Syllabus-to-O*NET Methodology

• Syllabus-to-O*NET is a rigorous methodology that converts academic syllabi into standardized O*NET skill vectors through a secure NLP pipeline.
• The process filters pedagogical content, embeds sentences with Sentence-BERT, and uses cosine similarity to map learning artifacts into detailed work activities.
• It underpins decentralized credentialing by providing privacy-preserving, bias-resistant skill attestations crucial for accurate job matching.

The Syllabus-to-O*NET methodology is a rigorous, validated approach for deriving standardized occupational skill vectors from academic syllabi and formal learning artifacts. It serves as a bridge between educational records and workforce credentialing, enabling automation of skill extraction from unstructured academic materials and their mapping to the O*NET taxonomy of Detailed Work Activities (DWAs). This methodology underpins AI-enabled, privacy-preserving decentralized Learning and Employment Record (LER) systems, providing robust, auditable, and bias-resistant skill attestations foundational to modern job matching and credentialing ecosystems (Xu et al., 6 Jan 2026).

1. Conceptual Overview and Operational Context

The methodology defines a secure pipeline whereby digitally signed academic records (syllabi, transcripts) are programmatically processed to project educational inputs into the O*NET skill space. The standardized O*NET taxonomy supplies the reference set of “skills” or DWAs, representing a comprehensive, structured inventory used by both researchers and workforce platforms.

The Syllabus-to-O*NET mapping is integrated into decentralized credentialing architectures, particularly those utilizing Trusted Execution Environments (TEEs) for privacy protection. This approach enables the derivation and attestation of skill credentials in a manner that is both verifiable and resistant to bias, with the raw academic records remaining enclave-confined throughout the pipeline (Xu et al., 6 Jan 2026). The methodology thus addresses a core automation bottleneck in bridging the educational and occupational data domains.

2. Pipeline Stages and Formal Mapping Procedures

The pipeline is implemented in four sequential stages within a TEE-adapted NLP stack:

1. Pedagogical Sentence Filtering: The input corpus (syllabi, transcripts) is parsed to identify sentences containing actionable learning content. Non-pedagogical sentences (such as logistics or administration) are filtered out using heuristic and regex-based rules, typically reducing the candidate set to approximately 14% of raw sentences.
2. Sentence Embedding: Pedagogical sentences are embedded using a pretrained Sentence-BERT model (“all-mpnet-base-v2” with default hyperparameters: max length 128 tokens, batch size 32, embedding dimension 768).
3. Similarity Scoring to O*NET: Each embedded sentence vector $\mathbf v_{\mathrm{sent}_j}$ is compared via cosine similarity to the embedding of each O*NET skill descriptor $\mathbf v_{s_i}$. The course-level skill vector is defined via maximal similarity over sentences for each skill:

$$v_{c,i} = \max_{1\leq j\leq n_c}\, \cos(\mathbf v_{\mathrm{sent}_j},\,\mathbf v_{s_i}),$$

yielding $\mathbf v_c = (v_{c,1},\ldots,v_{c,m})$ where $m$ is the number of O*NET skills.

1. Personalized Weighting and Aggregation: For a learner $H$ with course set $C_H$, the aggregate skill vector (attested skill profile) is computed as:

$$\mathbf v_H = \sum_{c\in C_H} w_{\mathrm{grd}(c)}\, w_{\mathrm{lvl}(c)}\, \mathbf v_c.$$

Here, $w_{\mathrm{grd}(c)}$ and $w_{\mathrm{lvl}(c)}$ are grade- and course-level-derived weights, allowing for contextualized and performance-sensitive aggregation across a learner’s formal record.

Thus, the overall mapping function is: $$f: \mathrm{(signed\ records\ } D) \mapsto \mathbf v_H \in \mathbb{R}^m,$$ with $\mathbf v_H$ the final O*NET skill vector.

3. Evaluation, Stability, and Benchmarking

The implemented pipeline demonstrates high stability and fidelity in extracting skill profiles:

• On repeated runs with identical learner data, the top-ranked O*NET skills demonstrate $<5\%$ variance, indicating robust and deterministic mapping under typical stochasticity introduced by NLP infrastructure.
• Empirical evaluation on R1 university computer science syllabi and public job descriptions reports Overlap@10 of $80\%$ (Java Developer) and $70\%$ (C# Data Mining roles) between top extracted skills and reference job requirements.
• Semantic similarity scores (mean of maximal sentence–skill cosines) exhibit high alignment, with all top-10 O*NET skills for target jobs achieving scores $\geq 0.80$.
• Pipeline latency benchmarks on TEEs (AWS Nitro Enclave): for small batches (10–40 records), $9$–$15$ seconds; for large batches (50–100), $<30$–$70$ seconds on commodity or cloud hardware; matching in the verifier enclave operates at $\sim$0.1 seconds per description and low memory/CPU overhead.

These results are supported by prior large-scale validation of the parent method (“Course–Skill Atlas”) on $3.16$M syllabi from $62$ academic fields (Xu et al., 6 Jan 2026).

4. Security, Privacy, and Attestation

All Syllabus-to-O*NET mapping occurs inside the holder-controlled TEE. Institutional credentials are verified by signature, and all ephemeral and persistent states within the enclave are cryptographically hashed (input, model measurement, policy), and bound to the derived skill-VC (verifiable credential). Selective disclosure enables the presentation of derived skills without leakage of raw academic records or auxiliary identifiers.

Formal privacy is underwritten by the TEE’s confidentiality and signature guarantees. Theorems are established that (i) adversarial distinguishing of original records from their derived skill-VCs reduces to breaking TEE confidentiality, and (ii) forgery of credentials reduces to breaking either the issuer’s or enclave’s signature schemes.

5. Job Matching and Bias-Resistant Scoring

The Syllabus-to-O*NET vectorization creates input to a “skill-only” matcher: employers and verifiers receive only the vector $\mathbf v_H$ for job-skill matching, with no access to non-skill resume fields. Job descriptions themselves are processed by the same pipeline to a corresponding skill vector $\mathbf v_J$, allowing for semantic-overlap-based matching: $$\mathrm{Overlap@}k = \frac{|\mathrm{Top}_k(\mathbf v_H) \cap R|}{|R|}, \quad \mathrm{SemSim} = \frac{1}{|R|} \sum_{s\in R} \max_{s'\in S} \cos(\mathbf v_s, \mathbf v_{s'}).$$ Where $R$ is the required skill set for a job, $S$ the full O*NET skill set. A match is declared if the semantic similarity or overlap clears configurable thresholds.

Non-skill invariance is formalized as $h(v,z) = h(v,z')$ for all additional fields $z,z'$; the corresponding bias-opportunity index $\mathrm{BOI}(h)$ for such a pipeline is zero, eliminating bias by construction (Xu et al., 6 Jan 2026).

6. Limitations and Prospects for Extension

Current evaluations are bounded to computer-science syllabi; generalization to broader disciplines remains to be empirically validated. Latency is dominated by NLP pre-processing within enclaves, but remains within practical bounds for batch-mode institutional applications. Side-channel resistance beyond standard TEE protection is not formally modeled.

Potential extensions include hybrid privacy mechanisms (TEE plus zero-knowledge for selective predicates), domain adaptation for non-CS syllabi, dynamic threshold adaptation based on job taxonomy, and continuous taxonomy updates incorporating human-in-the-loop verification (Xu et al., 6 Jan 2026).

7. Significance in Decentralized Credentialing

The Syllabus-to-O*NET methodology is a critical technical enabler for automated, privacy-preserving, and bias-resistant skill extraction in decentralized education and workforce records. By enabling direct, verifiable projection from heterogeneous educational experiences to standardized, employer-interpretable skill ontologies, it anchors the most sensitive component of AI-enabled LERs and online job-matching architectures (Xu et al., 6 Jan 2026). The method provides a stable, scalable bridge between formal curricula and labor market requirements, supporting robust self-sovereign credentialing and reducing institutional and algorithmic sources of bias.