Using Online Vacancy and Job Applicants' Data to Study Skills Dynamics

Major transformative phenomena, such as technological progress and trade, are shaping labor markets. Skills are an important factor in this process as they significantly influence how labor market transformations affect employment opportunities. For example, the demand for specific skills depends on whether these are complementary to or can be substituted by a newly introduced technology. As a result, workers' skills determine the comparative resilience of certain groups in contemporary labor markets, while affecting the relative vulnerability of others.

A rich literature has examined skills dynamics in Europe and, in particular, the United States. Computer technology has been found to replace tasks that can be routinized and complement nonroutine analytical and interactive tasks (Atalay et al., 2020; Autor et al., 2003), although the rising demand for nonroutine analytical tasks has been reversed more recently (Beaudry et al., 2016). Some studies predict that artificial intelligence and robotics will also replace nonroutine analytical labor and lead to large-scale job destruction (Frey & Osborne, 2017). Other analyses, in contrast, demonstrate the importance and nonreplicability of interactive skills within occupations, projecting smaller net job losses (Arntz et al., 2016). Given systematic differences in labor markets, it would be misleading to extrapolate these findings directly to other countries. Yet, the knowledge on skills dynamics is scarce outside of Europe and the United States, which is largely driven by the absence of adequate sources of data. Such knowledge is urgently needed today as it would contribute to understanding which policy responses are required to better prepare workers and respond to employers' needs in fast-changing labor markets.

In this article, we assess whether online data on vacancies and applications to a job board (or job portal) are a suitable source for studying skills dynamics outside of the high-income economies that the literature has so far focused on. The advantage of this approach is its reliance on data that is currently available in many countries across the world, thereby allowing for country-specific analyses that do not need to assume that occupational skills are the same across countries. Another distinguishable feature is the ability to study detailed skills dynamics across time, representative of both labor demand and supply. This results from the panel nature and granularity of the data, which is distinct from currently available survey data from emerging and developing economies. Finally, to the best of our knowledge, we are the first to explore this approach in the context of an emerging economy.

To devise our taxonomy, we build on the growing literature that uses online vacancy data from the United States to study skills dynamics, in particular those key contributions that have proposed seminal taxonomies to categorize skills (Deming & Kahn, 2018; Deming & Noray, 2020; Hershbein & Kahn, 2018). For our purposes, a limitation of such taxonomies is that they are specifically tailored to professional workers in the United States. Therefore, we complement such efforts, by further incorporating manual skills and broadening the underpinnings of socioeconomic skills, with a view to expanding the scope of possible analyses of online data beyond workers with high formal qualification levels. This is particularly relevant when studying the supply and demand of skills in emerging economies, and it is also relevant for high-income countries as it allows for analyses that are tailored to a more comprehensive set of workers. Specifically, to obtain a taxonomy that is comprehensive but succinct, we systematically aggregate three broad categories of skills – cognitive, socioemotional, and manual – and 14 commonly observed and recognizable skills subcategories. We identify these based on predefined unique tasks characterized through keywords and expressions stemming from our literature review and complemented by our own analysis. Each skills subcategory is therefore grounded in the academic literature, where we combine insights from seminal studies in economics (including also the literature on skill-biased technological change, see, for example, Acemoglu & Autor, 2011) and psychology (see, for example, Almlund et al., 2011).

We then develop a text-mining methodology that allows implementing our skills taxonomy in big online vacancy and applicants' data. This methodology provides a tool for researchers and practitioners to (i) preprocess raw information in textual data, and (ii) detect and classify the skills contained in a given country-specific dataset using natural language processing (NLP) techniques. We test this approach empirically, by exploiting information from the Uruguayan job board BuscoJobs. Thanks to the availability of these rich data, the setting of Uruguay is well suited for our purposes. In more detail, BuscoJobs is the biggest private job-search portal in the country and our data include the entire content uploaded between 2010 and 2020, capturing both the supply side (through applicants sharing their current profile and labor market biographies) and the demand side (in the form of vacancies posted by firms). Uruguay also represents an informative setting, since it allows studying a labor market – and skills supplied and demanded – that are clearly distinct from the cases the literature has so far focused on. To illustrate this, Uruguay is classified as a high-income country by the World Bank, but its current GDP per capita amounts to only 27% of the US American GDP per capita (see World Bank, 2023). In 2019, informal employment rates were at 24.5% in Uruguay, which is more than twice as high as the estimated 9.8% for Northern America (ILO, 2023). Moreover, the Uruguayan labor market regulation is comparatively strong, given the decisive influence of tripartite collective bargaining, which can impact the speed of labor market transformations (ILO, 2014).

To test our methodology, we use five types of metrics: first, the degree of representativeness and bias of these data in comparison to the overall labor market and how stable this is across time; second, the share of observations (vacancies and applicants' job spells) classified in terms of their skills content; third, the possibility of capturing a broader set of skills vis-à-vis taxonomies from seminal studies that have analyzed similar data in a high-income context and neglected, in particular, manual skills; fourth, the difference in results obtained when using Uruguayan data compared to what we would have obtained when relying instead on US O-NET data under the assumption that occupational skills are the same across these two countries; and finally, the applicability and suitability of our approach to study skills dynamics.

To preview our empirical results, we find that our NLP model is able to characterize the vast majority of vacancies and applicants' job spells in terms of the skills requirements and the skills that people have, respectively. The results significantly improve once we consider synonyms in addition to the keywords and expressions included in our predefined skills taxonomy. Then, 64% of applicants' job spells and 94% of vacancies are assigned to at least one of our 14 skills subcategories. While we lack a direct benchmark to compare these results (due to the limited availability of supply-side data on job boards and aggregators), we believe our model allows for a meaningful classification of skills that should enable credible analyses. This is especially notable considering the substantial heterogeneity in the quality of information available in the free-text descriptions we use to code the skills variables and gives hope to the possibility of replicating our methodology in similar data sources of other countries.¹

We also find that, although our data are not fully representative of the country's labor force, they allow for meaningful analyses of labor supply and demand pertaining to intermediate and even lower qualification levels. Based on the Uruguayan household survey, we show, for example, that applicants tend to be younger, more educated, and more likely to live in the capital than the overall labor force. They are also more likely to work in clerical support occupations and are less represented in craft and other trades. This is inherent to online vacancy and applicants' data and may require focusing the analysis on specific segments of the labor market and/or using weighting techniques to account for specific biases (Fabo & Kureková, 2022). At the same time, we show that in addition to highly qualified work, the data include meaningful numbers of vacancies requiring and applicants having intermediate and even lower qualification levels, which represents a characteristic that is not commonly found in online vacancy datasets from the US. We also establish that the bias in the BuscoJobs data remains stable across time for both vacancies and applicants' job spells, since the discrepancy between the occupational distribution in the BuscoJobs data and the household survey data does change significantly from 2010 through 2020. This means that the data are suitable for studying dynamics over time (for a similar argument, see Deming & Noray, 2020; Hershbein & Kahn, 2018).

We also show the importance of capturing a mixture of different skills, which might be especially relevant when analyzing skills dynamics outside of Europe and the United States. A large share of classified skills (61% in the case of the applicants' data) is attributable to keywords and expressions from studies analyzing online data, which neglect manual skills. However, additional sources also play a meaningful role. For the applicants' data, 31% of classified skills thus relate to keywords and expressions from studies using non-online sources and 9% to complementary keywords and expressions from the pilot version of O-NET Uruguay.

We demonstrate the importance of employing country-specific data when assessing skills dynamics outside of Europe and the United States, in line with recent literature (Lewandowski et al., 2020, 2022; Velardez, 2021). Within occupations we find considerable differences when comparing our findings to what we would have obtained, had we used O-NET data from the United States. Looking at the relative importance of the three broad skills categories (i.e., manual, socioemotional, and cognitive skills), we find that manual skills play a larger role in the composition of occupations using our Uruguayan data, than using the US O-NET. The importance of cognitive skills varies more strongly in our data from 60% for professionals to below 20% for plant and machine operators, while all occupations have in general lower scores in the socioemotional skills category in our approach than according to the US O-NET data.

Finally, we end with an illustration of how our approach can inform the analysis of skills dynamics. We illustrate the evolving supply and demand for specific skills narrowing our focus to a subset of the data, where we consider only applicants who applied to at least one vacancy and vacancies that received applications from at least one applicant. Notably, we observe a consistent gap between the skills sought by job vacancies and those explicitly indicated by applicants, and this gap widens over time. We posit that this gap may be related to employers' increasing, more complex, or more exigent demands placed on prospective employees, as reflected in the lengthier job postings. While these findings are not meant to offer conclusive results, they provide a first foundation for future research utilizing big data from job boards and our methodology.

In addition to relating to the literature for high-income countries as discussed further above, this study advances the literature on skills dynamics in emerging and developing countries. While the relative scarcity in empirical evidence in these economies is in large part due to the absence of appropriate data, some studies have addressed this data challenge in innovative ways. One strand of the literature studies skills dynamics in low- and middle-income countries by imputing occupational task information from the US O-NET data (Almeida et al., 2017; Bhorat et al., 2018; Reijnders & de Vries, 2018). This requires making the strong assumption that the task content of occupations is invariant across countries, which we do not need to make in our approach. A second strand of the literature measures tasks directly in the countries studied, combining survey data from the Programme for the International Assessment of Adult Competencies (PIAAC) and the Skills Measurement Program (STEP) with other longitudinal data sources. These studies identify significant differences in the skill composition of jobs and occupations across countries and highlight the importance of analyzing these questions using country-specific data (Carbonero et al., 2023; Caunedo et al., 2023; Lewandowski et al., 2020, 2022; Lo Bello et al., 2019). For example, Lewandowski et al. (2020) show that even within the same occupations, emerging and developing countries rely more strongly on routine work than developed economies. Moreover, Zapata-Román (2021) documents, in contrast to advanced economies, increases in earnings for routine occupations in Chile, while Xing (2021) finds the opposite for China.² Unfortunately, for developing and emerging economies, PIAAC and STEP surveys are currently available as single cross-sections only.³ In contrast to our study, this strand of the literature therefore neglects skills dynamics that occur within occupations over time, although findings for high-income countries show that within-occupational skills changes can be significant (Atalay et al., 2020; Spitz-Oener, 2006). Moreover, skills variables in the PIAAC and STEP surveys are limited in number, making it difficult to assess skills in a comprehensive manner. This stands in contrast to the granularity of the data that we propose.

This article is structured as follows. Section 2 develops the skills taxonomy and explains how we classify unique skills within categories and subcategories. Section 3 describes the BuscoJobs data. Section 4 summarizes our methodology to implement the skills taxonomy in the BuscoJobs data. It describes the process we follow to create skills variables through the identification of keywords and expressions and the use of NLP techniques. It also provides an illustration of how our approach can inform the analysis of skills dynamics. Section 5 concludes with a brief discussion of how online job data could be used in the future to analyze a range of research questions for emerging and developing countries.

We now devise our taxonomy. The discussion first motivates our objectives and general approach. Grounded in the literature from labor economics and psychology, we then present the conceptual underpinnings and definitions of our broad and more detailed skills categories.

Our taxonomy has to strike a balance between different objectives. To begin with, we want to categorize skills in a comprehensive way, capturing broadly defined skills with a view to understanding major labor market trends. In addition, we want these broadly defined skills categories to be suitable for a Latin American context and, more generally, to be representative of an emerging and developing-country labor market. This requires adjustments in comparison to approaches that were developed for a North American or European country context. Finally, our taxonomy needs to be suitable for an implementation using online data on job vacancies and applicants. This requires adapting the taxonomy to the mode of expression and vocabularies specific to this type of data.

With these objectives in mind, our taxonomy consists of the three broad categories of cognitive skills, socioemotional skills, and manual skills that we disaggregate into 14 subcategories (Table 1). Recent research confirms that these three broad skills categories represent vastly different attributes. Across workers' employment biographies, each category entails unique learning and adjustment patterns and therefore produces distinct returns (Lise & Postel-Vinay, 2020). This motivates our choice to organize the taxonomy around these three categories.

We deliberately focus on skills, rather than occupations. This is based on the realization that, first, skills are a central dimension for understanding how major transformative phenomena, such as technological progress or trade, affect employment opportunities (e.g., Acemoglu & Autor, 2011; Autor et al., 2003). Second, occupations are characterized by complex bundles of skills that change across time (e.g., Arntz et al., 2016; Atalay et al., 2020; Spitz-Oener, 2006). This implies that workers – across qualification levels – perform a combination of cognitive, socioemotional, and manual skills.⁴ The composition of these skills bundles is expected to shape how mega-trends (like technology and trade) affect workers' labor market situations. In a similar vein, ongoing policy debates emphasize the importance of skills that are portable across jobs and occupations (see ILO, 2021).⁵

As shown in Table 1, our taxonomy builds on Deming and Kahn (2018), who augment the task-based approach of Autor et al. (2003) to adapt it to the specific characteristics of online job data (BurningGlass data for the United States in their case). We have extended the approach of Deming and Kahn (2018) by adding information provided by a range of different studies (Atalay et al., 2020; Autor et al., 2003; Deming & Noray, 2020; Heckman & Kautz, 2012; Hershbein & Kahn, 2018; Kureková et al., 2016; Spitz-Oener, 2006) and O-NET Uruguay.⁶ Specifically, within the categories of cognitive and socioemotional skills, we include additional keywords that are meant to broaden the scope of descriptions used. This also allows us to capture terms whose popularity changed over time, but which refer to the same skill (see Deming & Noray, 2020).⁷ In addition, we add the entire category of “manual skills.” Manual skills were not the focus of Deming and Kahn (2018) but are important for our goal of obtaining a comprehensive skills taxonomy. Manual skills are, moreover, particularly relevant in the context of an emerging country's labor market, such as that of Uruguay. Overall, we make use of the various keywords used by seminal articles to be the most comprehensive possible in the definition of categories and subcategories, while devising our conceptual framework.⁸

The 14 subcategories of our taxonomy capture both skills that pertain to tasks that workers perform on the job and skills that refer to individuals' personal attributes.⁹ Both types of skills are important to provide a consistent representation for labor market developments in a wide range of contexts. In combination, they encompass skills that employers demand in vacancies and skills that workers supply and present in their online profiles. It is important to note that even if some skills subcategories are closely related, the words used to categorize them are mutually exclusive.¹⁰ This allows for the unique identification of skills categories in our granular data.

Our categorization has the advantage of being both complete and succinct. The comprehensiveness of this approach is necessary for it to be relevant to different countries' realities and is possible thanks to the granularity of online data. Meanwhile, its compactness into three broad categories and only 14 subcategories is tailored to research purposes, where overly detailed taxonomies, ontologies, or uncategorized lists of skills could introduce complexity that would be challenging to utilize effectively.

Within the category of cognitive skills, we define the subcategory “cognitive skills (narrow sense)” as the abilities or qualities needed to perform tasks that require analysis and calculation, problem-solving, intuition, flexibility, and creativity (Acemoglu & Autor, 2011, p. 1076; Autor et al., 2003, p. 1284). A large literature has studied the relative demand for cognitive skills and their returns in high-income countries, and cognitive skills are recognized in policy debates as being central for workers' resilience to labor market transformations (ILO, 2021).

One strand of this literature thus looks at the polarization of occupations by skill level vis-à-vis technological change and automation.¹¹ This literature finds that occupations largely requiring cognitive skills have thrived with the development of computer technology, both in terms of job creation and wage growth. This is driven by cognitive skills that are difficult to routinize (notably, those associated with non-repetitive cognitive tasks), where occupations relying on such tasks have been more resilient to technological shocks. Meanwhile, low-skilled occupations (also non-substitutable by technology, as discussed below) have likewise witnessed a broad-based increase in employment, relative to middle-skilled occupations. More recent studies suggest that the rise of high-skilled occupations has evolved, documenting a reversal in the demand for cognitive skills in the United States in 2000 (Beaudry et al., 2016). While the share of highly educated college graduates has increased, occupations intensive in cognitive skills have seen little wage and employment growth after 2000. The finding is consistent with the Information Technology (IT) revolution, and hence the introduction of IT as a general-purpose technology, having reached a maturity stage. This implied that high-skilled workers moved into occupations relying less strongly on cognitive skills (Beaudry et al., 2016). The reversal in the demand for cognitive skills therefore has consequences for the wage structure of occupations (see also Roys & Taber, 2019).

Another strand of the literature complements this knowledge by focusing on skill changes within occupations.¹² These studies take advantage of richer data sources with greater granularity, such as vacancy data. They find substantial variation in skill requirements within occupations. While cognitive skills are an important driver of this variation, socioemotional skills are taking on a key role (Kureková et al., 2016). In this context, Deming and Kahn (2018) show that a greater emphasis on cognitive skills, especially when they are demanded in combination with socioemotional skills, is associated with higher wages and firm productivity. Therefore, the considerable within-occupational variation in the demand for such skills contributes to explaining a significant share of existing patterns of wage inequality. In addition, changes in skills over time play a decisive role also within occupations (Atalay et al., 2020; Spitz-Oener, 2006). As one example, STEM occupations have seen pronounced changes in cognitive skills requirements – for example, STEM vacancies requiring skills linked to machine learning and artificial intelligence rose by 460% in the decade before 2017 – which can be attributed to the rapid diffusion of technological innovation in STEM (Deming & Noray, 2020). Against this background, the words we choose to identify cognitive skills aim to match the analytical job tasks defined by Autor et al. (2003) that are used by most articles studying routine-biased technological change and employment polarization. We add additional words suggested by Atalay et al. (2020), Deming and Kahn (2018), Deming and Noray (2020), Spitz-Oener (2006), and O-NET Uruguay.

Finally, we complete cognitive skills by adding five additional subcategories from Deming and Kahn (2018): computer skills, software skills, writing skills, financial skills, and project management skills. We also add one additional category suggested by Deming and Noray (2020), which refers to machine learning and artificial intelligence. These subcategories are intricately linked to the cognitive skills described before and represent topical subcategories that organize cognitive skills around themes. We categorize them separately because they are commonly listed in a wide range of online job vacancies and applicants' work experiences. The list of topical subcategories targets in particular white-collar jobs (Deming & Kahn, 2018), which means that this list can be consulted at the detailed level or condensed, depending on the focus of a given study on a particular segment of the labor market.

Our second broad category of socioemotional skills adds to the categorization a set of personal attributes that involve intellect, but more indirectly and less consciously than cognitive skills. The literature uses a range of expressions to refer to such skills, including noncognitive skills, soft skills, socioemotional skills, or personality traits. These different terms suggest different properties.¹³ First, we avoid the use of the term “noncognitive skills,” as all skills require some sort of cognition. Second, while “traits” gives a sense of immutability, “skills” connotes the possibility of learning (Kautz et al., 2014). A long debate in the psychology literature has discussed whether socioemotional skills are stable across different situations at a fixed point in time, as well as the degree of malleability of these skills over the life course (see Almlund et al., 2011, sec. 2). Today, most psychologists and economists support the notion of a stable personality across situations, which is supported by a large body of evidence showing that stable socioemotional skills exist and predict a variety of behaviors (Kautz et al., 2014).¹⁴ However, as pointed out by Heckman and Kautz (2012) “while traits are relatively stable across situations, they are not set in stone. They change over the life cycle.” The consensus that socioemotional skills are malleable and can be learned, is key in terms of policy. If this were not the case, there would be little room for training initiatives. Another decisive question is whether socioemotional skills can still significantly change during adult life. Disciplines diverge on this question. Economists like Carneiro and Heckman (2005) and more recent publications by Heckman and coauthors put emphasis on the malleability of these skills until adolescence and early adulthood. From this perspective, parenthood, education policies, and workplace-based internships and apprenticeships are the central means for developing socioemotional skills. Meanwhile, social psychologists argue that at least some of these attributes (e.g., emotional intelligence) can be learned at any age, which places greater emphasis on training, learning, and education policies for adults.¹⁵ Sociologists and management scholars agree, emphasizing the indispensability of experiential learning for some socioemotional skills such as problem solving, teamwork, and other social skills (see Green et al., 2001, sec. 2).

Measuring and classifying socioemotional skills is a challenge (Humphries & Kosse, 2017). Yet, the psychology literature has arrived at a well-accepted categorization of these skills, called the “five-factor personality model” or “big five” (McCrae & Costa, 2008). It includes agreeableness, conscientiousness, emotional stability, extraversion and autonomy, and openness to experience. A large literature on the labor market returns of so-called noncognitive skills has established that different noncognitive skills are important predictors of life outcomes, such as education and labor market success (Heckman & Kautz, 2012). Some of these personal attributes are correlated with cognitive skills (Kureková et al., 2016), but their explanatory power on labor market and education outcomes goes beyond this correlation.¹⁶ A study for the United States found that improvements in noncognitive skills are much more important for future earnings and employment than similar improvements in cognitive skills (Heckman et al., 2006). Similarly, a study for Sweden showed that both cognitive and noncognitive skills are strong predictors of future earnings, but that noncognitive skills have a stronger effect for people at the low end of the earnings distribution (Lindqvist & Vestman, 2011). In the emerging and developing world, four studies have assessed the causal effects of socioemotional skills on labor market outcomes, by experimentally evaluating the effect of job training programs focused solely on socioemotional skills and different categories of workers. The evidence is more mixed, with two studies finding positive effects in India and Togo (Adhvaryu et al., 2023; Campos et al., 2017), another study from the Dominican Republic finding positive effects for women but not men (Acevedo et al., 2020), and a final study for Jordan finding null effects (Groh et al., 2016).

In terms of our taxonomy, we divide the five-factor model between the subcategories “character skills” and “social skills.” Our subcategory “character skills” follows Deming and Kahn (2018), deviating somewhat from their definition (see Table 1). It includes conscientiousness (i.e., “the tendency to be organized, responsible, and hardworking,” American Psychological Association, 2020), openness to experience (i.e., “the tendency to be open to new aesthetic, cultural, or intellectual experiences,” American Psychological Association, 2020), and emotional stability (i.e., this is the contrary to “neuroticism.” “Emotional stability is predictability and consistency in emotional reactions, with absence of rapid mood changes,” American Psychological Association, 2020). It also encompasses dimensions such as being relaxed, independent, self-confident, and the degree of vulnerability to stress (Brunello & Schlotter, 2011; Heckman & Kautz, 2012).

There is no single agreed way to aggregate the five-factor personality model into broader categories. Most researchers bundle together various socioemotional skills depending on their specific research questions. We aggregate them into “character skills” and “social skills,” as distinguishing between all five categories separately would increase the probability of making mistakes at this finer level of aggregation, without necessarily adding considerable value to the analysis. Instead, the two groups of “character skills” and “social skills” are based on the similarities in their predictive nature that have been found by the empirical literature. As such, we bundle conscientiousness, emotional stability, and openness to experience together within our subcategory “character skills,” because these three factors appear to share the ability to predict education and labor market outcomes, although not with equal strength (Almlund et al., 2011; Brunello & Schlotter, 2011). For example, conscientiousness stands out as the most predictive trait of the five-factor model on future outcomes. It has been found to predict educational attainment, health, and labor market outcomes, in some cases, as strongly as measures of cognitive ability.¹⁷ While conscientiousness predicts performance and wages across a wide spectrum of jobs, the predictive power of cognitive skills decreases with job complexity (Almlund et al., 2011; Kautz et al., 2014). Meanwhile, attributes related to emotional stability ‒ especially internal locus of control, or the belief that one can determine one's success as opposed to believing that outcomes are the result of fate or luck ‒ positively predict earnings (Brunello & Schlotter, 2011) and job search effort (Almlund et al., 2011). Finally, openness to experience predicts finer measures of educational attainment, such as class attendance (Almlund et al., 2011) but also years of education (Borghans et al., 2008). It has also been associated with transversal competencies, such as sense of initiative and entrepreneurship, which are important factors for education and labor market success (Brunello & Schlotter, 2011). The words we use to define such “character skills” come from Deming and Kahn (2018), but also from Deming and Noray (2020), Kureková et al. (2016), and Heckman and Kautz (2012).

We complete the category of socioemotional skills with three of Deming and Kahn (2018)'s subskills groups, which capture social skills (as mentioned before), people management skills, and customer service skills. Within “social skills” we include the original words suggested by the authors (e.g., communication, teamwork, collaboration). We add the remaining two categories from the five-factor personality model, namely agreeableness and extraversion. Agreeableness is the tendency to be cooperative toward others (American Psychological Association, 2020). Extraversion is the orientation people show toward the outer world and social contact (American Psychological Association, 2020). In our taxonomy, we characterize these two personal attributes with words such as cooperation, flexibility, empathy, and assertiveness.¹⁸ We also add the subcategories of “people management skills” and “customer service skills,” which are determined by a large set of socioemotional skills, but usually listed separately in online job vacancies and workers' job experiences. As before, we use various references as sources for keywords with a view to arriving at a complete taxonomy and a more comprehensive use of words to define each category (see the last column of Table 1).

The importance the literature attributes to social skills and other socioemotional skills has considerably increased during the last decade and policy debates emphasize the core relevance of these skills across workers' life cycle (ILO, 2021). Research shows that employers in the United States and Europe rank some social skills above cognitive ones, particularly in low-skilled labor markets (Bowles et al., 2001; Kureková et al., 2016). Moreover, studies point to an increasing complementarity between cognitive skills and social skills (Arntz et al., 2016; Borghans et al., 2014; Deming & Kahn, 2018; Weinberger, 2014). As computers substitute for a wider set of noninteractive tasks, interpersonal skills are becoming increasingly central in a wide range of professional jobs (Autor, 2014; Lu, 2015). Interestingly, according to research in Europe, the increased demand for interactive skills does not substitute formal education, but appears in addition to it (Kureková et al., 2016). The implications for workers in emerging and developing countries are potentially significant, as large shares of employment in these countries are concentrated in sectors and occupations (e.g., services) that require disproportionately interpersonal skills.

Our third category, manual skills takes as point of departure how Autor et al. (2003) categorize manual tasks. In their analysis, manual tasks are divided between routine and nonroutine ones. The former are activities that necessitate finger-dexterity skills and the latter activities requiring “situational adaptability, visual and language recognition, and in-person interactions” (Acemoglu & Autor, 2011, p. 1077). Broadly speaking, routine manual tasks are more prevalent in production and operative occupations and nonroutine manual tasks in service occupations, but also in some production and operative positions requiring physical or situational adaptability (Acemoglu & Autor, 2011).

The model by Autor et al. (2003) captures the vulnerability of routine manual labor in light of automation and outsourcing. We implement this category in our taxonomy using words such as picking or sorting, repetitive assembly, and mixing or baking ingredients. However, not all manual tasks enter this category. Several tasks require motor processing and visual capabilities that, given the contemporary state of technological progress, cannot easily be programmable (e.g., driving a car through traffic, cleaning, and housekeeping). In the United States, lower-skilled workers have thus moved into service occupations that are associated with such tasks and have witnessed employment growth due to a concurrent change in consumer preferences that favored low-skilled services (Autor & Dorn, 2013; Lee et al., 2022). Outside of high-income countries, excess supply of labor is disproportionately absorbed by low-value-added urban services, which contributes to an underdeveloped manufacturing sector (e.g., Rodrik, 2018). Nonroutine manual skills relate to tasks such as accommodating, serving, or cleaning, but also attending cattle in farms, pruning and treating plants, driving, and even performing gymnastics and other sports requiring balance.

Finally, we complement the categorization of manual skills with a “physical skills” subcategory that pertains specifically to personal attributes, physical strength and effort, such as resistance, as well as related tasks such as those requiring walking and running, and carrying heavy loads.

Manual skills have been less studied in the literature than cognitive ones. One reason is that these tasks, especially the part that can be routinized, have been declining in the advanced world (Spitz-Oener, 2006). Another reason is that manual skills, and more generally low-skill jobs, are underrepresented in online sources, such as the BurningGlass data, and are therefore not the focus of studies using these data (Hershbein & Kahn, 2018). Yet, there are some findings from the advanced world that shed light on the importance of integrating this skills category into a research-oriented taxonomy, given that these tasks are important when considering workers at the lower end of the wage distribution (Autor & Dorn, 2013). While research is scant outside the United States, we expect these tasks to be at least as important in the emerging and developing world, if not more important. Recently, Lise and Postel-Vinay (2020) studied the returns of manual skills requirements using O-NET data from the United States. They find that manual skills have in general moderate returns, and that they are easily accumulated on the job but also relatively fast lost when not employed. Roys and Taber (2019) explore the payoff to different skills for low-skilled workers in the United States over the life cycle. They find that while the payoff related to interpersonal skills has increased over time, the returns of manual skills remain the most relevant factor determining the wages of low-skilled workers. From a policy perspective, the finding suggests that investments in manual skills are key for improving the wages of these workers. Importantly, however, while there is a clear understanding that education and training is a path to develop cognitive skills, the literature has not sufficiently explored how to develop manual skills further. Meanwhile, policy practitioners, through their experience implementing these policies, point to the importance of work-based learning to develop technical skills, including manual ones (Cedefop, 2015; European Commission, 2013; Kis & Catriona Windisch, 2018; ILO, 2017). Therefore, understanding these developments better has profound implications for low-skilled workers. This extends to large segments of workers who perform manual work in emerging and developing countries, including in informal labor markets. Yet, it is also relevant for high-income countries as it allows for analyses that are tailored to a more comprehensive set of workers.

Finally, as a cross-cutting dimension, the routine intensity of skills has been decisive for understanding the substitutability of routine-manual and routine-cognitive tasks by technology and hence the polarization of employment structures in high-income countries (Autor et al., 2003; Goos et al., 2014; Spitz-Oener, 2006). Research outside of high-income countries relies on survey data that is available for some countries, while imputing routine intensity for others. It shows that low- and middle-income countries have seen no, or a comparatively small, decline in routine tasks between 2000 and 2017 (Lewandowski et al., 2020). In addition, middle-skilled jobs are not necessarily intense in routine tasks, such that a study for Argentina finds some evidence of a reallocation of employment toward middle-skilled jobs (Maurizio & Monsalvo, 2021).

Empirically capturing routine intensity within our skills categorization would require implementing it as a cross-cutting dimension, as we expect that routine intensity interacts differentially with the various skills categories. The existing literature conventionally relies on keyword-based identification to discern terms associated with routine skills (see Atalay et al., 2020; Autor et al., 2003; Spitz-Oener, 2006). This approach, however, presents complexities that our available data and methodology so far could not definitely resolve. In particular, the nuanced nature of certain skills, such as financial skills which may manifest varying degrees of routine-intensity, requires rigorous validation. Developing a robust methodology to capture skills' intensity effectively in our framework, thus remains a promising focus for future research.

To implement our skills taxonomy empirically, we rely on online data from the Uruguayan, private online job board BuscoJobs, which contains detailed information on (i) workers searching for jobs, (ii) vacancies posted by firms, and (iii) applications jobseekers have made to the vacancies posted. Our data cover the entire content uploaded on the job board from January 2010 through December 2020 in Uruguay.¹⁹ In this section, we start by presenting the features of the job board (Section 3.1), then discuss conceptually issues of representativeness in job board data (Section 3.2), and finally assess the representativeness empirically, for both the applicants' and vacancy data (Sections 3.3 and 3.4).

To post online vacancies through BuscoJobs, enterprises need to create an account on the portal and pay a small fee. The portal offers several types of enterprise subscriptions, which cover different time spans and allow firms to publish between three and 600 vacancies. Firms can view users' curriculum vitaes (CVs) for the duration of their chosen subscription.

The data capture detailed information associated with each job vacancy. This includes characteristics of the firm posting the vacancy (e.g., name and location), firms' expectations and characteristics of the vacancy (e.g., desired age and sex of the candidate, required work experience and formal qualifications, the wage range offered), and a detailed text description of the job advertised. BuscoJobs generated upon request a variable indicating firms' industry, classified at the four-digit level of the International Standard Industrial Classification (ISIC), Revision 4.²⁰ Finally, we created a variable capturing occupational categories, following the two-digit ISCO-08 classification and using a machine-learning approach (see Section 4 for details).

As a distinctive feature, BuscoJobs not only allows firms to post vacancies but also enables jobseekers to register and apply for these vacancies directly through the portal. Jobseekers can create an account for free and include in their profile basic personal information (e.g., sex, birth dates, and places of residence), information about their educational attainment, the technical skills they possess, the languages they master, and, importantly, their entire history of employment (e.g., job spells' dates, position, location, and a detailed open-text description of the position). This allows us to observe rich longitudinal data on individuals' characteristics and employment biographies. Jobseekers can also express their job preferences, whether they are actively seeking a job, whether they are employed or unemployed, and whether they want their profiles to be fully visible for employers.²¹ As for the vacancies, we created a variable capturing the occupational categories pertaining to applicants' employment spells, where we followed the two-digit ISCO-08 classification and employed a machine-learning approach.

Another advantage of the BuscoJobs database, in comparison to other job boards and job aggregators in the country, is that BuscoJobs does not include duplicate job postings. Duplicate job postings are a common characteristic of most job boards as these republish vacancies to increase applications, and of most job platforms as these scrape postings to improve their coverage (Office for National Statistics, 2021). Jijkoun (2016) estimates that a typical vacancy is reposted between two and five times with the proportion of duplicates ranging as high as 50 to 80% in different countries. Such duplication inflates the number of job advertisements, and as these are often not identical, they cannot be easily identified. Duplication thus undermines the representativeness and usability of vacancy data (Evans et al., 2023). In contrast, BuscoJobs has a policy of not reposting. Instead, it publishes vacancies for as long as firms require, for example, a few remain open for up to two months.

BuscoJobs also presents a low volatility over time (Equipos Consultores, 2020), covering a comparatively long time series from 2010 through 2020. It captures vacancies, applicants’ employment histories, and applications. In contrast, other major job boards and aggregators in Uruguay and other data sources that are prominently used in the literature, such as the BurningGlass data, only contain vacancy data.

When assessing the demand and supply of skills and pursuing related analyses, online data have two general advantages, compared to other traditionally used data sources, such as labor force surveys (see also Hershbein & Kahn, 2018).

First, online vacancy and applicants’ data have a high degree of granularity. This allows identifying detailed trends in skills and skills needs by generating variables of workers' tasks performed on the job and the tasks requested in vacancies. These statistical indicators can be computed without recurring to proxies, such as occupations or industries, which do not capture changes to skills within the groups of these categorizations (Hershbein & Kahn, 2018). Furthermore, online vacancy and applicants' data tend to provide detailed information on geographical location, occupations, and industries. These variables can often be generated in a robust way at a highly disaggregated level, thanks to large sample sizes (ILO, 2020).

Second, online vacancy and applicants' data are typically collected at high frequency, making it possible to avoid time gaps associated with more common data collection methods, such as survey data, which are sampled in sometimes long intervals. Some of the delays associated with the processing and cleaning of survey data can also be avoided (ILO, 2020), in particular if the initial database associated with an online job board or job aggregator is well structured.

One of the major drawbacks associated with the use of online job board and aggregator data is their representativeness. As they are not based on random sampling, it is difficult to draw generalized conclusions about the overall working age population or the universe of firms. The data tend to be biased toward certain segments of the labor market, in particular higher-skilled occupations (Cedefop, 2021; Fabo & Kureková, 2022; Hershbein & Kahn, 2018).

The representativeness depends on several characteristics inherent to the national context. This includes the internet penetration rate. In Uruguay, 83.4% of the population used the internet in 2019.²² This is considerably above the global average, estimated at 53.7%, and the regional level, reaching 68.8% for Latin America and the Caribbean (LAC) (see World Bank, 2023). These figures suggest that exploiting web-based big data in the Uruguayan context could be associated with less bias, compared to other countries.

Nonetheless, also among more advanced countries where a high share of individuals uses the internet on a regular basis (e.g., the European Union, where 82.8% of the population used the internet in 2019, World Bank, 2023), there are major differences in the share of vacancies posted online. For example, the proportion of job vacancies published online in 2017 varied from below 50% in countries such as Romania or Denmark to close to 100% in Finland, Sweden, and Estonia (ILO, 2020). One explanatory factor is the share of the population living in urban or rural areas. In urban areas, the services sector tends to play a larger role and the incidence of online job advertisements is thus higher. In contrast, print media and word-of-mouth communication seem to be more relevant in rural areas (ILO, 2020). In Uruguay, a large majority of the working-age population (94.7%) resided in urban areas in 2019 (authors' calculation based on the household survey introduced in the next subsection). This is related to the prominent role of the capital Montevideo and the high degree of centralization in Uruguay.

Online vacancies also tend to be advertised by larger firms (e.g., international firms) and less by smaller firms, or firms operating in the construction, agricultural, and hospitality sectors (ILO, 2020). The incidence of informality also matters, since informal jobs often do not appear on public online job boards or aggregators (Cedefop, 2021). Uruguay has a significantly lower rate of informality than the LAC region as a whole. In 2019, 24.5% of all employment was informal in Uruguay compared to an estimated 53.6% in LAC (ILO, 2023).

Uruguay's labor market characteristics are thus amenable to the use of online portals by firms and jobseekers. Beyond this observation, it is important to empirically assess the degree and direction of biases of our data, which we do next.

The BuscoJobs applicants' database covers around 650,000 user profiles and a total of approximately 1,200,000 job spells. As shown in Fig. 1 (line with round markers), the data coverage is good from 2011 onward. The number of individuals joining the job board tends to increase up until 2017, to slightly decrease again until 2020.

To assess the representativeness of the database, we compare selected characteristics of the job board's users with information from the Uruguayan household survey, the Encuesta Continua de Hogares, which is representative of the country's population. Throughout this section, we present averages over annual results from 2010 through 2020.²³ The respective samples are restricted to individuals aged 15 and older, who are in the labor force, and live in Uruguay. Moreover, we transform the BuscoJobs employment spell data into a panel, where we consider each job board user's situation during the month of July in a given year.

Between 2010 and 2020, 47.4% of the applicants in our data are women, which compares well with the Uruguayan labor force (45.7%). However, 71.8% of the applicants are located in the Montevideo area, compared to only 41.6% of the overall labor force. The overrepresentation of the Montevideo area in the BuscoJobs applicants' database is likely related to the different job-search methods used in large metropolitan areas, compared with smaller urban areas or rural areas. This is consistent with the previously discussed finding of the higher relevance of print media and word-of-mouth communication in rural areas (ILO, 2020). According to the household survey, with 68.2%, a comparatively large share of those reporting to have used the internet to find a job, indeed lived in the Montevideo area during our period of analysis.

The applicants' data also contain a disproportionately high share of younger applicants (Fig. 2). This discrepancy is most pronounced among individuals aged 20–24 and 25–29, who taken together account for 48.5% of BuscoJobs users, compared to 22.3% of the Uruguayan labor force. More similar representations are observed for the age category 35–39 that accounts for around 11% of BuscoJobs users and the national labor force. The tendency is inverted as older age categories are considered.

The larger share of young portal users is intuitive given their stronger familiarity and use of IT tools, compared with older workers. It can also be rationalized by younger workers entering the labor market for the first time, and thus searching more actively for employment. For comparison, Marinescu and Skandalis (2021) report an average age of 31.8 for individuals using the online job-search platform of the French public employment services. This is very similar to the average age we find (31.1 years). Other studies directly focus on young individuals, perhaps because these are frequent users of online job-search engines when applying for jobs (Barbarasa et al., 2017; Kureková & Žilinčíková, 2018).

When looking at formal qualification levels, we find that the overall qualification structure compares well between applicants on the BuscoJobs portal and the national labor force (Fig. 3). Large shares of both populations report having completed neither a technical and vocational degree nor a university degree. Nevertheless, this is less likely among applicants (68.1%) compared with the national labor force (74.5%). Applicants are slightly less likely to have completed a technical and vocational degree, while being more likely to have obtained university degrees. Taken together, the findings indicate that applicants have above-average qualification levels.

In terms of the occupational distribution, Fig. 4 shows a certain divergence when comparing BuscoJobs applicants' employment spells with the national distribution. Notably, a disproportionately large share of the BuscoJobs users was employed in a clerical support profession (35.2%), while none of the users indicated any work experience in an agricultural profession.

This suggests that when analyzing the applicants' data, it may be appropriate to focus on certain groups of workers that are well represented without making inferences to the whole population and/or to use weighting techniques to improve representativeness. As a matter of comparison, while many studies exploiting online applicants' data do not provide information on individuals' occupations, Marinescu and Rathelot (2018) show that there is significant overlap ‒ but also some divergence ‒ between the online data from the job board CareerBuilder and national survey data for the United States.²⁴

Yet, while the applicants' dataset is not fully representative of the national occupational distribution, it has several comparative advantages. The dataset includes reasonably large samples for occupations requiring intermediate or lower levels of formal qualification, such as services and sales workers, clerical support workers, and even workers from elementary occupations. This suggests that the data can be used for analyses that extend beyond highly qualified workers, which one might not have thought at first glance.²⁵ Moreover, as mentioned above, the rich and granular information – in particular the information on tasks performed by workers – is an asset to obtain a better understanding of skills dynamics.

We are also able to document that the occupational biases are stable across time (Fig. 5). A primary concern when using these data is whether the representatives of the sample change over time; as this would be a threat to internal validity of any analysis exploiting a temporal dimension (Deming & Noray, 2020; Hershbein & Kahn, 2018). Fig. 4 shows that this is not the case. Based on Hershbein and Kahn (2018), the x-axis illustrates the BuscoJobs share in occupations in 2010 minus the share in the same occupation and year in the Uruguayan Household Survey, and the y-axis illustrates the same percentage point difference for each year from 2011 to 2020. Darker shades correspond to earlier years. The 45-degree line shows occupations where the shares between BuscoJobs and the household survey, did not change from 2010. The only group clearly deviating from this line are services and sales workers, where the year of 2010 seems to represent an outlier.

BuscoJobs is one of the leading job boards in Uruguay. According to the provider's calculations, it captures around 50% of online vacancies in Uruguay, providing information that is detailed and reliable. Vacancies are recent, have high turnover, and many of the fields available for each vacancy are complete (see Di Capua et al., 2020). As discussed before, the jobs ads posted are unique and are clean from repetitions, which is not a given among job portals. As other job portals and job aggregators in Uruguay post the same vacancies several times, it is estimated that the effective coverage of BuscoJobs is in fact closer to 60% of online vacancies.

Our data contain almost 87,000 vacancies, posted by more than 6,500 firms. As shown in Fig. 1 above (line with squared markers), the number of vacancies posted in the job board has increased constantly to reach a peak in 2014. While the observed numbers have remained at a high level also over the following years, the 2020 drop in the job postings might reflect the economic slowdown induced by the COVID-19 pandemic. It is important to note that one job posting can describe several vacant positions. In fact, 81.9% of the job advertisements posted on the job board are associated to one vacancy only, while 8.6% are associated to two vacancies, and 9.5% are linked to three or more open vacancies. Throughout the article, we conduct the analysis at the level of job postings.

For the assessment of the representativeness of the BuscoJobs online vacancy dataset, we again rely on the comparison with the national household survey data and present averages over annual results for years 2020 to 2010.²⁶ As before, the household survey analysis focuses on individuals aged 15 and older, who belong to the labor force. We also restrict our attention to individuals and vacancies located in Uruguay. Our comparison is indicative only, because vacancy data depend on turnover rates in the labor market and these may differ across regions, occupations, and economic sectors. In contrast, the household survey data represent a snapshot of workers' characteristics in a given moment. Nevertheless, this comparison is the best possible way to assess the representativeness of the BuscoJobs vacancy data.²⁷

We find that 43.0% of the vacancies are located in the capital of Montevideo, compared to 41.6% of the overall labor force.²⁸ Moreover, BuscoJobs online vacancies include a disproportionally high share of professional categories, and underrepresent occupations requiring lower formal qualification levels (Fig. 6). With 84.4%, the broad ISCO-08 categories 2–5 (i.e., professionals, technicians and associate professionals, clerical support workers, and service and sales workers) dominate in the BuscoJobs vacancy data. This suggests that the data allow for meaningful analysis especially of these groups of workers, which also cover a significant share of national employment (51.5%). Elementary occupations, which typically require low formal qualification levels, are underrepresented in the BuscoJobs data. Yet they still account for 7.3% of the vacancies (compared with 19.1% in the household survey). Furthermore, skilled agricultural occupations are absent from the BuscoJobs online vacancy database, while they account for 4.8% of national employment.

Overall, these patterns are not surprising as national employment distributions across occupations deviate in similar ways from other online vacancy sources studied in the literature. This includes the BurningGlass data, which capture vacancies for professional jobs in a more comprehensive way than other types of vacancies (Deming & Kahn, 2018; Hershbein & Kahn, 2018; ILO, 2020). We also note that the BuscoJobs vacancy data nevertheless cover a meaningful number of jobs requiring medium and even low formal qualification levels. Similar conclusions arise from the distribution across industrial sectors (see Table A1 and Fig. A1 in Appendix A).

As discussed above, we again look at changes in the representativeness of the BuscoJobs vacancy data over time to ensure the internal validity of temporal analyses. Reassuringly, Fig. 7 illustrates no major changes in the occupational distribution over time, when comparing it with the same distributions from the Uruguayan household survey. The only exceptions are again services and sales workers in 2010.

We next describe our method for creating skills variables (Section 4.1). Then, we evaluate the method's performance in four ways: first, by assessing the share of vacancies and applicants' employment spells that the model is able to classify (Section 4.2); second, by evaluating the relevance of relying on various sources of information to devise our taxonomy (Section 4.3); third, by assessing the significance of employing our approach instead of using US O-NET data to study skills dynamics in Uruguay (Section 4.4); and fourth, by showing an application of our approach to assessing skills dynamics (Section 4.5).

To create the skills variables in the BuscoJobs data, we employ a text-mining model that combines text preprocessing routines (in Python) from NLP with a rule-based classification method. The rules emanate from the taxonomy and the keywords and expressions that characterize each of the 14 skills subcategories that we predefined in Section 2.²⁹ Extended through pertinent synonyms, our model consists of 741 unique skills (i.e., keywords and expressions).³⁰ We define a skills subcategory as present in the vacancy or job spell of applicants whenever we identify at least one of its associated unique skills. We also coded related variables capturing how many unique skills pertain to each observation, as a proxy for skill intensity.

For the creation of the skills variables, we decided to rely on the open text-variables from the BuscoJobs data. For the vacancy data, we use the job title and vacancy description, while for the applicants' data, we focus on the applicants' description of each job spell. Compared to other variables in the database, the open-text variables contain the most detailed information on the skills demanded and supplied. These variables are available for almost all vacancies (99.9%) and most of the applicants' job spells (68.5%), in a functional manner as they contain a comparatively low share of missing or meaningless values. In addition, these variables are typically present in similar data sources and thus will allow to replicate our methodology using different databases.

Since we rely on free-text descriptions, we then preprocessed the text variables using NLP routines, with the aim of distilling information that facilitates the efficient mapping between the skills taxonomy and BuscoJobs data. The text data preprocessing involved translating the keywords and expressions into Spanish, tokenization, text normalization, and lemmatization of the skills variables and free text in the applicants' and vacancy databases.

We proceeded to create the skills variables, coding indicator variables for each of the 14 subcategories of skills, which take the value of one whenever a relevant keyword or expression was identified in the BuscoJobs data. We then refined the variable creation using synonyms through an automated web scraping method that targeted the website www.wordreference.com. Finally, we scrutinized all synonyms manually and excluded those whose meaning would have caused misleading classifications. This was especially true for some synonyms in the manual skills category, which might have mistakenly captured managerial activities. For example, “solucionar” is most relevant in the context of finding solutions but was identified as a synonym of “reparar” (“to repair”). To facilitate replications of our methodology, we provide details on its implementation, as well as the full set of initial keywords and expressions and additional synonyms in Appendix B.

In Table 2, we present the number of vacancies and applicants' job spells, respectively, that we could classify, and the corresponding number of skills assigned.³¹ Using only the initial keywords and expressions, we assigned on average 0.87 of our 14 skills subcategories to each applicant-job spell observation (column (1)), and 2.52 to each vacancy (column (3)). The model is thus able to assign skills to 53.0% of the applicants' job spells and 86.5% of vacancies (columns (1) and (3), inverse of the share with 0 skills). Once we additionally consider synonyms, the performance increases noticeably (columns (2) and (4)). The model is then able to assign an average of 1.44 skills subcategories to each applicant's job spell and 3.86 skills subcategories to each vacancy. The use of synonyms significantly increases the number of applicants' job spells that can be assigned a skill to 63.6%. While we lack a direct benchmark for job spells due to the limited availability of supply-side data on job boards and job aggregators, we believe 64% allows for a meaningful classification of skills, supporting credible analyses. This is especially pertinent given the substantial heterogeneity in the quality of information within the free-text descriptions used for coding the skills variables, particularly, the self-reported text from applicants, which does not follow any standardized format. In the case of vacancies, using synonyms even increases the classified share of observations to 94.3%. Overall, our results leave us optimistic that our methodology can be replicated in similar data sources.

Table 2 therefore allows drawing two central conclusions. First, the text-mining model performs significantly better when using a combination of keywords and synonyms to capture skills. This implies that the use of synonyms is a decisive step for implementing a skills taxonomy in online labor intermediation data. Second, the vacancy data tend to be richer than the applicants' job spells in terms of the information provided that can be used to capture skills. We note, however, that the statistics on applicants refer to a given employment spell. Once we aggregate skills over workers' employment biographies, as done in Appendix Table A2, the average number of skills subcategories increases to 2.74 (standard deviation of 2.71) per person when we use the method that is based on keywords and their synonyms. Also, the share of persons without any assigned skills subcategory is only 25.0%, compared to 36.4% of all employment spells that cannot be classified.

The primary reason why some observations cannot be classified is the brevity of the open text. The unclassified employment spells have only 5.5 words on average (i.e., with many not containing any meaningful description), whereas the number of words significantly increases for employment spells with one or several mapped skills subcategories (Table 3). This may in part be related to the fact that some applicants include descriptions that are largely self-explanatory. For example, there are observations that only include the description “biology teacher,” presumably because the tasks and skills of a biology teacher are considered as being common knowledge.

Fig. 8 further highlights the importance of including synonyms in the variable-coding process. By definition, the number of classified observations increases for each of the 14 subcategories. Interestingly, this effect is not uniform across categories. In the case of the applicants' data, synonyms make a substantial difference for capturing cognitive skills (narrow sense), social skills, people management skills, and finger-dexterity skills. For the vacancy data, synonyms are additionally decisive for significantly increasing the number of identified character skills. Moreover, cognitive skills (narrow sense), customer service skills, people management skills, social skills, and finger-dexterity skills are the five categories that appear most often in the applicants' data (i.e., when accounting for synonyms). These subcategories, likewise, play the greatest role for the vacancy data, although there are some differences in the ordering. In addition, character skills, computer skills, and software skills feature prominently in the vacancy data, but this trend is less obvious in the applicants' data.

For future work, complementing our text-mining model with a prediction model could have potential. However, a prediction model requires that a proportion of the data are classified using an external source of categorization (i.e., one that is not linked to our taxonomy or its implementation). In this way, the already classified part of the data can serve as a benchmark to train the prediction model against which the results obtained from our coding of the variables can be compared to. For example, this benchmark can be a classification carried out by a group of experts, based on clearly defined criteria and a review process of the classification results. Additionally, this benchmark must be long and varied enough to “teach” the computer and train the model properly, thus making it a demanding exercise. The lack of such a benchmark for skills variables currently prevents us from employing a prediction model. In the future, we will explore whether it might be possible to use a benchmark classification from O-NET Uruguay, which is currently in its pilot phase. We already followed a similar approach for coding two-digit ISCO-08 occupations for both the vacancy and job applicants' data, which is detailed in Appendix B.

We now investigate the usefulness of combining a comprehensive array of studies and sources in formulating our initial skills taxonomy (see Section 2). Alternatively, researchers could rely solely on prominent studies using online vacancy data. Table 4 reveals that the majority of identified keywords and expressions indeed stem from such studies (72.3% for the vacancy data and 60.6% for the applicants' data).³²

However, supplementary keywords and expressions from non-online data sources also hold significant relevance, contributing to 17.7% of identified unique skills in vacancies and 30.6% in applicants' job spells. Additionally, the O-NET Uruguay source adds 10.0% for vacancies and 8.8% for applicants, particularly enhancing our ability to capture manual skills, which remain comparatively more important outside of Europe and the United States. This underscores the effectiveness of our approach in achieving a comprehensive taxonomy by combining a diverse range of seminal sources.

We also assess how our approach compares to one that would have relied on imputing US O-NET data at the occupational level. For this purpose, we map US O-NET skills categories to our taxonomy. For both the US O-NET and the BuscoJobs applicants' data, we then compute scores that capture the relevance of cognitive, socioemotional, and manual skills at the one-digit occupational level, normalized to add to 100.³³ This comparison yields clear differences between the country-specific BuscoJobs results and the US-based O-NET results (Fig. 9). Across occupations, manual skills matter comparatively little in the composition of occupations of the United States. In contrast, manual skills play a larger role in the Uruguayan data, consistent with the expectation that manual skills matter more outside of high-income economies.³⁴ This is especially the case for plant and machine operators and assemblers, elementary occupations, and crafts and related trades workers. Accordingly, these occupations tend to have slightly lower scores in the socioemotional skills category and, especially, the cognitive skills category. Indeed, cognitive skills account for around 47%–56% of the skills composition of all occupations in the US, while in Uruguay, it varies from close to 60% for professionals to around 17% for plant and machine operators.

One consideration in this comparison is that the O-NET data are representative of occupations in the US, while we have noted in Section 3 that the BuscoJobs data are not fully representative of the Uruguayan labor market. However, even for occupations with high coverage in BuscoJobs, such as clerical support workers, there are discrepancies between the O-NET and BuscoJobs results: The Uruguayan data suggest a lesser prominence of cognitive skills and a greater emphasis on manual skills. Furthermore, our findings are in line with existing literature that uses the newly collected O-NET data in Uruguay (available for 22 occupations as part of the pilot phase) to analyze the composition of occupations. Velardez (2021) compares the occupational composition between the US and Uruguay using their respective country-specific O-NET data and finds that the largest differences between occupations pertain to the skills and abilities components. These findings resonate with prior research (Lewandowski et al., 2020, 2022). In summary, our findings and those from the existing literature clearly confirm one of the premises that motivated this study, namely that it is important to employ country-specific data when assessing skills dynamics beyond Europe and the United States.

We end this section with an illustration of how our approach can inform the analysis of skills dynamics. As such, Fig. 10 below illustrates the evolution of the supply of and demand for the skills subcategories that appeared more prominently in our analysis (see Fig. 8). The applicant trend presents the proportion of job applications containing keywords or expressions (including synonyms) related to each of the referenced skills subcategories, while the vacancy trend depicts this analogously for job vacancies. In general, we see a consistent gap between the proportion of vacancies demanding a certain skill, and the proportion of applicants with the stated ability to be using this skill in their current or during a previous employment spell (with the exception of customer service skills in 2010). This gap widens for all skills subcategories over the reference period, with the most dramatic increase realized by cognitive skills, social skills, and customer service, which rose from 23 to 45pp, 14 to 40pp, and −3 to 23pp, respectively. Thus, while the proportion of vacancies demanding certain skills has grown over time, this increase has not been met by applicants explicitly signaling the possession of these skills in their online CV.

The evolution of these gaps coincides with an increase in the number of words used to describe job vacancies over time and a concomitant decrease and subsequent stagnation in the number of words used by applicants to describe job spells over time. This notable contrast is presented in Fig. 11. The relevance of text length is corroborated by Fig. A2 in Appendix A, which presents a subset of the analysis in Fig. 10 for the period 2014–2017, where the average number of words used in both job vacancies and applications remains constant and the gaps between indicated skills supply and demand do not change significantly. The relationship between the demand for skills and the number of words used in job vacancies could be related to employers' need to reflect their increasing, more holistic demands required from prospective employees through an increase in the length of job postings. Delving into the intricacies of this relationship could be a promising direction for further research.

Many countries outside of Europe and the United States currently lack longitudinal data on skills, despite the importance of the topic for policymakers and academic debates. In this article, we assess whether data from online job vacancies and applicants' profiles, which are increasingly becoming available, can be a suitable resource for studying skills dynamics.

Drawing from the literature from the social sciences, in particular from labor economics and psychology, we derive a comprehensive yet succinct skills taxonomy. This taxonomy is designed to be adaptable to individual country-contexts and applicable to online data. It comprises the three broad categories of cognitive, socioemotional, and manual skills, along with 14 more detailed subcategories, which are defined in terms of keywords and expressions. In comparison to existing seminal skills taxonomies for online data, our taxonomy goes further by including manual skills and expanding the conceptual underpinnings surrounding socioemotional skills. This broader approach allows us to encompass a more comprehensive spectrum of workers, which represents a critical consideration for any labor market, but is particularly significant outside of the United States and Europe.

Based on a text-mining model that combines text preprocessing routines from NLP with a rule-based skill classification method, we then develop a methodology that serves as a tool for researchers and practitioners to classify skills variables in raw text data. We implement this methodology, exploiting data from the Uruguayan job board BuscoJobs. We are able to classify skills requirements and skills that applicants possess for 94% of job vacancies and 64% of applicants' employment spells. This goes beyond our initial expectations, especially when considering that the implementation is based on free-text descriptions that do not necessarily follow a standardized format.

We conclude that data from online job vacancies and applicants' profiles are a promising source for analyzing skills dynamics, even in countries where job boards and job aggregators have not traditionally been prevalent. This is a relevant finding, since these data capture country-specific developments, are available in numerous countries, and offer granular and longitudinal information often for both labor demand and supply.

However, such data may not fully represent a country's entire labor force, which might require weighting techniques and/or an analytical focus on specific labor market segments. The BuscoJobs data are no exception to this trend. Yet, these data effectively capture job vacancies and job seekers across a range of educational levels, including lower and especially intermediate educational levels in addition to highly qualified labor. Moreover, representativeness biases in the data appear not to fluctuate substantially across time.

Most importantly, BuscoJobs and similar sources of data enable the study of skills dynamics in countries where such analysis would otherwise be challenging due to the limited availability of alternative data sources. Indeed, our findings reveal considerable differences in the skills composition of occupations when compared to what we would have obtained had we relied on O-NET data from the United States. We also provide a first illustration of a possible analysis by assessing the evolving trends in the supply and demand for specific skills. The analysis illustrates a consistent gap between the skills sought by job vacancies and those explicitly indicated by applicants, which increases over time. This might be related to employers' more complex demands, reflected by an increase in the length of job postings, emphasizing the need for further research to explore this relationship in depth.

Going forward, this conceptual and methodological effort, the first carried out outside Europe and the United States, opens new doors for future research on skills dynamics. Such initiative can delve more deeply into empirical questions related to the role of skills in fostering transitions to better jobs and enhancing the resilience of both firms and individuals when these are confronted with fundamental labor market transformations. Moreover, there is an opportunity to explore the skills composition of occupations and the dynamics of skills within occupations at the national level. This research can also shed light on skills mismatches between labor demand and supply, the role of skills in job application behavior and the impact of shocks and regulations, as well as test for the widespread use of other classifications of skills inspired from high-income countries.

We extend our heartfelt gratitude to those who have contributed to this work. Fidel Bennett, a co-author of a previous version, played a pivotal role in the initial development of our text-mining approach. Valuable contributions were also made by Willian Adamczyk and Evgeny Gushchin. We are grateful to Claudia Abdallah, Johannes Brehm, Mariano Calderón, Angela Doku, Joana Duran-Franch, and Henry Stemmler for their outstanding research assistance throughout various stages of this article's production. Our sincere thanks go to Marcos Aguiar, Diego Alanis, Jorge Eguren, and Lucas Ng for providing data and addressing our enquiries. We are equally grateful for the invaluable insights provided by Janine Berg, David Deming, Gonzalo Graña, Cornelius Gregg, Piotr Lewandowski, Clemente Pignatti, Olga Strietska-Ilina, Michael Stops, Bolormaa Tumurchudur Klok, and Fernando Vargas. The responsibility for opinions expressed in this article rests solely with its authors, and publication does not constitute an endorsement by the International Labour Office of the opinions expressed in it. The authors have no declarations of interest to disclose.

This appendix first presents the methodology used to create skills and occupations variables. Section B.1 provides a comprehensive description of the text-mining methodology used to implement the skills taxonomy exploiting the open-text variables available in the BuscoJobs data. In the context of the vacancy data, these variables furnished information pertaining to job titles and vacancy descriptions, while for applicants' data, they provided descriptions of each job spell. Each of these variables required the open-text data to be structured in a manner that could be processed and interpreted by computer algorithms; as well as classified according to our pre-defined taxonomy. Section B.2 outlines the creation of ISCO-08 occupational variables. It starts by detailing a process of text processing similar to the one detailed in Section B.1. It then explains how we combined this process with a predictive model that employed BuscoJobs' existing classification as a benchmark.

The methodology used to create skills variables consists of a text-mining model/approach. Since we rely on free-text descriptions, we preprocessed the text variables using NLP routines, with the aim of distilling information that facilitates the efficient mapping between the skills taxonomy and BuscoJobs data. We then combined this approach with a rule-based skill classification method, where the rules emanate from the predefined taxonomy that we have described in the main text (Section 2). This involved the following steps:

• Translation of keywords and expressions: As explained in Section 2, our skills taxonomy initially defines 14 subcategories through keywords/expressions that we identified from the existing literature on skills dynamics. Since our online data are in Spanish, we first translated the keywords/expressions from English. Appendix Table B1 above provides this translation.

• Tokenization: We started the text preprocessing by this pivotal initial step, which involves the systematic dissection of textual data. In this process, the open text of the relevant variables (vacancies and applicants' job spells) was deconstructed into discrete units known as “tokens.” These tokens correspond to individual words or subword elements broken down by blank spaces, including punctuation marks or segments of words, which were later removed. By transforming raw, unstructured text into a structured format, tokenization allowed for the following computational analysis.

• Text normalization: As is typical for data analyses of this kind (see, e.g., Gentzkow et al., 2019), we normalized the text through text-mining techniques, mainly using the Natural Language Toolkit (NLTK) library in Python. This step was similarly performed for the relevant variables in the BuscoJobs data and for the list of keywords/expressions in the skills taxonomy. It included: (1) lowercasing of capital letters; (2) “unidecoding” to simplify characters and delete accents that are used in Spanish (for example, changing “á” to “a”); (3) eliminating all characters that are not actual text (for example, “%,” “$,” “&,” “NaN” or “Null” values, which are the common representation of missing data in Python); (4) eliminating stop words, which are commonly used prepositions and conjunctions that do not provide useful information by themselves (for example, the Spanish equivalents of “the,” “a,” “of,” and “in”); and (5) eliminating other concepts that do not add value, such as names of months, days, cities, and countries and single letters or words with less than three letters.¹ Exceptions, however, were the three-letter words “sql” and “web” that identify unique skills.

• Extended keywords and expressions in the skills taxonomy: In order to capture conjugations of words related to the same concept (e.g., when referring to “collaboration,” this could be done through the noun or through the verb “collaborate”), we relied on “lemmatization,” which replaces variations of words with a common lemma. This is equivalent to extending the set of words, so they capture their various forms of expression (i.e., as a noun or verb, feminine or masculine forms, singular and plural forms, etc.). To give one example, the Spanish words “estadística,” “estadístico,” “estadísticas,” “estadísticos” are all replaced by the form “estadistic” (i.e., “statistics”). The lemmas can be either keywords or words that are part of an expression.

• N-gram creation in the skills taxonomy: N-grams involve creating text segments by combining sequences of one or more words (tokens). These segments can be 2-grams, 3-grams, and n-grams (where “n” denotes the number of words included), and they are typically represented as separated elements in a list. To give an example, we had initially included in our skills taxonomy the expression “predictive analytics” as one of the terms characterizing “cognitive skills (narrow sense).” We manually identify and consider other (Spanish) versions of the same concept, such as “Análisis Predictivo de Datos” or – after applying the text-normalization described before – “analisis predictivo datos.” Once we carry out the n-gram creation, a total of seven elements are added to our list: the one-grams “analisis,” “predictivo,” and “datos”; the 2-grams “analisis predictivo,” “analisis datos,” and “predictivo datos”; and the original 3-gram “analisis predictivo datos.” In this way, we give our classification algorithms greater power to capture context within the free-text descriptions of vacancies and applicants' job spells. This process is carried out using the NLTK library but requires a manual revision of the results. For example, for “programming language” (included in the skills subcategory of “software (specific) skills and technical support”), we kept the token “programming” but deleted the token “language,” as the latter does not pertain to software skills and would have induced an error in the classification. The number of keywords/expressions in our taxonomy was still small enough to allow for such manual processing within a reasonable amount of time.

• N-gram creation in the vacancy and applicants' data: We similarly transformed the relevant BuscoJobs variables, using a process based on the NLTK library, to distil n-grams combinations from the data. We imposed two restrictions on this automated process: (1) We identified combinations of tokens, 2-grams, and 3-grams, but neglected combinations of four and more words. We imposed this restriction as our taxonomy includes normalized expressions with a maximum length of three words (such as the above example of “analisis predictivo datos”). (2) We kept the order of words as they appear in the original text description, as otherwise the task would not be manageable for a regular server to process.

With the text organized, we proceeded to create the skills variables:

• Implementing the skills taxonomy in the data: For the creation of the skills variables, we coded indicator variables for each of the 14 subcategories of skills, which take the value of one whenever a relevant token or n-gram was identified in the BuscoJobs data. We also coded related variables capturing how many times a relevant keyword/expression appeared in the data, as a proxy for skill intensity. Here, we did not count repetitions of the same keyword/expression, but considered each keyword/expression only once per observation.

• Refined variable creation using synonyms: We further expanded the initial list of keywords by also accounting for their synonyms. For this, we used an automated web scraping method that targeted the website www.wordreference.com and recorded, for each initial keyword from our taxonomy, direct, or first-order, synonyms.² Once we had identified these additional keywords, we again performed the steps above with the extended keyword list (mainly the “lemmatization” process) and recoded the skills variables.

• Manual correction of synonyms: We scrutinized all synonyms manually and excluded a few that were duplicates or whose meaning would have caused misleading classifications. This was especially true for some synonyms in the manual skills category, which might have mistakenly captured managerial activities. For example, “solucionar” is most relevant in the context of finding solutions but was identified as a synonym of “reparar” (“to repair”). For a similar reason, we manually changed “controlar” (“to control”) to “controlar máquinas,” “controlar aparatos,” and “controlar artefactos” (“control machines” etc.). We also added relevant synonyms that we identified based on our work with the BuscoJobs data and previous knowledge, including accountant software programs that frequently appear in vacancy texts and are relevant for capturing financial skills. Meanwhile, software names or technical words did not have synonyms attached. To facilitate replications of our methodology, we provide the full set of initial keywords and additional synonyms in Appendix Table B2.

In order to categorize the ISCO-08 occupation variables, we followed an analogous approach using the same open-text variables that were used to classify skills, and subsequently preprocessed the data using an NLP model. For the classification, we utilized the four key variables that contained the most valuable open-text information: job title (vacancies only), job description (vacancies and applicants), level of education (vacancies and applicants), and hierarchical level (vacancies and applicants).

The preprocessing of the variables followed the same stages described above, namely translation; tokenization; text normalization; and stemming. The application in the two datasets is slightly different, given that they differ in the available data. For the vacancy dataset, we applied the preprocessing also on the job title variable, and for both the vacancy and applicants’ dataset, the preprocessing was applied on the job description variable.

Following the preprocessing, we proceeded with the classification in three steps:

1. a classification based on a rule-based model and NLP, similar to that carried out in our classification of skills;

2. a classification based on a predictive model to classify ISCO-occupations at one-digit level, using an existing classification carried out by BuscoJobs, as a benchmark;

3. a second machine learning classification to create ISCO-occupations at a two-digit level, based on the one-digit predictions.

Step (1) – Rule-based model and NLP

As part of process (1), we created a dictionary of keywords and expressions that appear with high frequency in the data, using the free text of the subset of both the vacancies and applicants' job-spells variables, to which the corresponding ISCO-08 code was already assigned by BuscoJobs. The keywords and expressions used in this dictionary originate from the official ISCO-08 occupations classification.⁴ In specific situations, this required manual classification. This output provided us with a set of rules with clear indications on how to classify an ISCO code. Then, this dictionary was paired with the preprocessed texts achieving the classification based on the text matches that is available in Appendix Table B3.

In order to improve the classification performance, we used two auxiliary variables to clarify ambiguities from the classification: first, we used education level to differentiate between ISCO categories 2 and 3. Here, a professional or higher-education level would be classified as 2, and any other educational level as 3. Second, we used hierarchical level to identify jobs at Director or Managerial level (ISCO 1). When the rule-based model did not assign a nonambiguous classification, it used the hierarchical variable for confirmation.

Step (2) and (3) – Machine learning classification

After performing step (1), we used the approximately 5,000 observations already labelled at the four-digit level by BuscoJobs in each of the databases. For this process, we first applied a machine learning algorithm to classify observations that were not classified by step (1), or that diverged substantially from those labelled by BuscoJobs. Second, after performing several tests, we decided to combine the already labelled observations with the output from step (1) to train a predictive model to assign ISCO-08 occupations to job titles and combinations of additional open-text variables. This step (2) predicted the ISCO classification at the one-digit ISCO level.

Then, we moved to step (3) that used the input of step (2) in another machine learning process to classify observations at a two-digit level. In the case of the vacancy data, the variables used for the prediction were: job title, job description, CIIU, education level, sex, the rule-based model's predicted ISCO group, and first-stage ML predicted ISCO group. For the applicants' data, job title and CIIU were not available.

Ultimately, we tested three classification models: Random Forest, Support Vector Machine, and Gradient Boosting. A comprehensive strategy using training (two thirds of the data) and test samples (one third) allow us to test hyperparameters and various cross-validations for each combination. Based on this, we chose Gradient Boosting to code the one- and two-digit ISCO-08 occupations in the vacancy data, and Random Forest for the applicants' data.

The strategy resulted in the successful classification of the 87,030 observations in the vacancy dataset at one-digit level, and 94.8% at the two-digit level. Among the 1,230,588 observations in the applicants' dataset, only 70% have a text description. All of them were classified at one-digit level and 97.8% at the two-digit level.