VacancySBERT: Domain-Adapted Semantic Matching

• VacancySBERT is a family of domain-adapted semantic similarity models that align heterogeneous HR texts into structured embedding spaces for effective retrieval and classification.
• It employs transformer-based dual-encoder architectures with skill delimiter tokens and projection heads to achieve vocabulary-agnostic semantic comparisons.
• The model delivers substantial gains in recall, interpretability, and efficiency over traditional lexical baselines, excelling in multilingual job–CV matching and taxonomy linking.

VacancySBERT is a family of domain-adapted, neural semantic similarity models deploying the Siamese Sentence-BERT (SBERT) or related dual-encoder architectures to align heterogeneous HR texts—such as job titles, vacancy skill requirements, and candidate CVs—into structured embedding spaces for high-throughput, vocabulary-agnostic retrieval and classification in labor market analytics. Across its instantiations in multilingual job–CV matching, title normalization, curriculum–job alignment, and ontology linking, VacancySBERT has demonstrated substantial gains in recall, interpretability, and operational efficiency over lexical and generic neural baselines, while providing mechanisms for skill-informed matching, taxonomy compliance, and explainable analytics.

1. Model Architectures and Embedding Spaces

VacancySBERT is uniformly built upon variants of Siamese (bi-encoder) networks leveraging transformers like BERT, MiniLM, or domain-specific models. Architectures consist of two branches—often with shared weights—that separately encode inputs such as job titles and aggregated skill sets, or CV and vacancy full texts. Core architectural features include:

• Transformer Encoder Backbone: Most commonly BERT-base or mBERT (12 layers, 768-dim hidden), or all-MiniLM-L6-v2 for lightweight deployments (Bocharova et al., 2023, Turaev et al., 1 Jun 2026).
• Input Segmentation and Pooling: Each branch independently processes input tokens (title, skill set, candidate profile, vacancy description), with mean-pooling or [CLS] pooling to yield a fixed-dimensional vector (e.g., 768 or 384).
• Skill Delimiter Tokens: In skill-aware branches, a special [SKILL] token is prepended to each skill span before aggregation; German models introduce [JOB_TITLE_SEP], [QUALIFICATION_SEP], [SKILL_SEP] for structured queries (Hihn et al., 5 Sep 2025).
• Projection Heads: Lightweight MLP projections are often added, with dimensionality reduction (e.g., to 64 or 256 dimensions) and L2 normalization for efficient similarity computation and retrieval (Hihn et al., 5 Sep 2025, Lavi, 2021).

Each input (e.g., a job title, skill set, or CV) is thus mapped into a shared embedding space where semantic similarity is computed as cosine similarity.

2. Training Objectives, Loss Functions, and Data

VacancySBERT’s training objectives universally center on supervised or self-supervised contrastive learning tailored for retrieval:

• Multiple Negative Ranking (MNR) Loss: Given $N$ positive pairs in a batch, each anchor (e.g., title) must be most similar to its matching pair (e.g., skills), contrasting against $N-1$ in-batch negatives (Bocharova et al., 2023, Hihn et al., 5 Sep 2025, Lavi, 2021):

$$\mathcal{L} = -\frac{1}{N}\sum_{i=1}^{N} \log\frac{\exp(\mathrm{sim}(u_i, v_i)/\tau)}{\sum_{j=1}^{N} \exp(\mathrm{sim}(u_i, v_j)/\tau)}$$

• Classification or Regression Heads: For binary match/non-match prediction (e.g., resume–vacancy), extra fully connected classification or cosine-similarity regression heads are used (Lavi et al., 2021).
• Multi-scale Supervision: In ontology-aligned settings, Matryoshka meta-loss encourages semantic alignment at multiple embedding dimensionalities (Hihn et al., 5 Sep 2025).

Training data is large-scale and domain specific:

• 50 million job title–description pairs and 33,000 manually normalized titles for English title normalization (Bocharova et al., 2023).
• 270,000 labeled resume–vacancy interaction pairs, with positive and negative exemplars from placements, rejections, and random negatives (Lavi et al., 2021).
• Over 525,000 German occupational terms with KldB and ISCED codes, forming 1.5 million anchor-positive-negative triplets (Hihn et al., 5 Sep 2025).
• For curriculum–job alignment, formatted records of extracted competencies and job requirement clauses (Turaev et al., 1 Jun 2026).
• Preprocessing incorporates language filtering, boilerplate removal, skill phrase extraction, and document truncation to fit transformer input constraints.

3. Taxonomy Anchoring and Cross-Modal Mapping

VacancySBERT advances beyond generic semantic matching by normalizing inputs to shared taxonomies:

• ESCO (European Skills, Competences, Qualifications and Occupations): Used to map free-form curriculum and job texts into one of eleven computing-relevant domains, ensuring that coverage and gaps are defined according to a stable ontology (Turaev et al., 1 Jun 2026).
• KldB 2010 and ISCED 2011: German models embed job titles into a unified space aligned with both occupational and educational hierarchies, enabling semantic search and alignment across disparate vocabularies (Hihn et al., 5 Sep 2025).
• JSON Schema Enforcement: For curriculum–job alignment, extraction is governed by a seven-slot schema, enabling slotwise quality control and traceability (Turaev et al., 1 Jun 2026).

This approach enables vacancy–CV, title–skill, or curriculum–requirement matching to be interpreted and audited in the context of structured, expert-defined domains.

4. Evaluation, Metrics, and Benchmarking

VacancySBERT models are extensively benchmarked against both lexical and neural baselines:

• Recall@N and Mean Reciprocal Rank (MRR): For title normalization and candidate matching, top-k accuracy is reported. VacancySBERT achieves Recall@1 up to 30.1%, Recall@5 up to 42.5%, and Recall@10 up to 55.6%, representing 10–21.5% improvement over JobBERT and generic SBERT baselines (Bocharova et al., 2023, Lavi, 2021).
• Macro-F1 and Hierarchy-Level Accuracy: German models report macro-F1 up to 0.961 at requirement level, and mAP@3 >0.97 across taxonomic hierarchies (Hihn et al., 5 Sep 2025).
• Skill-Domain Coverage and Gap Ratios: In curriculum–job alignment, domain-level supply, demand coverage, and coverage gaps are quantified mathematically, e.g., a 25.0% gap in general and transversal skills, 1.8% in AI/data science (Turaev et al., 1 Jun 2026).
• Cohen’s Kappa for Extraction Reliability: Inter-annotator agreement for extraction slots reaches κ = 0.79 (substantial) on skill (Turaev et al., 1 Jun 2026).
• Ablation and Robustness Analyses: Perturbation and cross-lingual tests quantify the invariance of embeddings to superficial variation and vocabulary gap (Lavi et al., 2021, Hihn et al., 5 Sep 2025).
• Latency and Scalability: End-to-end ANN (e.g., FAISS/HNSW) search yields sub-100 ms latency for >1M candidates or titles (Lavi et al., 2021, Hihn et al., 5 Sep 2025).

The open-source 33k-title normalization benchmark and codebase further supports replicability and comparative research (Bocharova et al., 2023).

5. Cross-Lingual and Multilingual Capabilities

VacancySBERT consistently leverages multilingual pretraining and supervision for robust cross-lingual semantic mapping:

• mBERT/XLM-R Backbones: Cover 100+ languages, supporting CV–vacancy matching in multilingual markets (Lavi et al., 2021, Lavi, 2021).
• Cross-lingual Positive Pairs: Models are explicitly exposed to title–requirement or CV–vacancy pairs that span different languages, forcing the alignment of semantic content despite surface linguistic variation (Lavi et al., 2021, Lavi, 2021).
• Zero-Shot Generalization: Models transfer to new languages or unseen job domains without retraining, with no external thesaurus (Lavi et al., 2021).

Empirical results demonstrate that cross-lingual positive pairs are embedded closely (cosine similarity >0.78), while unsupervised models fail to cluster semantically similar but lexically distinct mentions (Lavi et al., 2021).

6. Applications, Deployment, and Interpretability

VacancySBERT is deployed in multiple operational workflows:

• Title Normalization and Skill Taxonomy Linking: Models automatically map noisy, ambiguous job titles and skills from resumes or ads to normalized taxonomies (e.g., ESCO, KldB) for retrieval, reporting, and compliance (Bocharova et al., 2023, Hihn et al., 5 Sep 2025).
• Resume–Vacancy and Curriculum–Job Matching: Embeddings support scalable ANN retrieval (e.g., FAISS, HNSW) and, when combined with structured business features, power large-scale recommender systems (Lavi et al., 2021, Turaev et al., 1 Jun 2026).
• Gap Analysis and Curriculum Governance: The SBERT-alignment pipeline provides fully auditable supply–demand gap matrices, including domain-level coverage, Bloom-level depth differentials, and temporal adoption lags (Turaev et al., 1 Jun 2026).

The pipeline architecture supports routine index refresh, human-in-the-loop correction, and provides evidence fields (raw cosine, provenance) for interpretability. Single multilingual models can be maintained and rolled out across markets with minimal overhead.

7. Comparative Significance and Limitations

VacancySBERT marks a substantial advancement over prior approaches:

• Semantic Bridging of Heterogeneous and Multilingual Inputs: Directly fine-tuning dual-tower networks on real-world co-occurrence data (placements, skill sets, ontology tags) allows bridging of vocabulary gaps unaddressed by surface or generic neural methods (Bocharova et al., 2023, Lavi et al., 2021, Lavi, 2021).
• Evaluation Against Lexical and Skill-Aware Baselines: Robust, statistically significant gains (up to 47.3 pp ROC-AUC) over TF-IDF, off-the-shelf SBERT, and domain-specific models like JobBERT are consistently observed (Bocharova et al., 2023, Lavi et al., 2021).
• Domain-Specific Generalization and Adaptability: The contrastive objective and modular input structure allow adaptation to new taxonomies, languages, or verticals with only minimal data engineering (Hihn et al., 5 Sep 2025, Turaev et al., 1 Jun 2026).

A plausible implication is that fine-tuned, taxonomy-anchored SBERT models will increasingly replace manual mapping and rigid classification for labor market analytics, enabling rapid, explainable alignment across disparate HR data sources. However, real-world deployment necessitates careful threshold calibration, schema-constrained extraction, and ongoing domain-anchoring to ensure interpretability and fairness.