DRC: Democratic Republic of the CongoMINESU: Ministère de l’Enseignement Supérieur et UniversitairePCA/CBA: Approche par Compétences (Competency-Based Approach, CBA)LMD: Licence—Master—Doctorat

The relationship between higher education and employability has long been central to human capital theory, which argues that education increases productivity and thus labour-market value. However, in the context of developing economies, the translation of educational attainment into employment is far from automatic. The growing complexity of production systems, the expansion of the digital economy, and the shift toward knowledge-intensive services have created new skill requirements that traditional academic programmes often fail to meet [1][6].

Employability is no longer determined solely by formal qualifications but by a mix of technical, cognitive, and transversal competencies, such as problem-solving, communication, adaptability, and digital literacy [3]. Consequently, the analysis of graduate employability must go beyond simple enrollment or unemployment statistics and focus on the alignment between education and actual job requirements.

In Sub-Saharan Africa, several studies have highlighted persistent skill mismatches between university training and labour-market expectations. The African Development Bank and IFEF/UNESCO emphasize that most universities continue to operate within discipline-based systems with limited integration of soft and digital skills [7]. Employers across the continent report difficulties finding graduates with the competencies required for emerging sectors such as ICT, finance, renewable energy, and logistics.

In the Democratic Republic of Congo (DRC), this challenge is particularly pronounced. Research by Bwanga and Arionzi reveals that while the number of graduates has increased substantially since the implementation of the LMD (Bachelor–Master–Doctorate) system, their transition to formal employment remains weak. The authors attribute this gap to curricula that emphasize theoretical knowledge over practical application and to insufficient collaboration between universities and the private sector. Similarly, Christelle and Moïse note that employers frequently report deficits in language, analytical, and ICT skills among recent graduates, despite their academic credentials [8].

Most empirical studies on skill mismatch in Africa rely on survey data and descriptive analyses of graduates’ employment status. While informative, such approaches often remain subjective and limited in scope. Only a few recent contributions have applied quantitative text-based techniques to measure how well educational content corresponds to occupational needs. The OECD proposed a semantic mapping framework comparing skill taxonomies across education and employment databases, showing that semantic similarity can serve as a reliable proxy for skill adequacy [6].

In the African context, however, such methods remain under-used. Existing research generally focuses on employability perceptions rather than the linguistic content of training and job descriptions. To date, no study in the DRC has combined natural language processing (NLP) with probabilistic modeling to quantify employability outcomes based on skill alignment. This study therefore fills a critical methodological gap by integrating semantic similarity metrics (via the spaCy model) into a logistic regression predicting the probability of graduate employability.

The literature consistently points to a need for evidence-based frameworks capable of translating qualitative observations about skill mismatch into measurable indicators. By adopting a semantic–probabilistic approach, this study provides a reproducible tool for diagnosing how well academic programmes align with market requirements. It builds on international advances in text analytics while adapting them to the Congolese context, where statistical data are scarce but textual information (curricula and job postings) is abundant.

The main contribution of this work is therefore twofold:

1) It quantifies the semantic proximity between university competencies and job-market skills using large-scale textual data;

2) It models the probability of employability as a function of that semantic alignment and the academic domain.

This approach bridges the gap between educational research and labour-market analysis, offering practical insights for policymakers seeking to modernize higher education and enhance graduates’ professional integration in the DRC.

Drawing from the semantic–probabilistic approaches identified in the literature, this study proposes an operational framework adapted to the Congolese context for assessing the adequacy between university training and labour-market requirements. The methodological logic follows six interconnected stages combining semantic analysis and probabilistic modelling.

Step 1:Corpus design and data integration.

The protocol begins with the systematic collection and structuring of two complementary corpora: university curricula officially published by the Ministry of Higher and University Education (MINESU), and LinkedIn job offers (2023-2025) representing real market demands. This dual-source design ensures representativeness of both the academic supply and the employment demand for skills [4][8].

Step 2:Text preprocessing and semantic normalization.

All textual data are cleaned and standardized to manage linguistic variation between French, English, and local usage. This involves lemmatization, translation, and terminological harmonization through the spaCy library, following recommendations from lexicometric studies on bias reduction and linguistic clarity [9]-[11].

Step 3:Semantic similarity measurement.

University and job-market skills are transformed into numerical vector representations and compared using spaCy’s word-embedding model. The semantic proximity between the two corpora provides a quantitative indicator of skill alignment, inspired by previous educational alignment research [9][12].

Step 4:Construction of alignment indicators.

For each field of study, an overall correspondence index is produced by averaging individual similarity scores. Competencies are classified as aligned when their similarity exceeds an empirical threshold of 70 %, corresponding to the upper confidence boundary observed in the dataset [13][14]. Missing or unaligned skills are then ranked by frequency and priority for curricular revision [15].

Step 5:Probabilistic modelling of employability.

The degree of alignment is integrated into a logistic regression model to estimate the probability of employability by academic domain. This approach quantifies the effect of skill adequacy on labour-market insertion, in line with established econometric and competency-based frameworks [1].

Step 6:Validation and feedback.

Model robustness is assessed through cross-validation and performance metrics (AUC, accuracy, F1-score). The results are compared with recent national employability surveys and validated through discussions with stakeholders from universities and private-sector organizations [16][17]. Each of these steps is documented or inspired by uses and feedback identified in the literature (lexicometric analyses, scale studies, econometric models applied to the labour market), which ensures the methodological robustness of the proposal, [1][9][11][16].

3.2.1. Educational Programmes (Provision of Skills)

Collection of official curricula published by the Ministry of Higher and University Education (MINESU).

Extraction of targeted competencies in the 8 training areas defined by the ESU (Health Sciences, Engineering Sciences, Economics and Management Sciences, Social and Political Sciences, Education Sciences, Arts and Letters, Agronomic Sciences, Basic Sciences).

3.2.2. Labour Market (Demand for Skills)

Scraping of 1560 offers published on LinkedIn for the DRC (2023-2025), with the help of the official SALESQL.¹ Extraction of skills, job titles, sectors, and associated requirements.

Elimination of duplication and redundant skills.

Lemmatisation and linguistic standardisation with the help of Cy (fr_core_news_md) to standardise terms (e.g. “web development” vs. “web developer”).

Machine translation to harmonize the terms with Python’s Google library.

To evaluate the correspondence between the skills developed in university curricula and those required in the labor market, this study relies on the semantic similarity measure provided by the spaCy library.

The similarity() function in spaCy is grounded in vector-space mathematics from Natural Language Processing (NLP), allowing the representation of word meanings as high-dimensional numerical vectors [18].

3.4.1. Vector Representation of Words

Each word w i is transformed into a dense vector v i ∈ ℝ d , where d denotes the dimensionality of the semantic space (typically d = 300 ). These vectors, known as word embeddings, are produced by deep-learning models such as Word2Vec [19] or GloVe [20]. Thus, two semantically similar words are represented by vectors that are close to each other in this multidimensional space:

This approach follows Firth’s (1957) distributional hypothesis: “You shall know a word by the company it keeps.”

3.4.2. Cosine Similarity Computation

To measure the proximity between two semantic vectors v 1 and v 2 , spaCy applies the cosine similarity, defined as:

where

v 1 ⋅ v 2 = ∑ i − 1 n v 1 i v 2 i is the dot product, and ‖ v 1 ‖ = ∑ i = 1 d v 1 i 2 is the Euclidean norm.

The cosine similarity ranges from −1 to 1:

1→ identical meaning (same orientation),0→ no relation, −1→ opposite meanings [21].

3.4.3. Extension to Sentences and Documents

For larger linguistic units such as sentences or documents, spaCy computes the mean vector of all word embeddings in the text:

Then, for two competences x (university) and y (LinkedIn), the similarity is given by:

where V x and V y are the embedding vectors of both competences,

Sim ( x , y ) ∈ [ 0 , 1 ] measures semantic proximity (1 = identical, 0 = no relationship)².

An overall index of correspondence (ICG) will be calculated for each field of training:

With:

N d = number of skills in the field d ,

c i u n = 1^st university competence,

c j e m p = j^th competency from LinkedIn offers.

Notation of the Logistic Model

Let Y i ∈ { 0 , 1 } denote employability, S i the semantic similarity score (spaCy), and D i k the dummy variables for academic domains. The model estimated is:

where β 1 measures the marginal effect of semantic similarity on employability and β k the effect of belonging to domain k . All variables were standardized or dummy-coded before estimation.

Then, the logistic regression model used to estimate the probability of a student finding a job by field is defined by:

where:

X k represents explanatory variables (field of study, number of skills aligned, skills missing);

β k are the estimated coefficients.

The dependent variable Y is determined by matching LinkedIn profiles (Congolese graduates).

To dichotomize the similarity scores, a binary variable Y i is constructed as follows:

The cutoff of 70% corresponds to the upper confidence boundary under a normal approximation of the similarity scores, i.e., T = p ¯ + 1.96 σ .

Empirically, this value was observed around 0.70 in the dataset, implying that only skills with a semantic proximity significantly higher than the mean (at a 95% confidence level) were classified as aligned with labor-market demands.³

X k represents explanatory variables (field of study, number of skills aligned, skills missing), β k are the estimated coefficients.

The dependent variable Y is determined by matching LinkedIn profiles (Congolese graduates) with the corresponding job offers.

The results will be presented in the form of probabilities of insertion (%) by field of training.

The number of aligned skills is equal to the total number of skills weighted by the simultaneity rate.

where, n s a is the number of skills aligned and n s the number of skills. Sim is the similarity rate.

Indeed, as the number of aligned skills is precisely estimated by multiplying the number of academic skills by the simultaneity rate, it follows that the number of missing skills will be the complement to one of the latter. There is a very strong collinearity between these variables and the explained variable, which is employability, itself defined by the simultaneity of skills. These variables therefore belong to the same affine hyperplane. The model will thus focus solely on explaining this employability through skills (via the simultaneity rate and the domains).

A list of skills presents in job vacancies but absent from university programs will be generated for each area. A priority score (SP) will be assigned according to the frequency in the tenders:

where f j is the frequency of occurrence of competence c j LinkedIn.

Python (spacy, Pandas, Scikit-learn, statmodels, Gtrans) for the extraction, cleaning and analysis as well as the translation of texts into French with the aim of standardizing the contents.Excel (Google Sheet) for the translation of texts into French in order to standardize the contents.

After encoding and cleaning, the analytical dataset contained 103,620 valid observations and 8 variables, all in numeric format (float64 and int32). No missing or non-numeric values were detected, ensuring data consistency and reliability for statistical modeling.

A multicollinearity test was conducted using the Variance Inflation Factor (VIF) for all explanatory variables, including the semantic similarity index and the dummy-coded academic domains.

The results are summarized as follows:

Table1. Ariance inflation factor (VIF) results.

All VIF values are well below the threshold of 5 (see Table 1), confirming the absence of multicollinearity among the predictors.

This means that the semantic similarity variable and the academic domains provide independent information to the model and can be used simultaneously without risk of statistical redundancy.

A binary logistic regression model was applied to 103,620 observations to estimate the probability that a graduate is employable based on the semantic similarity between university-acquired and market-required skills, and the academic domain of training. The model converged successfully after nine iterations, with a Pseudo R² = 0.6899, indicating that nearly 69% of employability variability is explained by the predictors. The likelihood-ratio test (p < 0.001) confirms the global significance of the model.

Before regression modeling, the global correspondence between academic competencies (from the 8 ESU domains) and LinkedIn job offers (2023-2025) was measured using the semantic matching algorithm (SpaCy cosine similarity).

Table2. Summary correspondence of competences.

Table3. Average correspondence of competences.

Only 49.14% of the competencies taught in Congolese universities correspond effectively to what employers demand (see Table2). This means that one out of two skills developed in academia has no direct relevance to the labor market, confirming a persistent but measurable skills gap.

Economics and Management sciences are the most suited to the world of employment, with a 60% match rate. Conversely, Psychological and Educational Sciences is the least suitable with a ratio of 44.34 (see Table3).

Skills Missing Priorities

Textual and semantic analysis of profiles shows a recurring deficit in some key competencies. Among the most frequently absent are:

Accounting and financial management (198 occurrences).Human Resources (HR) and Engineering (132 occurrences each).Purchases and sales, procurement, marketing and customer service.Digital skills: Microsoft Office, Excel, SQL, HTML, Java, PHP, JavaScript, React.js, Laravel, etc.Languages and communication: English, professional communication, teamwork.These gaps point to the need for a realignment of university curricula to integrate more cross-cutting and technical skills directly related to employability.

The logistic regression model reveals that graduate employability in the DRC primarily depends on the degree of alignment between university-acquired skills and those demanded by the labor market.

Table4. Coefficients and statistical significance.

Pseudo R² = 0.6899; Log-Likelihood = −17,970; p < 0.001.

Semantic similarity emerges as the dominant determinant: each 1% increase in skill correspondence raises the likelihood of employment by about 70% (see Table 4). All academic domains contribute positively but with varying magnitudes—Technology, Economics and Management, and Law and Administration show the highest employability probabilities, while Health, Psychological, and Social Sciences exert more moderate effects. These findings confirm that the key challenge for Congolese higher education is not the choice of academic field, but the effective alignment of curricula with labor market skill demands.

The semantic similarity alone explains more than two-thirds of the total predictive variance, showing that employability primarily depends on the degree of match between educational content and labor market expectations.

The analysis of variable importance highlights the dominant predictive role of semantic similarity in explaining graduate employability in the DRC.

With an importance weight of 5.55, this variable alone accounts for 68% of the total predictive power of the model (see Table 5), confirming that the degree of alignment between university training and labor market skills is by far the main determinant of employment likelihood.

Table 5. Predictive performance and Importance weight.

The remaining 32% of explanatory power is distributed among the academic domains, hose effects, although positive, are secondary and relatively modest: Sciences et Technologie (8.5%) and Economic & Management Sciences (6.5%) are the most influential fields after similarity. These domains correspond to disciplines that are directly connected to the digital and managerial demands of the modern labor market.

Figure1.Confidence intervall of employability.

Psychologiques & Education Sciences (6.1%) and Health Sciences (5.1%) also contribute positively, reflecting moderate labor absorption in education and health-related professions.Legal, Political & Administrative Sciences (4.7%) present a stable but less pronounced contribution, consistent with employment patterns in public administration.Finally, Human & Social Sciences (2.1%) show the weakest influence, confirming a limited alignment between social science curricula and current job market needs.

In summary, employability in the DRC is primarily driven by the semantic proximity between academic and professional competencies (≈70%), while disciplinary orientation contributes only marginally (≈30%). This finding reinforces the conclusion that curricular alignment rather than domain specialization is the key to improving graduates’ integration into the labor market.

In the context of evaluating the alignment between university and professional skills, these results suggest that our classification model is highly reliable (see Figure 1):

It captures nearly all truly aligned competencies (high recall),and rarely labels unaligned ones as aligned (high precision).

The logistic regression model, optimized with C = 100, L2 regularization, and lbfgs solver, achieved AUC = 0.977 ± 0.001 and accuracy = 91.8%, demonstrating outstanding predictive quality.

The combination of high AUC, low misclassification rates, and consistent cross-validation performance provides solid evidence that the model is statistically robust and generalizable.

Table6. Summary of model validation and robustness.

It can therefore serve as a reliable analytical tool for forecasting graduate employability based on skill alignment and domain-specific training.

The model demonstrates high predictive accuracy (AUC = 0.977), strong robustness, and generalizability. Employability depends mainly on semantic skill alignment (≈70%), while disciplinary effects are secondary (Table 6).

The model shows that mixed similarity plays a decisive role in the probability of employability. The fields of training do not contribute significantly to this prediction, stressing that the match between the profile and the professional requirements takes precedence over the disciplinary field.

The results indicate a very strong improvement of the model compared to the null model. The extremely high null deviance (115,910) shows that a model without explanatory variables fits the data very poorly, whereas the very low residual deviance (2.61) reflects an excellent fit of the final model, suggesting that the explanatory variables account for almost all the variability in the outcome. This substantial reduction in deviance highlights the high explanatory power of the model. Moreover, the low AIC value (18.6) indicates that this strong fit is achieved without unnecessary complexity, supporting the parsimony and robustness of the model (see Table 7). Finally, convergence after 25 iterations suggests stable parameter estimation with no apparent numerical issues.

Table 7. Model quality.

Table 8 represents the predicted employability probabilities across the main academic domains in the Democratic Republic of Congo, together with their 95% confidence intervals obtained via bootstrapping (1000 replications).

The predicted probabilities indicate that graduates from Economics and Management Sciences have the highest employability prospects (82.4%), followed by Law and Political Sciences (79.0%) and Arts and Languages (78.8%).

Table 8. Predicted probabilities by fields.

At the other end, Technology (68.9%), Health (73.7%), and Psychological/Educational Sciences (73.9%) record the lowest probabilities.

The non-overlapping confidence intervals (e.g., between 68.9% and 82.4%) confirm that these inter-domain differences are statistically meaningful and robust.

In other words, there is a 13-point employability gap between the best-aligned (Economics & Management) and least-aligned (Technology) domains.

The results show that employability is strongly correlated to the similarity between academic profile and market requirements, rather than to the field of study. The economic, legal and management fields perform best, while technical, digital and cross-cutting skills are the main shortfall to be filled in order to improve the vocational integration of graduates.

The results confirm a marked disparity between the skills developed in Congolese universities and those demanded by the labour market. The global adequacy rate of 49.14% highlights a structural misalignment consistent with earlier observations in African and OECD contexts [1]. In other words, only one out of two competencies taught in the Congolese higher education system is directly relevant to current professional needs.

This gap underscores the limited responsiveness of university curricula to the dynamics of labour demand, particularly in sectors such as digital technologies, data analytics, and language proficiency—areas identified by the ILO [3] and UNESCO [7] as key determinants of employability in the digital age.

At the same time, the high predictive power of semantic similarity (≈68% of variance explained) empirically confirms that employability is not primarily determined by the field of study but rather by the degree of skill alignment between academic and occupational domains, a result that echoes [11] and [16], who emphasized the central role of transferable and context-relevant competencies.

The inter-domain analysis shows strong contrasts: Economics & Management (mean employability = 82.4%) and Legal-Political Sciences (79%) achieve the best outcomes, while Technology (68.9%) and Psychological-Educational Sciences (73.9%) lag behind.

These differences can be attributed to several structural and pedagogical factors:

Sectoral structure of the Congolese economy.

Tertiary services (administration, banking, commerce, NGOs) dominate the national economy, favouring graduates from managerial and administrative fields (World Bank, 2023).

Professionalization and exposure.

Curricula in business and law often include internships, case studies, and partnerships with employers—enhancing experiential learning and work-integrated education [12][22].

Transferability of skills.Management, accounting, and legal literacy are cross-sectoral and therefore more easily redeployable across industries, unlike highly specialized scientific or educational skills [17].

Conversely, Science and Technology programs face structural barriers, insufficient laboratory resources, weak industrial ecosystems, and limited collaboration with private firms [5][8]. Similarly, graduates in education and health often face saturated public sectors and outdated pedagogies, echoing UNESCO’s warning about the mismatch between traditional didactics and modern labour expectations [7].

The logistic regression results (Pseudo R² = 0.6899; AUC = 0.977; Accuracy = 0.918) demonstrate an exceptionally strong predictive model, confirming the robustness of the semantic–probabilistic framework. The coefficient for semantic similarity (+53.27, p < 0.001) indicates that each 1% increase in skill alignment boosts employability probability by approximately 70%—an effect size rarely observed in social data.

This finding supports the theoretical stance of employability models proposed by [16] and [23], according to which professional insertion results from a combination of personal, institutional, and contextual factors, but where skill adequacy remains the most decisive micro-determinant.

In this perspective, the DRC case confirms the hypothesis advanced by [24]: structural labour-market constraints (weak private-sector absorption, rigid public employment) can be partially mitigated when training is explicitly competency-driven and aligned with market language and technologies.

Additional empirical and methodological perspectives from recent studies further reinforce the interpretation of our findings. Several authors highlight that labour-market outcomes in African and emerging contexts depend not only on domain-specific knowledge but also on structural factors such as public employment dynamics and macroeconomic constraints [25][26]. Similarly, research in health sciences and professional training shows that competency development and assessment practices strongly shape the transition from academic training to workplace performance [27]-[30]. The pedagogical literature also emphasizes the importance of alternative instructional models and professional development of university teachers in enhancing student readiness and adaptability [31]-[33]. From a methodological standpoint, recent NLP-based studies confirm the reliability of vector-space linguistic models for analysing educational and professional corpora [34]-[36] while foundational works in statistical NLP remain central to understanding the structure and behaviour of language models used in this study [37]. Moreover, the literature documents the value of analysing the interaction between education systems and labour-market absorption capacity, particularly in African countries facing demographic pressure and structural unemployment [38]-[40]. These complementary insights align with our results and justify the adoption of a semantic–probabilistic approach for analysing employability in the DRC.

The DRC results mirror trends observed across Africa:

In Kenya and Ghana, more than half of graduates report a disconnect between their studies and labour-market needs [1].In South Africa, employability is higher in applied scientific fields where universities collaborate with tech industries [25].In Morocco and Tunisia, the LMD reform’s success relied on integrating English, entrepreneurship, and digital literacy into university curricula [7].

Internationally, the World Development Report [2] stresses that adaptability, digital skills, and communication competencies exert a stronger influence on employability than disciplinary specialization; exactly what this study’s probabilistic model confirms for the Congolese context.

The empirical findings suggest several strategic implications:

a)Curriculum reform and modernization.

Universities should embed digital, linguistic, and entrepreneurial modules into all curricula, reinforcing transversal competencies.

b)Creation of Innovation and EmployabilityCentres(IECs).

Following African Development Bank recommendations, IECs can coordinate career guidance, internships, and partnerships with employers.

c)Strengthening university–industry partnerships.

Joint development of professional pathways (e.g., Economics + Data Science, Law + Digitalization) can directly improve employability outcomes [15][27].

d)Evidence-based governance.

The Ministry of Higher Education should institutionalize an Observatory of Skills and Employability, ensuring continuous data collection and semantic monitoring [8].

This research contributes methodologically by demonstrating that semantic similarity measures, grounded in vector-space models [18]-[20], can be effectively coupled with logistic probability modelling to quantify employability.

It bridges the gap between textual analysis (lexicometry) and econometric inference, thereby operationalizing the call by [11] and [16] for hybrid, data-driven approaches to educational adequacy.

In sum, this study provides empirical evidence that employability in the DRC depends primarily on the semantic alignment between university and market competencies, not on disciplinary background.

By quantifying this relationship through a hybrid semantic-probabilistic model, it offers a reproducible framework for evaluating educational relevance in other developing economies.

Future research should expand this approach to include longitudinal data (career trajectories) and integrate machine-learning classifiers capable of dynamically tracking emerging skills on digital platforms.