A skill term alone is not yet proof of competence
“Problem-solving,” “data literacy,” and “teamwork” appear in job profiles, curricula, and competency models around the world. Nevertheless, there is no guarantee that different organizations, industries, or countries understand them in the same way. An identical label may refer to different activities, proficiency levels, and forms of evidence. Conversely, skills described in different terms may represent almost identical behaviour in practice.
This is precisely where the Open Skills Consortium’s published work-in-progress begins. The OSC describes a globally interoperable Knowledge Graph ontology in which skills are modelled not merely as terms, but as relationships among activities, contexts, qualifications, evidence, and observable behaviour. Vector-based methods are intended to make semantically similar content discoverable across systems and languages. Behaviour Anchors are intended to specify how different proficiency levels of a skill can be recognized in practice. Experts review the proposed mappings, while a GraphRAG-enabled access layer is intended to provide transparent, source-grounded answers.
The approach therefore follows the chain formulated in the SupraTix research portfolio:
Workflow event → Skill hypothesis → Human review → Learning intervention → Outcome
Behaviour Anchors translate abstract skills into observable action
Competency models deliver value only when abstract concepts are translated into concrete performance expectations. Campion and colleagues describe 20 best practices that connect the analysis, structuring, presentation, and use of competency information. Competency modelling complements traditional job and task analysis, but does not replace it [Campion et al., 2011].
The idea of behavioural anchors dates back to the development of Behaviorally Anchored Rating Scales. Smith and Kendall proposed anchoring ambiguous rating scales with concrete behavioural examples derived from practice [Smith & Kendall, 1963]. The OSC does not simply transfer this basic principle as a personnel assessment scale, but introduces it as a semantic layer within a skills ontology: a Behaviour Anchor describes observable action, the context of application, the required proficiency level, and the admissible evidence.
A purely illustrative example for the skill “diagnosing technical faults,” which has not yet been approved by subject-matter experts, might look like this:
At a basic level, a person identifies documented error codes and correctly performs a prescribed diagnostic step.
At an advanced level, the person prioritizes multiple fault hypotheses based on measurement data, tests them systematically, and documents the decision-making process.
At an expert level, the person develops a cross-system diagnostic plan, rules out alternative causes, and translates the findings into improved diagnostic procedures.
Such anchors do not create objectivity by themselves. They do, however, make clear what is to be assessed. The quality of the expert review is crucial. In a quasi-experiment, trained raters improved the discriminant validity, agreement, and accuracy of competency assessments compared with an untrained control group. The best results were achieved by trained individuals with additional experience in competency modelling [Lievens & Sanchez, 2007]. Expert validation is therefore not a retrospective quality check within the OSC, but an integral part of the data model and approval process.
Why the ontology needs a vector layer
Traditional ontologies are effective at representing explicit meanings and relationships. For example, they can specify that a skill is required for an activity, developed through a learning outcome, or demonstrated through a particular work sample. The challenge becomes greater when millions of texts, synonyms, regional expressions, and emerging skill terms need to be matched.
Vector-based representations complement the graph with a statistical similarity layer. Early Knowledge Graph embedding methods such as TransE demonstrated that entities and relations in multi-relational data can be modelled in low-dimensional vector spaces and used for link prediction [Bordes et al., 2013]. Later work on Entity Alignment combined names, relations, and attributes to map corresponding entities between different Knowledge Graphs more robustly [Zhang et al., 2019].
Cross-language alignment is also essential for a global skills infrastructure. Reimers and Gurevych demonstrated how sentence embeddings from more than 50 languages can be mapped into a shared vector space, positioning translations and semantically similar statements closer together [Reimers & Gurevych, 2020].
For the OSC, this leads to one central rule:
Vector similarity produces a mapping proposal, not semantic truth.
Two skill descriptions may be close together in vector space while still differing in their levels of responsibility, safety requirements, or occupational contexts. Their authoritative meaning therefore remains anchored in the versioned Knowledge Graph through explicit relations, sources, scopes of validity, and approval statuses.
Automated skill recognition remains a proposal mechanism
Research on skill extraction demonstrates both the potential and the limitations of automated methods. The peer-reviewed SkillSpan dataset contains 14,500 sentences and more than 12,500 skill spans from English-language job advertisements annotated by domain experts. Domain-adapted language models performed significantly better than non-adapted baselines [Zhang et al., 2022]. This emphasizes that skill recognition is not merely a general language problem; it requires domain-specific data and clear annotation rules.
Nor is the automated structuring of entire competency models a substitute for human modelling work. In an exploratory study, competency frameworks developed by humans were generally rated as more coherent than frameworks generated using Natural Language Processing. Only for the most complex model examined was there no longer a discernible advantage for human structuring [Garman et al., 2021]. For the OSC, this supports a Human-in-the-Loop approach: machines scale search, clustering, and proposal generation, while qualified experts decide on meaning, proficiency level, context, and approval.
The proposed OSC architecture combines five layers
The published work-in-progress can be described as a hybrid architecture:
Explicit ontology and Knowledge Graph
Skills, activities, roles, qualifications, learning opportunities, credentials, Behaviour Anchors, and sources receive stable identities and typed relationships. Every change remains versioned and traceable.
Multilingual vector and alignment layer
Embeddings identify semantically similar descriptions, translation variants, potential duplicates, and new mapping candidates. They facilitate crosswalking between existing taxonomies, but do not replace an expert decision on equivalence.
Behaviour Anchor and evidence layer
Skills are linked to observable actions, application contexts, proficiency levels, and admissible types of evidence. The ontology therefore describes not only what a skill is called, but also how its application can be recognized.
Expert validation and governance
Domain Owners review proposals using defined criteria. Approvals, rejections, disagreements, evidence, language versions, and scopes of validity are stored as auditable provenance. High-impact decisions must not be based solely on unconstrained LLM classifications.
GraphRAG-enabled access and reasoning logic
User questions are answered not only against isolated text passages, but against relevant subgraphs, relationships, and sources. The generated answer should be able to reference the nodes, edges, and evidence used.
What GraphRAG can do—and what it cannot
Retrieval-Augmented Generation connects language models with externally retrievable knowledge. Foundational RAG research explicitly addresses challenges such as updatable knowledge and provenance [Lewis et al., 2020].
More recent peer-reviewed research extends this principle to graph structures. G-Retriever retrieves and highlights subgraphs relevant to a question and demonstrated advantages over comparison methods across several textual graph tasks, as well as a reduction in hallucinations [He et al., 2024]. KG²RAG uses relationships between knowledge elements to expand and organize semantically retrieved seed passages through a graph-based process. On HotpotQA variants, the method improved retrieval and answer quality [Zhu et al., 2025].
For the OSC, GraphRAG is therefore primarily an evidence-grounded access layer. One possible question might be: “Which observable actions demonstrate the skill of process diagnostics at an advanced level, in which qualification frameworks does it appear, and which learning opportunities address an identified gap?” To answer this, the system could retrieve skill nodes, Behaviour Anchors, qualifications, evidence types, and learning opportunities as a connected subgraph.
GraphRAG does not, however, guarantee either truth or fairness. Incorrect edges, inadequate sources, or culturally biased anchors may simply be presented in a particularly convincing manner. Retrieval coverage, source precision, answer faithfulness, and the accurate representation of uncertainty must therefore be evaluated separately.
“Global” means validatable, not identical everywhere
A multilingual ontology is not automatically valid worldwide. Translation equivalence proves neither cultural equivalence nor identical occupational requirements. Research on measurement invariance shows that comparisons between groups require a construct to be measured in a sufficiently equivalent manner across the groups being compared [Vandenberg & Lance, 2000].
The term “global Behaviour Anchors” should therefore not be understood as a single universal description. A more appropriate model is a shared, open core with controlled local variants:
a globally referenceable skill concept,
language- and region-specific terms,
context-dependent Behaviour Anchors,
documented correspondences and differences,
separate approvals for each professional domain, language, and legal or educational context,
empirical validation before results are compared across groups.
This allows interoperability to emerge without making local differences invisible.
Research and validation programme for the next step
The specific OSC architecture has not yet been peer-reviewed or validated worldwide as an integrated system. A robust pilot should therefore combine at least five validation tracks:
Semantic alignment
Experts assess whether the mappings proposed by the vector model are genuinely equivalent, related, or merely superficially similar. Metrics include Precision@k, Recall@k, and error types by language and domain.
2. Quality of the Behaviour Anchors
Expert groups develop and review observable anchors. They assess clarity, contextual appropriateness, proficiency level, evidentiary support, and inter-rater agreement.
3. Language and cultural validation
Translated anchors are not only back-translated linguistically, but also tested in real occupational contexts. Where comparisons are intended, measurement invariance is examined.
4. Graph and provenance quality
Every relevant statement must be traceable to approved nodes, relations, and sources. Versions, conflicts, and uncertainties remain visible.
5. GraphRAG evaluation
The system is tested using domain-specific questions. Relevant criteria include retrieval coverage, source and citation precision, answer faithfulness, handling of conflicting evidence, and the ability to refrain from producing a seemingly confident answer when the available evidence is insufficient.
The potential value for the skills ecosystem
Such an infrastructure could connect different stakeholders through a shared but verifiable data foundation:
Companies could link role profiles to concrete activities, evidence, and learning pathways.
Education providers could relate learning outcomes more precisely to observable occupational behaviour.
Chambers and recognition authorities could transparently present similarities and differences between qualifications without automatically claiming equivalence.
Labour-market institutions could identify emerging skill patterns earlier and review their classification transparently.
Employees and learners could receive more portable credentials that communicate not only a label, but also context and evidence.
HR technology, LMS, and talent marketplace providers could access versioned, machine-readable, and source-grounded skill relationships.
Current status and invitation to expert review
The approach described here is a published OSC work-in-progress and part of the research agenda on Evidence-linked Skills Intelligence and the SupraGraph Skills Twin. The scientific literature supports key components: behaviour-based competency descriptions, expert validation, multilingual embeddings, Knowledge Graph alignment, and graph-supported retrieval. It does not yet demonstrate that the specific OSC implementation is already globally valid or production-ready.
The next appropriate step is a structured review and participation process involving experts in vocational education and training, competency assessment, occupational psychology, recognition, ontology development, Knowledge Graphs, multilingual AI, and labour-market research. Together, they should examine which Behaviour Anchors can be referenced globally, where local variants are required, and what evidence is sufficient for robust approval.
The goal is not to create the next closed skills list. The goal is an open, transparent, and adaptive infrastructure that connects skills across systems and languages without confusing similarity with equivalence.
Sources
1. Smith, P. C., & Kendall, L. M. (1963). Retranslation of expectations: An approach to the construction of unambiguous anchors for rating scales. Journal of Applied Psychology, 47(2), 149–155. [DOI](https://doi.org/10.1037/h0047060)
2. Lievens, F., & Sanchez, J. I. (2007). Can training improve the quality of inferences made by raters in competency modeling? A quasi-experiment. Journal of Applied Psychology, 92(3), 812–819. [DOI](https://doi.org/10.1037/0021-9010.92.3.812)
3. Campion, M. A., et al. (2011). Doing competencies well: Best practices in competency modeling. Personnel Psychology, 64(1), 225–262. [DOI](https://doi.org/10.1111/j.1744-6570.2010.01207.x)
4. Vandenberg, R. J., & Lance, C. E. (2000). A review and synthesis of the measurement invariance literature. Organizational Research Methods, 3(1), 4–70. [DOI](https://doi.org/10.1177/109442810031002)
5. Garman, A. N., et al. (2021). Developing competency frameworks using natural language processing: An exploratory study. The Journal of Competency-Based Education, 6(3), e01256. [DOI](https://doi.org/10.1002/cbe2.1256)
6. Bordes, A., et al. (2013). Translating embeddings for modeling multi-relational data. Advances in Neural Information Processing Systems 26. [Proceedings](https://proceedings.neurips.cc/paper_files/paper/2013/hash/1cecc7a77928ca8133fa24680a88d2f9-Abstract.html)
7. Zhang, Q., et al. (2019). Multi-view knowledge graph embedding for entity alignment. IJCAI 2019, 5429–5435. [DOI](https://doi.org/10.24963/ijcai.2019/754)
8. Reimers, N., & Gurevych, I. (2020). Making monolingual sentence embeddings multilingual using knowledge distillation. EMNLP 2020, 4512–4525. [DOI](https://doi.org/10.18653/v1/2020.emnlp-main.365)
9. Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems 33. [Proceedings](https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html)
10. Zhang, M., et al. (2022). SkillSpan: Hard and Soft Skill Extraction from English Job Postings. NAACL 2022, 4962–4984. [DOI](https://doi.org/10.18653/v1/2022.naacl-main.366)
11. He, X., et al. (2024). G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering. Advances in Neural Information Processing Systems 37. [Proceedings](https://proceedings.neurips.cc/paper_files/paper/2024/hash/efaf1c9726648c8ba363a5c927440529-Abstract-Conference.html)
12. Zhu, X., et al. (2025). Knowledge Graph-Guided Retrieval Augmented Generation. NAACL 2025, 8912–8924. [DOI](https://doi.org/10.18653/v1/2025.naacl-long.449)
