UniSkill: A Dataset for Matching University Curricula to Professional Competencies

Nurlan Musazade; Joszef Mezei; Mike Zhang

cs.CLMar 3, 2026

UniSkill: A Dataset for Matching University Curricula to Professional Competencies

Authors: Nurlan Musazade, Joszef Mezei, Mike Zhang

Organizations: Åbo Akademi University, Finland · University of Copenhagen, Denmark · Pioneer Centre for AI, Denmark

Abstract

Skill extraction and recommendation systems have been studied from recruiter, applicant, and education perspectives. While AI applications in job advertisements have received broad attention, deficiencies in the instructed skills side remain a challenge. In this work, we address the scarcity of publicly available datasets by releasing both manually annotated and synthetic datasets of skills from the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy and university course pairs and publishing corresponding annotation guidelines. Specifically, we match graduate-level university courses with skills from the Systems Analysts and Management and Organization Analyst ESCO occupation groups at two granularities: course title with a skill, and course sentence with a skill. We train language models on this dataset to serve as a baseline for retrieval and recommendation systems for course-to-skill and skill-to-course matching. We evaluate the models on a portion of the annotated data. Our BERT model achieves 87% F1-score, showing that course and skill matching is a feasible task.

Figures & tables

Figure 1: The preparation of the original pairs.

Non-match (Class 0)
Label	Train	Validation	Test
Sentence-Skill	434	36	171
Title-Skill	331	14	145
Match (Class 1)
Sentence-Skill	261	66	130
Title-Skill	364	88	156

Table 1: Data Distribution. Class distribution across training, validation, and test splits for course sentence-skill, course title-skill, and combined pairs.

ST	Approach	Class	F1	Prec.	Rec.	Avg F1	Acc.
Without	Course sentence-Skill	0	0.87	0.84	0.91	0.84	0.85
	Course sentence-Skill	1	0.81	0.86	0.77	0.84	0.85
	Course title-Skill	0	0.87	0.83	0.90	0.87	0.87
	Course title-Skill	1	0.87	0.90	0.83	0.87	0.87
	Combined	0	—	—	—	—	—
	Combined	1	—	—	—	—	—

Table 2: Separate Model Results. Performance comparison of separate sentence-only, title-only, and combined predictions with and without input labels. All models were trained with synthetic data. We either use Special Tokens (ST) or not. Class 0 = Non-match, Class 1 = Match.

Synthetic	ST	Class	F1	Prec.	Rec.
Yes	No	0	0.89	0.92	0.87
Yes	No	1	0.82	0.78	0.86
		Avg	0.86	0.85	0.87
Yes	Yes	0	0.90	0.93	0.87
Yes	Yes	1	0.83	0.79	0.89
		Avg	0.87	0.86	0.88

Table 3: Combined Approach performance with different training configurations. We either use Special Tokens (ST) or not. Class 0 = Non-match, Class 1 = Match. Best performance (bold averages) achieved with synthetic data and input labels.

ST	Class	F1	Prec.	Rec.	Acc.
No	0	0.63	0.60	0.67	0.68
	1	0.72	0.75	0.69
	Avg	0.68	0.68	0.68
Yes	0	0.68	0.69	0.67	0.74
	1	0.78	0.77	0.79
	Avg	0.73	0.73	0.73

Table 4: Combined Approach performance on 100 hard validation pairs with high semantic similarity. We either use Special Tokens (ST) or not. Class 0 = Non-match, Class 1 = Match. Both models trained with synthetic data.

Model	Class	F1	Prec.	Rec.	Acc.
ESCOXLM-R_ENG	0	0.89	0.87	0.91	0.86
	1	0.79	0.83	0.76
	Avg	0.84	0.85	0.84
ESCOXLM-R	0	0.88	0.90	0.85	0.84
	1	0.79	0.76	0.82
	Avg	0.84	0.83	0.84

Table 5: Domain-specific model performance on the test set (300 pairs). All models trained with synthetic data, input labels, and the Combined Approach. Class 0 = Non-match, Class 1 = Match.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Model architecture	AutoModelForSequenceClassification
Learning rate	$2\times 10^{-5}$
Optimizer	AdamW
Training batch size	16 (shuffle=True)
Evaluation batch size	32
Epochs	7

Appendix

Table 6: Training hyperparameters.

Statistics	Value
Total pairs	301
Positive pairs	108
Negative pairs	193
Avg. sentence length	19.32
Median sentence length	12
Std. sentence length	25.69

Appendix

Table 7: Descriptive statistics of the evaluation dataset.

Statistics	Value
Total pairs	102
Positive pairs	60
Negative pairs	42
Avg. sentence length	18.07
Median sentence length	12.5
Std. sentence length	19.14

Appendix

Table 8: Descriptive statistics of the additional validation dataset.

Original pairs
Total pairs	695
Positive pairs	230
Negative pairs	465
Avg. sentence length	17.52
Median sentence length	11
Std. sentence length	25.62

Appendix

Table 9: Descriptive statistics of the training dataset.

Course Title	Course Sentence	ESCO Skill	Label
Food Metabolomics and Multivariate Analysis	Introduction to different statistical methods (e.g.	statistics	0
Artificial Intelligence for Teachers Open University Studies	Jupyter notebooks can be edited and run in a standard browser, using freely available tools such as Google Collaboratory.	TypeScript	0
Statistical Thinking and Performing Basic Analyses Using JMP	The objective is to understand the basic concepts and methods of descriptive and inferential statistical analysis.	statistics	1
Cybersecurity D	Cyber security concepts.	attack vectors	1
Project Management and Consulting Practice	The project management part deals with advanced issues of project work.	Agile project management	1

Appendix

Table 10: Example course - skill pairs from the real (evaluation) dataset.

Educational Data Mining

Educational Data Mining studies computational methods for analyzing learning data to understand learners, improve instruction, and support educational decisions.

See all 74 papers

JobMatchAI-An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI
Mar 15, 2026Mayank Vyas, Abhijit Chakraborty, Vivek GuptaJob-Candidate Matching Hybrid IR
Arizona State University
Chinese-SkillSpan: A Span-Level Dataset for ESCO-Aligned Competency Extraction from Chinese Job Ads
Apr 24, 2026Guojing Li, Zichuan Fu, Junyi Li +6Named Entity Recognition
City University of Hong Kong · Renmin University of China
An NLP-Driven Framework for Curriculum-Labor Market Alignment: Schema-Constrained LLM Extraction, ESCO-Anchored Semantic Matching, and Multi-Dimensional Gap Quantification
Jun 1, 2026Sherzod Turaev, Mary John, Mamoun Awad +2Document Information Extraction LLM Information Extraction
Department of Computer Science & Software Engineering, College of Information Technology, United Arab Emirates University, Al Ain, Abu Dhabi, United Arab Emirates · Academic Support Department, Abu Dhabi Polytechnic, Abu Dhabi, United Arab Emirates · Department of Information Systems and Security, College of Information Technology, United Arab Emirates University, Al Ain, Abu Dhabi, United Arab Emirates

Mar 15, 2026J/K move · Enter open · S save

JobMatchAI-An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI

Mayank Vyas, Abhijit Chakraborty, Vivek Gupta

Arizona State University

Job-Candidate Matching Hybrid IR Semantic Search Explainable Recommendation Graph-Based Retrieval

Open paper AskPDF arXiv

Recruiters and job seekers rely on search systems to navigate labor markets, making candidate matching engines critical for hiring outcomes. Most systems act as keyword filters, failing to handle skill synonyms and nonlinear careers, resulting in missed candidates and opaque match scores. We introduce JobMatchAI, a production-ready system integrating Transformer embeddings, skill knowledge graphs, and interpretable reranking. Our system optimizes utility across skill fit, experience, location, salary, and company preferences, providing factor-wise explanations through resume-driven search workflows. We release JobSearch-XS benchmark and a hybrid retrieval stack combining BM25, knowledge graph and semantic components to evaluate skill generalization. We assess system performance on JobSearch-XS across retrieval tasks, provide a demo video, a hosted website and installable package.

Abstract

Educational Data Mining

Explore similar work

JobMatchAI-An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI

Chinese-SkillSpan: A Span-Level Dataset for ESCO-Aligned Competency Extraction from Chinese Job Ads

An NLP-Driven Framework for Curriculum-Labor Market Alignment: Schema-Constrained LLM Extraction, ESCO-Anchored Semantic Matching, and Multi-Dimensional Gap Quantification