Ontology-based Skill Description Learning

• The paper presents a framework that integrates advanced NLP, semantic embeddings, and logic-based reasoning to achieve high-precision skill extraction.
• Ontology-based skill description learning employs state-of-the-art embedding techniques and contrastive losses to map unstructured texts to formal skill representations.
• The approach enables adaptive applications in HR analytics and personalized learning by transforming abstract skill data into machine-interpretable ontologies.

Ontology-based skill description learning refers to automated or semi-automated frameworks for extracting, formalizing, and aligning descriptions of skills with formally defined ontologies. This approach is central to structuring unstructured information about skills—derived from documents, logs, resumes, or educational resources—so that downstream applications such as workforce analytics, personalized learning, and automated production planning can operate with machine-interpretable, verifiable skill representations. Core methodologies integrate advanced NLP (e.g., transformers, LLMs), semantic embeddings, logic-based induction, and matching against domain-standard ontologies such as ESCO, leveraging both statistical and symbolic reasoning for robust, scalable, and interpretable skill modeling.

1. Formal Frameworks and Ontology Structures

Skill modeling ontologies encode skills as atomic or composite entities, defined in description logic (DL), and related through formal axioms and object properties. The EduCOR ontology provides a prototypical structure, decomposing the concept Skill into linked subcomponents such as KnowledgeTopic, links to Job, and connections to EducationalResource and UserProfile. Formal axioms assert patterns including:

• $$\text{Skill} \sqsubseteq \exists\,\text{requiresKnowledgeTopic.KnowledgeTopic}$$
• $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$
• $$\text{EducationalResource} \sqsubseteq \exists\,\text{coversTopic.KnowledgeTopic}$$

These constructs enable automated reasoning for skill acquisition, recommendation, and learning-path generation, bridging abstractions of knowledge, labor-market requirements, and user learning goals (Ilkou et al., 2021).

Industrial settings further leverage OWL-DL ontologies: each skill is specified as a class expression or first-order logic predicate, relating operations, modules, materials, and positional parameters. Logic-based representations—coupled with inductive logic programming (ILP)—support the extraction of candidate skill descriptions from operational logs, with CELOE-based refinement over the space of class expressions to discover formalized, machine-actionable skill definitions (Himmelhuber et al., 2021).

2. Extraction Pipelines and Semantic Embedding

State-of-the-art pipelines ingest diverse, unstructured documents, applying:

• Robust text preprocessing: domain-adapted tokenization, lemmatization, normalization, segmentation (e.g., into 120-token chunks for local context preservation).
• Semantic embedding: Models such as SentenceTransformer (“all-MiniLM-L6-v2”, 384-D) map text chunks and ontology skill definitions to a shared vector space. Embeddings are L2-normalized: $\hat v = \frac{v}{\lVert v\rVert_2}$.
• Skill extraction via vector search: Skills are indexed (e.g., in FAISS IndexFlatIP for cosine search), and document segments are matched to ontology skills using a similarity threshold ($\tau=0.35$ typical), with per-skill frequency aggregation:

$$f_i = \sum_{j=1}^N \mathbf{1}\bigl\{\mathrm{sim}(\hat v_j,\hat u_i) > \tau\bigr\}$$

This enables both explicit (direct mention) and implicit (contextual inference, e.g., verb–object patterns) skill identification, with near-human performance (F1 $>0.95$ explicit, F1 $>0.93$ implicit) on standardized test sets (Koundouri et al., 13 Mar 2025).

In XMLC formulations for job ad skill extraction, contrastive bi-encoders (BERT backbone with BiLSTM and attention pooling) further align unstructured sentences to ESCO definitions, enabling zero-shot, cross-lingual skill retrieval. Margin-based contrastive losses and synthetic training data—generated by LLMs and structured to preserve ESCO hierarchy—drive retrieval F1@5 up to $0.72$, outperforming TF-IDF and standard transformer baselines (Sun, 14 Jan 2026).

3. Ontology Alignment, Decomposition, and Verification

Ontology-based alignment ensures that extracted skill mentions are mapped to canonical concepts. Typical pipelines associate each candidate with an ontology node by maximizing cosine similarity over precomputed embeddings for ESCO concepts, enforcing high-confidence ($s_{ij} > \tau$) mapping and supporting further aggregation at the occupation or curriculum level (Koundouri et al., 13 Mar 2025).

Automated skill decomposition—splitting a broad skill node into constituent sub-skills—has been systematized with LLM-driven, ontology-anchored frameworks. Controlled prompt engineering (zero-shot/few-shot with leakage-safe exemplars) and surface-level normalization feed candidate sub-skills into embedding-based verification against the ontology’s child nodes. Metrics include semantic F1 (cosine-aligned paraphrastic matching) and hierarchy-aware F1 (structural fidelity in the ontology graph), allowing quantitative benchmarking (e.g., F1 $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$0–$$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$1 for different LLM strategies) (Luyen et al., 13 Oct 2025).

Logic-based and neural approaches (e.g., RNN encoder–decoders for DL axiom induction) further expand expressive ontology learning, capturing complex constructs such as cardinality constraints and existential-role restrictions by transducing raw text to DL-formulas. These pipelines can be adapted to skill domain vocabularies, leveraging silver/gold training pairs from existing taxonomies (Petrucci et al., 2016).

4. Integration with Downstream Systems and Applications

Skill ontologies, once populated, underpin:

• Automated learning-path recommendation: Reasoners infer requisite topics, fetch resources per topic, and assemble sequences tailored to user learning goals (SPARQL queries operationalize this logic) (Ilkou et al., 2021).
• Occupation mapping: Skills aggregate to occupational profiles, with rank scores combining overlap in required skills and semantic similarity of occupation descriptions:

$$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$2

where $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$3 measures skill set overlap, $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$4 textual similarity, and $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$5 is a weight parameter (Koundouri et al., 13 Mar 2025).

• Course alignment: Embedding-based retrieval matches skills to courses indexed similarly to skills, supporting personalized upskilling and curriculum advice.
• Visualization and analytics: Real-time dashboards construct tripartite graphs of skills, occupations, and courses. Node sizes and edge colors reflect instance frequencies and similarity strengths, exposing labor-market and training pathway structures.

5. Evaluation Methodologies and Metrics

Rigorous evaluation protocols combine explicit and implicit annotation, synthetic and real-world benchmarks, and multiple axes of performance:

Subdomain	Metric/Result	Key Source
Skill Extraction	F1: $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$6 (explicit), $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$7 (implicit)	(Koundouri et al., 13 Mar 2025)
Skill Decomp.	Sem. F1: $$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$8–$$\text{Job} \sqsubseteq \exists\,\text{requiresSkill.Skill}$$9	(Luyen et al., 13 Oct 2025)
Job Ads Retrieval	F1@5: $$\text{EducationalResource} \sqsubseteq \exists\,\text{coversTopic.KnowledgeTopic}$$0 zero-shot	(Sun, 14 Jan 2026)
ILP Skill Induction	Recall: $$\text{EducationalResource} \sqsubseteq \exists\,\text{coversTopic.KnowledgeTopic}$$1, Prec: $$\text{EducationalResource} \sqsubseteq \exists\,\text{coversTopic.KnowledgeTopic}$$2–$$\text{EducationalResource} \sqsubseteq \exists\,\text{coversTopic.KnowledgeTopic}$$3 (Top20)	(Himmelhuber et al., 2021)

Metrics include token-, chunk-, and concept-level precision, recall, F1; semantic and hierarchy F1 (embedding+structure aware), and coverage/consistency tests for reasoning frameworks. Evaluation scripts compare predicted skills/axioms to ground-truth, and ablation studies assess sensitivity to hyperparameters, hierarchy constraints, or prompt type.

A plausible implication is that near-supervised accuracy in real-world, cross-lingual labor market settings can be achieved without manual annotation by synergistically combining LLM-driven synthetic data, taxonomy-grounded filtering, and contrastive neural architectures (Sun, 14 Jan 2026).

6. Challenges, Limitations, and Future Directions

Open challenges include managing overlapping or fine-grained taxonomy labels (“taxonomy misalignment”), improving precision in class-expression induction (expert filtering remains essential), and scaling to billion-scale ontologies (necessitating FAISS-IVFPQ or contrastive regularization). The reliance on single-model LLMs for data generation introduces potential bias and phrasal rigidity. Enhanced scalability is possible via distributed embedding precomputation and adoption of cross-lingual ontological frameworks. Integration of user/expert feedback (for threshold adaptation and candidate validation) and curriculum refinement (e.g., by mining labor-market trends) represent active areas of research.

The explicit formalization of skills as ontology nodes with machine-actionable properties positions ontology-based skill description learning as a central enabler for adaptive, explainable, and interoperable computational systems in HR analytics, curriculum design, production automation, and AI-driven lifelong learning (Koundouri et al., 13 Mar 2025, Himmelhuber et al., 2021, Ilkou et al., 2021, Sun, 14 Jan 2026, Luyen et al., 13 Oct 2025, Petrucci et al., 2016).