Privacy-Preserving Skill Extraction

• Privacy-Preserving Skill Extraction is a validated NLP pipeline that transforms academic syllabi into quantifiable, machine-readable O*NET skill vectors.
• It integrates TEE-based decentralized systems to securely process signed educational records and ensure bias-resistant, verifiable credentialing.
• It leverages cosine similarity, advanced embedding models, and grade-level weighting to reliably align academic learning outcomes with job skill requirements.

The Syllabus-to-O*NET methodology is a validated NLP pipeline for the automated extraction and quantification of occupational skills from educational syllabi and formal academic records, mapping these to the U.S. O*NET taxonomy of job-related Descriptors of Work Activities (DWAs). This methodology serves as the backbone for several privacy-preserving, decentralized Learning and Employment Record (LER) systems, enabling the derivation of verifiable, machine-readable skill credentials from semi-structured educational artifacts within secure compute environments. The approach provides stability, transparency, and invariance to non-skill attributes, supporting bias-resistant credentialing and automated job-skill matching within decentralized architectures (Xu et al., 6 Jan 2026).

1. Definition and Motivation

The Syllabus-to-O*NET methodology operationalizes the transformation of course content—especially syllabus text—into quantifiable skill vectors anchored on the O*NET standard. It addresses the need for automatable, externalizable, and verifiable skill recognition, bridging formal academic credentials and job market language by mapping textually-described learning outcomes to occupational skill sets. This mapping—$f$—is formally defined as

$$f:\text{(signed records }D)\longmapsto\mathbf v_H\in\mathbb R^m,$$

where $D$ is an authenticated bundle of educational records and $\mathbf{v}_H$ is the resulting person-specific skill vector aligned with the $m$ O*NET descriptors (Xu et al., 6 Jan 2026).

2. Pipeline Stages and Algorithms

The Syllabus-to-O*NET workflow, as embedded within TEE-based LER systems, follows a multi-stage pipeline:

1. Pedagogical-sentence filtering: Syllabus (and optionally transcript) text is filtered to isolate pedagogical content, excluding administrative or non-learning sentences. Regex-based and heuristic filters retain approximately 14% of raw sentences as semantically relevant for skill mapping.

2. Sentence embedding: Each retained sentence is embedded into $\mathbb R^{768}$ using models such as Sentence-BERT "all-mpnet-base-v2" with default hyperparameters (max length 128 tokens, batch size 32). This reduces unstructured text to a dense vector, capturing latent semantic relationships.

3. Skill similarity scoring: Let $S=\{s_1,\dots,s_m\}$ denote the set of O*NET skills. For each sentence vector $\mathbf v_{\mathrm{sent}_j}$ and each skill embedding $\mathbf v_{s_i}$, compute

$$\mathrm{sim}(\mathbf v_{\mathrm{sent}_j},\mathbf v_{s_i}) = \cos(\mathbf v_{\mathrm{sent}_j},\mathbf v_{s_i}).$$

The course-level score for skill $i$ is

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

accumulated into $\mathbf v_c = (v_{c,1},\dots,v_{c,m})$.

4. Grade and level weighting/aggregation: Across the set $C_H$ of an individual's courses, skill contributions are modulated via grade and level weights $w_{\mathrm{grd}(c)}$, $w_{\mathrm{lvl}(c)}$: $$\mathbf v_H = \sum_{c\in C_H} w_{\mathrm{grd}(c)}\,w_{\mathrm{lvl}(c)}\,\mathbf v_c.$$ This produces an attested skill vector for the student.

The mapping is robust: repeated runs on fixed data yield skill vector variance $<5\%$ in top-ranked skills. The methodology inherits extensive validation from the "Course–Skill Atlas" project (>3M syllabi across 62 fields) (Xu et al., 6 Jan 2026).

3. Integration into Decentralized and Privacy-Preserving Credentialing

The methodology is implemented inside a Trusted Execution Environment (TEE) enclave as part of privacy-preserving LER infrastructure. The enclave ingests digitally signed transcripts, syllabi, and informal artifacts, verifies issuer signatures, runs the Syllabus-to-O*NET pipeline to yield $\mathbf{v}_H$, and binds this output cryptographically:

• Inputs and artifact hashes are incorporated into an attestation measurement $H_{\mathrm{inputs}}$.
• The resultant skill vector and its provenance are encapsulated as a self-issued verifiable credential (VC), signed by the enclave's private key.
• Selective disclosure policies ensure that only the derived skill vector is exposed; raw records and keys remain enclave-confined.

Presentations of skill credentials to verifiers (e.g., employers) are accomplished via verifiable presentations, and the attestation group guarantees provenance, integrity, and recency of the computation (Xu et al., 6 Jan 2026).

4. Skill-Only Job Matching and Bias Invariance

A defining property of the Syllabus-to-O*NET methodology, as embedded in LER systems, is its strict “skill-only” matching:

• Verifiers convert job descriptions into O*NET skill vectors ($\mathbf v_J$) using an identical NLP pipeline.
• Matching proceeds solely by comparison of $\mathbf v_H$ (from the applicant) and $\mathbf v_J$ (from the job), invoking semantic-overlap and cosine similarity metrics:

$$\mathrm{Overlap@}k = \frac{|\mathrm{Top}_k(\mathbf v_H)\cap R|}{|R|}, \quad \mathrm{SemSim} = \frac{1}{|R|}\sum_{s\in R}\max_{s'\in S}\cos(\mathbf v_s,\mathbf v_{s'})$$

where $R$ is the required skill set for the job (Xu et al., 6 Jan 2026).

• Threshold-based rules (e.g., $\mathrm{SemSim}\geq \tau$) yield an invariant decision, independent of any non-skill attributes $z$ (gender, age, etc.). Formally, the bias-opportunity index is provably zero:

$$\mathrm{BOI}(h) = \mathbb{E}_v \mathbb{E}_{z,z'}[h(v,z) - h(v,z')]^2 = 0.$$

5. Empirical Performance and Stability

Evaluation details indicate:

• For a benchmark corpus (computer-science syllabi + job postings), skill extraction and mapping display $<5\%$ variance across repeated pipeline runs.
• Example job-matching for a Java Developer role: Overlap@10 = $80\%$, SemSim scores of $>0.80$ for all top skills.
• NLP pipeline throughput: On AWS Nitro Enclave, 10–40 input files processed in $9$–$23$ seconds; for 100 files, processing completes within $50$ seconds. Job matching inside the verifier enclave executes in $\sim0.1$ s per job, consuming minimal resources (Xu et al., 6 Jan 2026).

The primary bottleneck is NLP ingestion; attestation and cryptographic verification incur $<0.05$ s overhead.

6. Security, Privacy, and Limitations

The Syllabus-to-O*NET methodology, by operating wholly inside attested TEEs, provides strong guarantees regarding:

• Confidentiality: Raw educational records and private keys remain strictly within the enclave.
• Unforgeability: Verifiable credentials issued are cryptographically bound to enclave attestation and original issuer signatures; adversarial attempts at generating valid credentials reduce to signature forgeries.
• Selective disclosure: Only attested skill vectors are exposed externally; repeated presentations guarantee unlinkability.
• Stability: Empirical results confirm $<5\%$ output variance on fixed input, supporting reliability for longitudinal or batch processing.

Limitations noted include domain generalizability (currently validated primarily on computer-science syllabi) and the absence of formal side-channel leakage modeling; upstream disparities in background skill distributions are also not explicitly addressed (Xu et al., 6 Jan 2026).

7. Significance and Future Directions

Automating the translation from syllabus to O*NET skill profiles enables standardized, audit-friendly, and privacy-preserving skill credentials suitable for decentralized, employer-verifiable matching. The methodology underpins non-biasable, selectively disclosable, and reproducible mechanisms for linking education to employment. Proposed research avenues include hybrid TEE/zero-knowledge privacy models, human-in-the-loop calibration, and broadening to new academic domains for robust generalizability (Xu et al., 6 Jan 2026).