotions (Boyd et al., 2022). It has been developed based on the judgement of human raters, where words were added or kept in or deleted from a category if 2 out of 3 raters agreed (Tausczik & Pennebaker, 2010). The dictionary is a proprietary part of the LIWC software (Pennebaker, 2025) and does not contain specific categories for personality.

Interestingly, LIWC has been used frequently to assess personality in texts (Azucar et al., 2018; Farnadi et al., 2016; Iatan, 2017; Mairesse et al., 2007; Park et al., 2014; Tadesse et al., 2018; Tay et al., 2020; Yarkoni, 2010). However, no high correlations between the LIWC analysis of the stream of consciousness student essays and their corresponding personalities derived from classical self-assessment were identified (Pennebaker & King, 1999). Over the last decade, several studies showed significant statistical interrelations between the dimensions of personality and the categories as measured by LIWC (Hickman et al., 2024). Contrary evidence is reported that the relationship between LIWC and personality is not generalisable in cross-validation studies (Martínez-Huertas et al., 2022). Thus, some researchers started to develop a specific closed vocabulary for personality research based on HEXACO, exploring spoken and written text in (asynchronous) job interviews (Holtrop et al., 2022). Others suggested following an open vocabulary approach implementing NLP and deep learning for text classification and matching (Cutler & Condon, 2023), where some of these open vocabulary approaches challenge their model against the LIWC (Aslam et al., 2025).

On the one hand, using a specific and closed vocabulary for measuring personality seems more preferable than applying the multiple-construct and multi-dimensional LIWC. On the other hand, implementing deep learning to extract word weightings instead of simple word counts also seems more preferable, as predefined loading of individual words on specific dimensions is much better captured when word weighting is used in advance (Iwai et al., 2020). While the latest closed vocabulary repeats the LIWC methodology of grouping words into categories (Holtrop et al., 2022), one of the first attempts to join the positive from both perspectives, especially in the field of personality research, stems from Ostendorf (1990, 1994). He developed trait descriptors (Norman, 1967) and introduced 823 words as a closed vocabulary for the Big Five domains of personality. He additionally offered SME weightings for each word in each domain based on inter-rater agreement (n = 16). Unfortunately, his vocabulary has only been published in a German textbook and has not yet found its way into international scholarly debate. Thus, according to our knowledge, there are no studies that applied this dictionary to any type of data. More recent research in the field introduced 526 words with individual loading as a closed vocabulary; however, only word and weight samples have been published in Japanese (Iwai et al., 2020).

Thus, to answer the two empirical research questions above, there is a methodological research question to be answered first. As there are no studies at hand that try to make use of the weighted closed vocabular, the question is whether the weighted closed vocabulary approach of Ostendorf (1994) can be used to identify personality in OJAs (MRQ).

Method

In the following section, we describe the dataset, the details of Ostendorf’s (1994) vocabulary, and the procedure for applying our weighted closed vocabulary-based text analysis.

Dataset and Descriptive Stats

When analysing online job advertisements (OJA), a large dataset is required if done quantitatively (Chung & Chen, 2019; Pejic-Bach et al., 2020). While one source could be the webpages of different companies, we followed the signalling approach of OJAs and scraped data from an online job portal (Saurkar et al., 2018). Online job portals often limit the design of OJAs to a small set of data fields that can be automatically extracted and stored in a structured dataset (Vargiu & Urru, 2012). Scraping OJAs is quite common in the field (De Mauro et al., 2018), although we focused on the German job market. Our web-scraping procedure was protected under German copyright law and was in line with the German Consortium for Social, Behavioural, Educational, and Economic Sciences.

We scraped 150,956 OJAs (n1.a) via generic categories of the online job portal, consisting of 112,845 unique job titles (n1.b) from January to May 2022 from a large German job portal that formed our sample. The frequency per unique job title ranged from 1 (min) to 443 (max). The median was 1, and the mean was 1.3. All OJAs consisted of four separate sections: the company’s description, job task, expected personal profile, and offer. In relation to our research questions, we focused solely on our analysis of the text within the section “expected personal profile” of the OJAs.

As both the mean and median across the job titles were close to 1, numerous unique job titles made it difficult to find answers to our empirical research questions about profile comparison and clustering. Thus, we decided to generate a sub-sample of job functions based on Porter’s generic value chain (M. E. Porter, 1998). It does fit better to an applicant’s perspective in search of a generic job in an organisation, than the US American O*Net or the European ESCO, which are focused on occupation frameworks, skills, or competences unknown to the applicant. We picked 9 job functions (n2.b) instead of the job titles. As supportive functions, we made a keyword search for procurement, human resources (HR), information technology (IT), research and development (R&D), or finance. For primary functions, we made a keyword search for production, logistics, marketing, and sales. A summary of the keyword searches for generating the job functions is provided in Table A1. Table 1 shows the descriptive statistics for our job functions with the corresponding number of OJAs identified in the final subsample (n2.a).

Table 1.

Descriptive Statistics for the Sub-Sample (n2.a*) of the Selected OJAs, Based on the Word Count (WC) of the Section “Expected Personal Profile.”

Value chain	Job function	Count	Min	Max	Mean	Median	Std	Skew	Kurt.
Supportive functions	Procurement	426	7	158	57.60	53	22.15	1.12	2.02
	HR (referee)	440	18	241	64.34	60	24.97	1.98	8.72
	HR (clerk)	370	10	297	57.41	52.5	26.15	3.06	20.76
	IT	420	18	177	66.06	62	26.01	1.09	2.14
	R&D	219	13	220	70.20	62	31.03	1.17	2.32
	Finance	334	9	156	55.18	53	21.43	0.96	1.63
Primary functions	Production	138	8	137	41.38	37	21.78	1.33	2.45
	Logistic	190	7	115	44.86	40	18.77	0.80	0.67
	Marketing	328	19	276	78.86	72	33.03	1.89	6.63
	Sales	374	27	75	73.45	73	3.61	−9.49	102.29

*

Overall stats for n2.a are sum (3,239), mean (323.9), and standard deviation (106.14) across the job functions.

There are three specialties in our selected subsample. First, one prior study regarding vocabulary-based text analysis suggested that context-specific text length may be taken into account, where only texts larger than the average word count of a given corpus of text data should be included (Eichstaedt et al., 2021). As there is no further evidence that this kind of pre-selection is necessary, especially when using predefined vocabularies, we did not further reduce our dataset. Second, for the HR function, 810 OJAs have been identified; that is more than twice the number of procurement or IT. Thus, we carefully checked the OJAs identified and realised that two different job levels were identified via our keyword search. On the one hand, a clerk with mostly a commercial education/apprenticeship, and on the other hand, an officer/referee with mostly a university degree. Therefore, we split the HR function into two job functions. Third, the job function of sales shows a very low standard deviation of words in comparison to the other job functions. This is due to numerous OJAs stemming from one large sales company, which lowers the deviation and increases the skewness and kurtosis. Results from that job function have to be interpreted carefully.

Weighted Vocabulary and its Descriptive Stats

As introduced in the theoretical section, Ostendorf (1994) developed a weighted vocabulary of 823 adjectives based on prior studies that are used as descriptors of the Big Five dimensions of openness (referred to as culture, CU), conscientiousness (CO), extraversion (he named it surgency, SU), agreeableness (AG), and neuroticism (he used the opposition emotional stability, ES). Sixteen subject matter experts (SMEs) rated all adjectives on a 7-point Likert scale ranging from −3 (good descriptor for a negative pole) to +3 (good descriptor for a positive pole) across all dimensions, where 0 means that the word is neither a descriptor for a positive nor a negative pole and thus has no impact on the dimension. In Table 2, we provide the results from Ostendorf (1994) according to the minimum and maximum average word ratings. The expected mean average of 0 [−0.08, 0.17] and standard deviation average of 1 [0.91, 1.25] representing the indicators of a standard normal distribution were closely achieved per dimension. The correlation analysis among the raters showed high agreement with a high intraclass correlation (ICC1, [0.97, 0.98]), where the Eigenvalue-analysis per dimension revealed only 1 Eigenvalue (λ) or main factor within the raters’ feedback. Thus, Ostendorf (1994) assumed high validity and provided the mean and standard deviation for each word per dimension based on the 16 raters. The quantile report shows that all dimensions have words either close to a full loading [2.94, 3.0] or to a contrary loading [−3.0, −2.81]; thus, the vocabulary seems to cover all dimensions adequately.

Table 2.

Descriptive Statistics of the SME Ratings for the Weighted Vocabulary Across All Five Dimensions for the Means Provided.

Big 5 dimensions	Adj. num. a	Min	Max	Avg. mean	Avg. std	ICC	λ
Culture (CU) ∼ openness	767	−2.94	2.94	0.17	1.09	0.97	1
Conscientiousness (CO)	764	−2.81	3.00	0.10	1.07	0.97	1
Surgency (SU) ∼ extraversion	745	−3.00	3.00	0.14	1.04	0.97	1
Agreeableness (AG)	739	−2.88	2.94	−0.08	1.25	0.98	1
Emotional stability (ES) ∼ neuroticism	737	−2.94	2.94	−0.03	0.91	0.97	1

a

Only the adjusted number of adjectives per dimension is provided which included those adjectives with a mean and standard deviation unequal to 0. ICC = intra-class correlation; λ = Eigenvalues equal or larger than 1 per dimension.

There are two major critical aspects regarding the vocabulary of Ostendorf (1994). First, there is no weight threshold around 0 applied (e.g. <0.1) to exclude “lightweight” or “neutral” words from individual dimensions. As we did not want to adjust the SME rating but to apply their word weighting, we accepted these “lightweights” and reported the adjusted number of adjectives per dimension with a mean and standard deviation unequal to zero. Thereby, we are also able to show how many words can be finally used when a word-weighting is applied (as a multiplication by 0 would result in a score of 0 as well). Second, applying positive (>0) and negative (<0) word weights within the same dimension, they will outweigh themselves (e.g. one word in agreeableness counts −2.88, and another +2.88; the dimensional weight would be 0). Again, we did not want to adjust the initial idea of Ostendorf (1994) that language contains contradicting poles of a single dimension and thus reduces or neglects the impact of the words. Both aspects emerge from the use of a weighted word count approach and are seen as advancements in research.

Procedure of Analysis With a Weighted Vocabulary

We applied this weighted vocabulary to our subsample in the following steps using a Python/Pandas/SkLearn/NLTK environment:

First, we preprocessed the data as suggested in several NLP papers (e.g. Evangelopoulos et al., 2012) by turning the job profile to lowercase and excluding punctuation. We skipped the stop-word reduction because we applied a given vocabulary and did not analyse all words in the corpus, as is usually conducted in uninformed data mining (e.g. Cutler & Condon, 2023). The last step of pre-processing is stemming or lemmatisation, where the endings of words have to be cut off based on algorithms or machine learning (Balakrishnan & Lloyd-Yemoh, 2014). While using vocabularies with predetermined wildcards “*” (e.g. Gaucher et al., 2011; Pandey & Pandey, 2019), stemming or lemmatisation is not necessary. However, the vocabulary from Ostendorf provided full words without wildcards, so endings had to be cut off in the dataset as well as in the vocabulary itself. Although research suggests that lemmatisation outperforms stemming for the English language in principle (Balakrishnan & Lloyd-Yemoh, 2014), the German language is more difficult to handle in NLP. At the date of our analysis, lemmatisation was not available in the most common NLP toolkit for Python, namely NLTK (Wartena, 2019), and the German stemmer of the NLTK package offered better results than the German lemmatisation of the alternative Spacy package. This is due to the training corpora of the latter, which is rooted in news and Wikipedia data. Therefore, we decided to remain with the German stemmer.

Second, we conducted an unweighted word count similar to the LIWC approach (Pennebaker & King, 1999) but counted the stemmed words of the vocabulary used in a specific stemmed OJA. Thereby, we were able to see whether different words per dimension were identified (differences across the dimensions) and to focus on frequencies of words identified per job function. For example, if the word “abenteuerlust” was found in an OJA, it counts 1 for each dimension in the case of the unweighted approach, but it has different loadings (mean/standard deviation) on the weighted approach (SU: 2.81/0.75; AG: 0.06/0.25; CO: −0.25/0.58; ES: 0.38/0.72; CU: 0.06/0.25). We report descriptive statistics on the frequency of words identified across the OJAs grouped per job function as a percentage, show the mean, standard deviation, median, and minima/maxima. Finally, we apply a mean/median split to dichotomise our dimensions of personality, as it was conducted with one of the core datasets for studying personality with the use of LIWC (Celli et al., 2013; Pennebaker & King, 1999), and to uncover whether specific job functions ask for a high or low profile in relation to others. While one may argue this is an oversimplification with arbitrariness instead of reality (Schütz et al., 2013), others argue that it is the most common approach of categorisation (Norlander et al., 2002).

Third, we applied the weighted word count approach by multiplying the frequency of individual words in one dimension for one OJA by the word-weight mean divided by its standard deviation, as found in Ostendorf (1994). As an example, the word “abenteuerlust” was used once in an OJA; thus, the frequency was multiplied by the 16 raters’ average word-weight of 2.81 and divided by its standard deviation of 0.75, with a final value of 3.74 for the dimension of extraversion (SU). In case it was found twice in one OJA, the result would be (2 × 2.81/0.75) = 7.49. If an additional word like “ansprechbar” (mean: 1.75; SD: 1.00) was found, its weighting (1 × 1.75/1.0 = 1.75) would be added to the dimensional result (7.49 + 1.75 = 9.24). While this looks similar to term-frequency inverse-document frequency (TF-IDF) weighting, this approach is based on raters’ word weights but not the OJA-wide document frequency; thus, it is an SME-based rather than a machine-learned weighting. We report the mean statistics across the weighted word frequencies per dimension across job functions and then apply a mean split per row and per column. Therefore, we cross-checked the prior outcomes of the unweighted approach and were able to answer the methodological research question.

Fourth, we applied a min-max scaler from SkLearn to the results of the weighted word count after removing OJAs where no words could be identified. Thereafter, scaling the results of the dataset into a predefined point range was necessary, as different word counts and weights across the dimensions resulted in different maxima and minima. For example, across all OJAs within the dimension of extraversion, the weighted word range lies between −2.89 and +6.45. The SkLearn scaler was used to translate each dimension individually, but across all OJAs, into a simple range [0; 1]. With these reduced and scaled data, we conducted a profile analysis of the multivariate data using the R/profileR package (Bulut & Desjardins, 2018). With the profile analysis, parallelism, equality, and flatness could be tested between two grouped data for job profiles, where parallelism was given when the profiles were congruent and equality was given if the values of the profile were at the same level. Flatness was a cross-check of a unique but exclusive case in which the profile showed all zeros. H0 tests are reported to test profile parallelism with Wilks and Pillai tests, equality, and flatness (Bulut & Desjardins, 2018). Significant results would reject H0, where, in our case, the combination of parallelism and equality (twice non-significant results) and non-flatness (significant results) was the target for answering empirical research question 1.

Fifth, we performed a cluster analysis with the SkLearn and ClustersFeatures packages in Python to determine whether groups of job profiles could be extracted by matching with the job function-based value chain sorting or profiling results of step 4. The analysis is conducted only with those OJAs containing at least information about one personality trait. Several methods can be used to achieve a 2D distribution for clustering, such as PCA, NMF, ICA, or SVD. For more complex structures, the t-distributed stochastic neighbour embedding (t-SNE) has been suggested (van der Maaten & Hinton, 2008), so we only applied this method. Thereafter, we performed k-means clustering in scikit-learn as it offers a linear complexity and works well on isotropic data (Rokach & Maimon, 2005), to identify the optimal number of clusters with the elbow method (noise as the sum of the squared distances). We expected non-overlapping spherical subsets of clusters, as we assumed clearly differentiated personality types across the jobs. We then predicted the corresponding cluster per OJA. Thereafter, we calculated silhouette scores applying ClustersFeatures for the individual clusters as well as the full clustered dataset. Finally, we provided cluster averages across the dimensions and calculated the mean split to differentiate between high and low dimension loadings. Therefore, we were able to answer empirical research question 2.

Results

At first, we aim to answer the methodological research question (MRQ) regarding the applicability of Ostendorf’s (1994) weighted closed-vocabulary approach. Table 3 shows that across all job functions, 141 stemmed words of the vocabulary could be identified, which is equal to 17.1%. Having analysed the distribution of the stemmed words across job functions, the results could be grouped into those covering no words of the vocabulary, one to three words, four to six words, seven to nine words, or more than nine words. The majority of OJAs per job function used one to three words, ranging from 52.7% to 66.3% (mean: 60.1%), excluding sales (0.8%). The second rank was covered by four to six words, ranging from 13.8% to 30.5% (mean: 22.6%), excluding sales (99.2%).

Table 3.

Descriptives Statistics for the Unweighted Approach of n2.a = 3,239 OJAs, Same Across All Dimensions (SU/AG/CO/ES/CU).

Value chain	Job function	Count	Num. of words and % of coverage					Med	Mean	Std	Min	Max
			0	1–3	4–6	7–9	>9
Supportive functions	Procurement	426	15.5	59.2	22.5	2.8	0.0	2	2.46	1.81	0	8
	HR (referee)	440	7.7	59.1	27.5	5.2	0.5	3 a	2.93 b	1.91	0	10
	HR (clerk)	370	5.7	61.6	28.1	4.3	0.3	3 a	2.97 b	1.81	0	10
	IT	420	18.8	63.6	14.5	3.1	0.0	2	2.12	1.80	0	9
	R&D	219	21.9	58.4	16.0	3.7	0.0	2	2.15	1.95	0	9
	Finance	334	11.7	53.9	30.5	3.9	0.0	3 a	2.82 b	1.95	0	8
Primary functions	Production	138	19.6	65.9	13.8	0.7	0.0	1	1.76	1.54	0	7
	Logistic	190	11.1	66.3	20.5	2.1	0.0	2	2.52	1.74	0	9
	Marketing	328	11.9	52.7	30.2	4.3	0.9	3 a	1.74	2.06	0	12
	Sales	374	0.0	0.8	99.2	0.0	0.0	4 a	3.98 b	0.25	1	5

a

Median larger as median split (1).

b

Mean larger as mean split (0.94).

When simply taking the unweighted frequency of words into account, each word identified showed the same impact on each dimension of the Big Five. Thus, the mean, standard deviation, and median did not vary in the unweighted approach, and the results only showed the frequency of the personality descriptors used. A closer look at these results across the job functions in Table 3 shows that the suggested median split across all medians (2.5) resulted in five job profiles in which personality played a greater role: HR (referee and clerk), finance, marketing, and sales. The mean split across the means (2.65) also supported these results.

After applying the word weights of the vocabulary from Ostendorf, the mean statistics changed per dimension, as explained in the methodological section. As shown in Table 4, regarding the individual dimensions of personality grouped by job functions, the mean split across the row means (3.22) revealed five job functions above the level. Interestingly, they were not the same as in the case of the unweighted approach. While HR (referee and clerk), finance, and sales were equally above the mean, as in the unweighted approach, marketing showed a lower level and logistics a higher level than the mean. We then applied the mean split across each dimension individually (per column mean), and it could be concluded that only HR (referee and clerk) showed an importance for all dimensions, followed by sales covering all but one (conscientiousness) dimension of personality. Focusing on the highest (Logistics, CO) and the lowest (Production, CU) reported means (5.97; 0.99) and standard deviations (6.15; 1.25) revealed that a high standard deviation was observable.

Table 4.

Descriptive Statistics for the Weighted Approach of n2.a = 3,239 OJAs, Individual Means per Dimension.

Value chain	Job function	Count	SU	AG	CO	ES	CU	Row means
			Mean/SD	Mean/SD	Mean/SD	Mean/SD	Mean/SD
Supportive functions	Procurement	426	2.46 a /2.39	3.28/3.30	3.57/4.21	3.69 a /3.19	2.05 a /2.00	3.01
	HR (referee)	440	2.84 a /2.35	5.02 a /4.45	4.52 a /4.93	4.03 a /3.23	2.16 a /1.92	3.72 b
	HR (clerk)	370	2.70 a /2.70	4.78 a /4.21	5.93 a /6.00	4.08 a /3.13	2.08 a /1.68	3.91 b
	IT	420	1.95/2.19	2.63/3.21	3.09/4.09	3.04/3.04	1.80/2.04	2.50
	R&D	219	1.93/2.24	2.85/3.84	2.36/3.19	2.79/3.25	1.90/2.27	2.37
	Finance	334	2.21/2.29	3.68/3.24	5.55 a /5.75	3.89 a /3.36	2.21 a /1.96	3.51 b
Primary functions	Production	138	0.61/1.16	1.70/2.44	5.89 a /5.97	2.01/1.95	0.99/1.25	2.24
	Logistic	190	1.96/2.26	3.71/3.57	5.97 a /6.15	3.76 a /3.12	1.28/1.53	3.34 b
	Marketing (Manager)	328	2.64 a /2.52	3.48/3.91	2.91/3.34	3.44/3.54	3.25 a /3.08	3.14
	Sales	374	3.88 a /0.29	8.66 a /0.66	3.04/0.21	4.68 a /0.47	2.25 a /0.24	4.50 b
	Column mean		2.32	3.98	4.28	3.54	2.00	3.22

a

Mean split per column by the column means with higher values marked.

b

Mean split (3.22) across the row means with higher values marked.

We now focus on the empirical research questions. For answering the first one (ERQ1), whether a similar personality structure can be identified across specific OJAs for a given job, we identified significant similarities across the profiles for each job function by applying profile analysis. As seen in Table 3, there was a percentage of OJAs per job function that did not cover any of the words from the Ostendorf vocabulary (0 words). Thus, for profile analysis, only OJAs have been included that made use of at least one word in the vocabulary (see Table 5, n3). Table 5 shows the results of H0 tests of profile parallelism on equal levels and curve flatness in an aggregated form; in the Table A2, all values are included. As described above, the targeted combination showed non-significant results for parallelism and equality to accept the H0 hypothesis but significant results for flatness to reject H0 (represented as -|-,-,*). 4 general result types across the full dataset have been uncovered: (a) our best case of parallelism and equality but no flatness (-|-,-,*; 4x), (b) a rejection of parallelism, equality, and flatness (*|*,*,*; 19x), (c) a rejection of parallelism and flatness but equality of levels (*|*,-,*; 21x), and (d) curve flatness at different levels (*|*,*,-; 1x). There were only four cases in which our most expected type “a” appeared, which positively answered our first empirical research question. The first case was our previously discussed linkage between the HR clerk and referee job profile, and the second was a group formed of IT, procurement, and R&D with interlinkages.

Table 5.

Profile Analysis Results (n3 = 2,865).

Value chain	Job function	Count	Procurement	HR (referee)	HR (clerk)	IT	R&D	Finance	Production	Logistic	Marketing
Supportive functions	Procurement	360	–	–	–	–	–	–	–	–	–
	HR (referee)	406	(*|*,-,*)	–	–	–	–	–	–	–	–
	HR (clerk)	349	(*|*,-,*)	(-|-,-,*)	–	–	–	–	–	–	–
	IT	341	(-|-,-,*)	(*|*,*,*)	(*|*,*,*)	–	–	–	–	–	–
	R&D	171	(-|-,-,*)	(*|*,*,*)	(*|*,*,*)	(-|-,-,*)	–	–	–	–	–
	Finance	295	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,*,*)	(*|*,-,*)	–	–	–	–
Primary functions	Production	111	(*|*,*,*)	(*|*,*,*)	(*|*,*,*)	(*|*,-,*)	(*|*,-,*)	(*|*,*,*)	–	–	–
	Logistic	169	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,*,*)	–	–
	Marketing	289	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,-,*)	(*|*,*,*)	(*|*,-,*)	–
	Sales	374	(*|*,*,*)	(*|*,*,*)	(*|*,*,*)	(*|*,*,*)	(*|*,*,-)	(*|*,*,*)	(*|*,*,*)	(*|*,*,*)	(*|*,*,*)

Note. All tests were conducted on p level <.0001. Targeted test results are (-|-,-,*) with non-significant results for parallelism and equality to accept the H0 hypothesis, but significant results for flatness to reject H0. Other results are (*|*,*,*) representing a rejection of all tests regarding parallelism, equality, and flatness, (*|*,-,*) representing a rejection of parallelism and flatness but equality of levels, and (*|*,*,-) representing a rejection of parallelism and equality but a curve flatness.

Finally, we answer the second empirical research question (ERQ2), whether groups of personality structure (similar within a group and different between groups) can be identified across all OJAs independent of the job. Therefore, we also relied on the reduced dataset for profile analysis (n3), covering more than 88% of the OJAs sub-sample. First, dimensionality reduction was conducted using t-SNE across the five dimensions. Thereafter, we estimated the number of clusters using the k-means algorithm and the corresponding elbow for inertia in the t-SNE distance matrix. As shown in Figure 1a, we selected five clusters as the best choice. We then predicted the clusters using the k-means algorithm on all OJAs, with the results shown in Figure 1b, and calculated the average profile within the clusters, as shown in Figure 1c. To evaluate the cluster quality, we calculated the silhouette score with a total average of 0.37 for the full dataset and separate cluster silhouette scores of 0.34 (cluster 0), 0.45 (cluster 1), 0.28 (cluster 2), 0.37 (cluster 3), and 0.37 (cluster 4).

Figure 1.

(a) K-means elbow diagram of inertia, (b) K-means clustering with 5 clusters, (c) cluster profiles of personality.

While in the profile analysis we observed similarities between profiles on their dimension averages, the cluster analysis provided five generic personality profiles inherent in all OJAs across different job functions. We again applied a mean split (0.23) across scales and clusters, which revealed that all clusters showed high surgency/extraversion (SU) and low conscientiousness (CO), two clusters showed high agreeableness (AG; 0 and 3) and high emotional stability (ES; 0 and 1), and three clusters showed high culture/openness (CU; 0, 1, and 2). As shown in Table 6, when analysing the group members, there was only one job function clearly belonging to a specific cluster (with minor exceptions): sales, with 98%. While most of the other job functions can be linked to one cluster based on their highest percentage, only finance is bimodal, with twice the rate of 25%.

Table 6.

Percentage of OJAs for a Specific Job Function per cluster.

Value chain	Job function	n3	Cluster (%)
			0	1	2	3	4
Supportive functions	Procurement	360	16.39	30.56	20.00	13.33	29.72
	HR (referee)	406	28.6	24.88	14.29	11.58	20.69
	HR (clerk)	349	28.37	18.34	21.49	12.61	19.20
	IT	341	14.37	24.63	17.01	22.29	21.70
	R&D	171	15.20	29.24	15.20	24.56	15.79
	Finance	295	24.75	21.36	25.08	13.22	15.59
Primary functions	Production	111	18.92	5.41	44.14	14.41	17.12
	Logistic	169	18.34	13.02	30.18	16.57	21.89
	Marketing (Manager)	289	15.92	38.75	16.96	13.84	14.53
	Sales	374	0.27	1.60	0.00	98.13	0.00
	Total cluster size	2,865	521	618	512	747	467

Note. Highest percentage per job function in bold. Total amount of OJAs reduced to 2,865 excluding those with flatness.

Thus, we can conclude that cluster analysis fully supported our second empirical question because five groups of personality structure (similar within a group and different between groups) could be identified across all OJAs independent of the job function. Thus, human resource managers seem to look for only five different profiles.

Discussion

Being rooted in the concept of the supplementary perceived A-J fit research, we addressed the central question of whether the information provided in OJAs offers enough clues for the applicants to draw conclusions about personality. Therefore, we applied an up-to-date, untested closed-vocabulary on personality to identify whether a personality profile can be identified across specific OJAs for a given job and different personality profiles can be identified across all OJAs with a profiling and clustering approach.

Methodological and Theoretical Contributions

We discuss 4 perspectives within our methodological contribution, all drawn from our theoretical focus on the supplementary perceived A-J fit in the context of the P-E fit theory: (a) the central methodological impact of our research findings, (b) the supportive results in line with current evidence, (c) the contradictory results against current evidence, and (d) the new insights from our cluster analysis.

Starting with the central methodological impact, we can state that the idea from Ostendorf (1990, 1994) using a specific weighted closed vocabulary describing personality was very useful in analysing our corpus of OJAs. As shown in Table 4, different loadings on the individual dimensions of personality can be identified across different job functions for the weighted approach. Even the unweighted approach of the specific vocabulary, as shown in Table 3, resulted in different numbers of words used across different job functions when analysing OJAs. Thus, our method is a better approach for personality research in the context of the A-J fit perspective (Sekiguchi, 2004) based on OJA’s in contrast to the LIWC approach (Pennebaker & King, 1999). In the latter, the linkages between LIWC factors and personality dimensions were proven only via significant but weak factor scores (ES: −0.16 to 0.13, SU: −0.14 to 0.15, CU: −0.16 to 0.11, AG: −0.15 to 0.07, CO: −0.15 to 0.07). From the perspective of the broader P-E fit theory (Kristof-Brown et al., 2005) linked to signalling theory (Guest et al., 2021), we suppose that organisational culture (Herold & Roedenbeck, 2025) or corporate image (Wei et al., 2016) about gender and diversity are also communicated via OJAs. Thus, we recommend further development of specific closed dictionaries for OJA-based research, such as for culture (Herold & Roedenbeck, 2025; Pandey & Pandey, 2019; Ponizovskiy et al., 2020), gender (Gaucher et al., 2011; Hentschel et al., 2021), diversity (Konnikov et al., 2023; Remedios et al., 2010), and sentiment (Young & Soroka, 2012), by weighting the words included in the dimensions of the corresponding theoretical frameworks. The procedure suggested by Ostendorf (1994), who used a set of 16 SMEs to rate all words used, could be applied. Alternatively, 3 SME may be the minimum requirement (Tausczik & Pennebaker, 2010) as long as they follow Ostendorf’s weighting in the range of −3 to +3. Finally, weights can be achieved by calculating the mean and standard deviation. Besides the Big Five personality research, new domains for text analysis could be the core self-evaluation traits (Bono & Judge, 2003) or the dark triad (Paulhus & Williams, 2002) using different corpora than OJA’s.

Focusing on supportive results in relation to prior evidence in the context of the A-J fit perspective, we draw back on Antonie et al. (2024). They found only partial support for the idea taken from the A-J fit perspective that applicants can use OJAs to draw assumptions about personality requirements (Koçak et al., 2022; Newman & Lyon, 2009), as 66% of higher education OJAs did contain corresponding keywords. With the closed vocabulary from Ostendorf, we could show that 88% of the OJAs (linked to job functions in the generic value chain) did contain personality-related keywords. Thus, we found stronger support for the cited idea that personality-related information can be drawn from OJAs. In the context of the A-J fit perspective, there is also evidence about a significant, positive relationship between an applicant’ personality dimension and OJA’s listing this personality dimension as a requirement (Stevens & Szmerekovsky, 2010). As 88% of our OJAs contained personality-related keywords only, in the case of the remaining 12%, organisations cannot expect a high A-J fit. Additionally, past research showed that personality plays an important role in HR (Lounsbury et al., 2008), sales (Conte & Gintoft, 2005), and finance (Baker et al., 2022) job functions. Our analysis applying the weighted (mean-) median-split, as shown in Table 4, highlighted the job functions of HR (referee and clerk), sales, and finance, with higher-than-average loadings across all dimensions. Interestingly, we uncovered logistic professionals as an additional job function in which personality plays an important role in OJAs.

However, we also offer contradictory findings from the A-J fit perspective within the P-E fit theory. There are two different sources that provide generic insights in the context of the P-J fit that both contradict each other: on the one hand, there is evidence that extroversion (SU) and agreeableness (AG) have a higher level of P-J fit as compared to employees with lower levels in both dimensions (Goh & Lee, 2016). On the other hand, there is evidence that neither extroversion (SU) nor openness (CU) is significantly related to P-J fit, while conscientiousness (CO) and emotional stability (ES) are (Ehrhart & Makransky, 2007). Our results suggest that although there is only a few personality clusters across various job functions, the differences between the clusters may be an explanation why the two generic findings contradict each other. Thus, our results call for more detailed research within the P-E fit theory differentiating between job functions and managerial level, especially via clustering. This contribution is even more strengthened via one of the latest literature reviews about P-E fit theory-based research. It showed that cluster analysis was only used once for studying work role proficiency and proactivity (Vleugels et al., 2023). In addition to the two contradictory findings, there is evidence that personality is more important at higher levels of management (L. W. Porter & Henry, 1964). Our results show that the two HR levels of clerks and referees both have higher than mean levels across all dimensions of personality, and both profiles are significantly parallel on equal levels, as shown in Table 5. While we would argue that there is no such expected difference observable, possibly the hierarchical levels of the two HR jobs are not diverse or high enough to be used in regard to that question.

Linking all profiles across job functions via profile and cluster analysis offers new insights into personality research from the perspective of the supplementary perceived A-J fit, because there is a group of significantly related profiles (IT, procurement, and R&D). A closer look at the OJAs included in the dataset and their job environment description offers the assumption that the included profiles of IT administrators, lower-level procurement referees with small budget responsibilities, and R&D employees all have individual/single workplaces in which personality plays a subordinate role.

Practical & Social Implications

As the weighted closed dictionary of Ostendorf (1990, 1994) worked well on our text corpus of OJAs, HR professionals may make use of it to reflect whether the announced personality profile in their job advertisements fits with their personality profile in mind before the online publication. Alternatively, HR professionals can check whether the use of incorrect words may guide applicants in an incorrect A-J fit reflection. Additionally, HR professionals may also use this approach to extract personality from cover or motivational letters to contrast these results either with a survey-based candidate assessment or against their ideal profile in mind. While this is of greater value for large enterprises with their own IT departments, the programming skills of HR professionals in small and medium-sized enterprises without any IT departments might not be as adequate as needed. Thus, our study supports the call for (standardised) HR analytic skills (Kulikowski, 2024; Marler & Boudreau, 2017). Besides this HR impact, there is also evidence about sales professionals that showed that agreeableness is not related to sales job success (Barrick et al., 2002). However, our results in Table 4 show that sales provided the highest value (M = 8.66, SD = 0.66) across all job functions and dimensions of agreeableness. Thus, the question arises, why are companies searching for highly agreeable salespeople if they are not successful?

When summarising the cluster analysis via a mean (0.25) ± standard deviation split (0.14), as shown in Table 7, our findings demonstrate the possible risk that companies are currently searching for only four different types of personalities to hire. While this may be understandable from the perspective of efficiency, it is irritating from the perspective of diversity. Interestingly, although all clusters showed a low to average level of conscientiousness (CO), others have shown evidence that conscientiousness is an important predictor of performance (Barrick & Mount, 1991; Barrick et al., 2001; Hurtz & Donovan, 2000; Salgado, 1997; Schmidt et al., 2008). This means that practitioners do not seem to consider this evidence at hand.

Table 7.

Major Personality Clusters Identified in OJAs Across Different Job Functions Based on Mean ± Std split.

Cluster number	SU	AG	CO	ES	CU
0	High	High	Average	Average	Average
1	High	Average	Low	Average	High
2	Average	Average	Average	Low	Average
3 and 4	Average	Average	Low	Average	Average

Additionally, cluster 3, covering nearly all the OJAs of sales (98%), only showed average levels of surgency/extroversion (SU) and agreeableness (AG). This was unexpected in relation to the findings of Conte and Gintoft (2005), who reported that surgency/extroversion (SU) is positively correlated with sales performance (r = .20) and overall performance (r = .25). Thus, we would have expected a high level.

Limitations and Suggestions

As our study is the first to apply the closed vocabulary from Ostendorf, there are several limitations given by the decisions made in our paper, offering several areas of further research. From a sample-based point of view, first, we conducted our analysis on a subsample of around 3,000 OJAs linked to generic job functions based on Porter’s Value Chain to better address the applicant’s perspective in search of a generic job in an organisation (A-J fit). However, our dataset offers the opportunity to repeat the analysis without focusing on specific generic job functions for more than 151,000 OJAs. Thus, the OJAs could be grouped to study industry (NAICS [US] or NACE [EU]) or job skill-based (O*Net [US] or ESCO [EU]) heterogeneity. While the negative effect would be that the job functions are more spread, the positive effect would be that the words covered from Ostendorf’s (1994) vocabulary would increase from 17 to 41%, as revealed by a first trial. In the case of increasing the sub-sample size, the impact of Eichstaedt et al. (2021) suggests that only texts of a corpus with a total word count higher than the average may be tested. Second, our sample only covered German OJAs that probably have a cultural bias. Thus, Ostendorf’s (1994) vocabulary can be tested on other corpora in terms of a comparative study. These may form either an uninformative or a reference distribution. In the case of an uninformative distribution, the assumption is that in the corpora, no or only a few words are used to describe personality (such as law cases or cooking recipes). In regard to the reference distribution, the assumption would be that in the corpora an equal or higher number of words is used to describe personality (such as OJAs from different languages, Pennebaker’s Essay data, or political debates). If the corpora stem from different languages, the vocabulary needs to be translated accordingly.

From a methodological point of view, there are also improvements that need to be made. First, the simple algorithmic-based stemming of a given vocabulary could be replaced with machine-learned lemmatisation for German texts (e.g. the new HANTA Python package). Thereby, the hit ratio of the words may be further increased and analysed in a comparative study. Second, the k-means clustering algorithm applied for the elbow diagram has the limitation that within-cluster variation is similar across all clusters, while DBSCAN (Schubert et al., 2017) addresses uneven cluster sizes with outlier removal, and OPTICS (Ankerst et al., 1999) addresses the uneven density of clusters. Thus, a comparative analysis of the same dataset but with different clustering approaches may offer different insights. Third, the rather old vocabulary of Ostendorf may, on the one hand, be tested against the classical LIWC approach, where the linkages of the categories to personality must be indirectly drawn from correlational studies. On the other hand, it may also be extended with the latest research for identifying personality in text (Holtrop et al., 2022) or by the use of part-of-speech (POS) tagging (Church, 1992) for a given corpus. Therefore, adjectives can be extracted from the specific corpus of OJAs to analyse which additional adjectives need to be included in the vocabulary for a given text class. In both cases, all additional adjectives needed to be evaluated via SMEs using the Ostendorf (1994) scale (−3 to +3) for all traits.

Conclusion

This study is the first that successfully applied Ostendorf’s (1994) weighted closed vocabulary approach to analyse OJAs, identifying different loadings on personality dimensions across different job functions.

For future research, we suggest a different type of sampling of OJAs to address industry or job skill-based heterogeneity, using OJAs from different countries, translating the vocabulary for cross-lingual studies, or using different German corpora, leaving the context of the A-J fit for cross-content studies. Additionally, we suggest different methodological approaches like lemmatisation instead of stemming, alternative clustering methods like DBScan, a comparative analysis against LIWC, and an extension of the vocabulary with the latest research from Holtrop and colleagues. Generally, we suggest applying Ostendorf’s idea to all existing closed vocabularies for culture, gender, sentiment, diversity, and so forth.

For a practical application of the vocabulary, we support the request for training in HR analytic skills, as the techniques used stem from natural language processing. In case the analytic skills are available, HR shall analyse their job advertisement for match and mismatch before putting them online and apply the vocabulary also to applicant cover letters to see which personality the applicants do communicate. Additionally, because across 3,239 OJAs only 4 personality types could be identified, it is suggested that HR should better take care of the words used to differentiate the personality profiles depending on the job function.