Job Title-Based Skill Prediction

Updated 6 July 2026

Job title-based skill prediction is a method that infers relevant skills from title strings using multi-label ranking and embedding techniques.
It employs direct prediction, dense retrieval, and graph-based inference to map job titles to a comprehensive skill taxonomy like ESCO.
Applications include labor-market intelligence, personalized career pathfinding, and targeted learning initiatives, bolstered by multilingual transfer and weak supervision.

Searching arXiv for the cited papers to ground the article and ensure up-to-date references. Job title-based skill prediction is the problem of inferring, from a job title string, which professional skills are relevant to the associated occupation and, in some settings, whether those skills are essential or context-dependent. The task has been instantiated as multi-label prediction over a fixed skill vocabulary, semantic retrieval and ranking in a shared embedding space, and link prediction in occupation-skill graphs enriched with job-posting data. Across these formulations, the dominant resources are ESCO for skills, ISCO for occupations, and benchmark-style evaluation with ranking metrics such as MAP, MRR, nDCG, and Precision@k (Gasco et al., 30 Jun 2026). The resulting systems are used for skills-based matching, prioritizing learning and development, labor-market intelligence, and related analyses such as career pathfinding and occupation-specific skill distinctiveness (Groot et al., 2021).

1. Formal problem statements

A standard formulation takes as input a job title $J$ and a fixed repository $S$ of candidate skills, usually represented by canonical ESCO strings or identifiers. The output is a ranked list $L \subseteq S$ of the top $k$ skills by predicted relevance to $J$ . In TalentCLEF 2026 Task B, the formulation is extended with a skill-type classification function $t(s) \in \{\text{core}, \text{contextual}\}$ over the retrieved skills, so that the system must both rank skills and distinguish universally required skills from those that depend on sector, technology stack, or organizational practice (Gasco et al., 30 Jun 2026).

An alternative but equivalent view is to learn a scoring function

$y = f(q,s),$

where $q$ is a job title, $s$ is a skill, and $y \in [0,1]$ is either a relevance score or a binary decision variable. Classification-oriented systems interpret this score as $S$ 0, typically through a sigmoid head over a title representation, whereas retrieval-oriented systems interpret it as a similarity value between job-title and skill embeddings (Anand et al., 2022).

The task should be distinguished from job-title matching. In job-title matching, the target space is a set of occupation titles; in job title-based skill prediction, the target space is a skill inventory, often much larger and semantically finer-grained. Competition reports explicitly note that mapping job titles to skills is a more semantically demanding problem than title-to-title matching (Zhang et al., 23 Jun 2025).

2. Data resources and supervision regimes

The field uses several supervision regimes, ranging from exact lexical matching in job descriptions to weak supervision and manually curated benchmark judgments. In SkillRec, 6,000 unique job postings were harvested over a six-month period from three major online job boards, and a vocabulary of 589 distinct skill names was compiled from two massive-open-online-course catalogs. Ground-truth labels were assigned by exact string search for each skill name in the cleaned job description, yielding a binary label vector $S$ 1 for each posting (Ong et al., 2023).

Weak supervision replaces direct annotation with distributional signals over similar job titles. In the LaBSE-based method, the starting point is approximately 170 K unique English job titles, each tagged with a noisy set of skills from an English-only in-house taxonomy of approximately 37 K distinct skills. Sentence-BERT embeddings are first used to identify a neighborhood of semantically similar titles with cosine similarity threshold $S$ 2, after which skill frequencies inside that neighborhood are normalized into weak targets $S$ 3. The core assumption is that important skills for a title tend to recur across highly similar titles (Anand et al., 2022).

A third regime derives supervision from structured taxonomies and live vacancies. The knowledge-graph approach defines occupations $S$ 4 from ISCO-08 and skills $S$ 5 from ESCO, then constructs an undirected weighted graph $S$ 6 with $S$ 7 and edges drawn from both canonical ESCO $S$ 8ISCO mappings and co-occurrences extracted from vacancies. In the reported pipeline, approximately 600,000 Dutch vacancies labeled with an ISCO level-4 code are processed with a Textkernel entity-extraction parser, free-text skill mentions are normalized by character $S$ 9-gram Jaccard matching with threshold $L \subseteq S$ 0, and each observed occupation-skill pair increments a co-occurrence count $L \subseteq S$ 1 (Groot et al., 2021).

Shared-task datasets introduce a different type of supervision. TalentCLEF 2025 Task B uses a training split automatically generated from ESCO’s job-to-skill relevance tables, with development and test subsets manually curated from a live recruitment database. The development set contains 304 job titles and 1,439 distinct skills, with 85.2 relevant skills per query on average; the test set contains 125 queries and 1,986 skills, with 101.8 relevant skills per query on average. The skill vocabulary contains 7,493 unique skills, all mapped and normalized to ESCO v1.2.0 lexical entries (Gasco et al., 17 Jul 2025). TalentCLEF 2026 reformulates the same general problem with graded labels, providing a development set of 304 job-title queries and 9,052 candidate skills, and a test set of 125 queries and 5,358 candidate skills, with average relevant-skill counts of 185.6 and 184.2 per query respectively (Gasco et al., 30 Jun 2026).

3. Modeling paradigms

Three modeling families dominate the literature: direct multi-label prediction, dense retrieval in a shared embedding space, and graph-based inference.

Direct multi-label prediction maps a title representation to independent skill probabilities. In the LaBSE system, a title $L \subseteq S$ 2 is encoded as $L \subseteq S$ 3, followed by a single linear layer $L \subseteq S$ 4 with $L \subseteq S$ 5 and a sigmoid over each output dimension. The predicted score for a skill is

$L \subseteq S$ 6

and training minimizes binary cross-entropy against the weak supervision signal (Anand et al., 2022). SkillRec adopts the same multi-label perspective but uses either FastText title embeddings of dimension 300 or BERT-Base embeddings of dimension 768 as input to a three-layer feed-forward network with hidden sizes 1280 and 640, again trained with binary cross-entropy over 589 skills (Ong et al., 2023).

Dense retrieval models instead encode job titles and skills into a common vector space and score them by similarity. JobBERT-V3 uses a shared multilingual MPNet-based transformer backbone for both titles and skill phrases, producing 768-dimensional vectors from the penultimate layer. For skill prediction, the asymmetric 1024-dimensional projection used in the original title-matching setup is omitted, and relevance is computed by cosine similarity,

$L \subseteq S$ 7

The model pre-encodes approximately 30,000 ESCO alias strings and collapses them back to approximately 13,000 skills by keeping the highest-scoring alias per skill (Decorte et al., 29 Jul 2025).

Contrastive learning is the standard training objective for many retrieval systems. In JobBERT-V3, title-skill pairs are optimized with a temperature-scaled InfoNCE loss over positive pairs and in-batch negatives. TalentCLEF 2025 system reports describe the same general pattern for multilingual encoder-based bi-encoders and decoder-based embedding models, using cosine similarity together with InfoNCE or GISTEmbed-style objectives (Gasco et al., 17 Jul 2025). The best-performing 2026 systems extend this paradigm into multi-stage retrieval architectures that combine fine-tuned dense encoders, fusion, and reranking (Gasco et al., 30 Jun 2026).

Graph-based inference models relations between occupations and skills explicitly. In the ESCO/ISCO knowledge graph, link prediction can be performed with Preferential Attachment,

$L \subseteq S$ 8

or with Node2Vec embeddings learned from biased random walks of length $L \subseteq S$ 9, 2500 walks per node, $k$ 0, $k$ 1, and embedding dimension $k$ 2. Inference then uses either the normalized Preferential Attachment score or a dot product in the learned embedding space to rank candidate skill links for an occupation (Groot et al., 2021).

A further distinction concerns discriminative versus embedding-only pipelines. NLPnorth’s TalentCLEF 2025 submission reports that a fine-tuned classification model based on mdeberta-v3-base achieved higher test MAP than its own contrastive and zero-shot prompting variants, using a paired input format of "[JOB_TITLE] [SEP] [SKILL_LABEL_OR_ALT_LABEL]" and standard cross-entropy over relevant versus not relevant (Zhang et al., 23 Jun 2025).

The base relevance score is often modified to privilege specialized skills over ubiquitous ones. The clearest example is the inverse-document-frequency boost used in the LaBSE approach. After training, each skill receives

$k$ 3

where $k$ 4 is the total number of titles and $k$ 5 is the number of titles whose raw skill list contains $k$ 6. The final inference score becomes

$k$ 7

The reported qualitative effect is a shift from generic skills such as “Sales,” “Communication Skills,” and “Excel” toward more specialized outputs such as “Financial Markets,” “Securities,” and “Equities” for the title “Stock Broker” (Anand et al., 2022).

An analogous occupation-level weighting appears in the knowledge-graph setting. Each ISCO group is treated as a document and each skill as a term, with

$k$ 8

Sorting by this weight yields the most distinctive skills for an occupation group. Reported examples include office-software terms for “Managers” and electronics-related terms for “Technicians” (Groot et al., 2021).

Multilingual transfer is now a central property rather than a peripheral one. The LaBSE system applies the same learned skill head to titles in any of 109 languages without additional training, and qualitative examples include German and French inputs such as “Arzt” and “Docteur” producing sensible outputs aligned with the English title “Doctor” (Anand et al., 2022). JobBERT-V3 extends this logic with explicit multilingual contrastive training over English, German, Spanish, and Chinese titles. Because job titles and ESCO skill aliases occupy the same embedding space, a Chinese title can retrieve English ESCO skills directly. The reported top-5 outputs for “Software Developer,” “Desarrollador de Software,” and “软件开发人员” are closely aligned, including “Program software applications,” “Write computer code,” and “Debug software faults,” together with their Spanish and Chinese counterparts (Decorte et al., 29 Jul 2025).

Advanced benchmark systems further refine ranking with augmentation and reranking. TalentCLEF 2025 reports that prompting LLMs to generate short natural-language definitions or contexts for skills and concatenating those texts to the original skill string yielded up to $k$ 9 MAP over baseline fine-tuned encoders of comparable size (Gasco et al., 17 Jul 2025). TalentCLEF 2026 winning systems use pseudo-document generation for terse titles, maximum inner-product search over fused embeddings, and listwise or cross-encoder rerankers operating on the top retrieved candidates; organizers also note that ESCO graph embeddings such as RGCN were used to capture hierarchical relations in difficult synonym-heavy cases (Gasco et al., 30 Jun 2026).

5. Evaluation regimes and empirical results

Evaluation protocols vary with task formulation. Early systems such as SkillRec report exact-match accuracy together with micro-averaged precision, recall, and F1 over skill labels (Ong et al., 2023). Retrieval-style systems report ranking metrics instead: the LaBSE model uses MAP@20 on a held-out test set (Anand et al., 2022), JobBERT-V3 reports MAP, MRR, nDCG, Precision@5, and Precision@10 on TalentCLEF validation data (Decorte et al., 29 Jul 2025), and TalentCLEF benchmarks standardize on MAP in 2025 and graded NDCG in 2026 (Gasco et al., 17 Jul 2025).

Setting	Evaluation	Headline result
SkillRec (Ong et al., 2023)	Accuracy, F1	BERT + NN: 0.9870 accuracy, 0.4973 F1
Weakly supervised LaBSE (Anand et al., 2022)	MAP@20	Fine-tuned LaBSE: 0.722; zero-shot: 0.621
KG link prediction (Groot et al., 2021)	Positive-class F1	PA: 0.78; Node2Vec: 0.65 at 1:1 negatives
JobBERT-V3 (Decorte et al., 29 Jul 2025)	MAP, MRR, nDCG, P@5, P@10	Validation MAP 0.2449; blind test MAP 0.255
TalentCLEF 2025 (Gasco et al., 17 Jul 2025)	MAP	Best team: 0.360; baseline: 0.196
TalentCLEF 2026 (Gasco et al., 30 Jun 2026)	NDCG(graded), NDCG(binary)	Best team: 0.8068 graded, 0.8340 binary

Several empirical patterns recur across studies. In the knowledge-graph setting, Preferential Attachment is stronger than Node2Vec under a balanced 1:1 negative-to-positive ratio, with positive-class $J$ 0 values of 0.78 and 0.65 respectively, but Node2Vec overtakes Preferential Attachment as the negative-to-positive ratio grows from 1 to 7, which was interpreted as robustness in realistic candidate-ranking settings (Groot et al., 2021). In weakly supervised title-only modeling, fine-tuning the LaBSE encoder improves MAP@20 from 0.621 to 0.722, and the IDF boost increases the precision of specialized skills (Anand et al., 2022). In TalentCLEF 2025, the overall leaderboard places pjmathematician first at MAP = 0.360, followed by moali at 0.345 and TechWolf at 0.265, while the competition abstract emphasizes that training strategies have a larger effect than model size alone (Gasco et al., 17 Jul 2025).

The 2026 shared task introduces a finer-grained relevance structure. Skills marked “essential” in ESCO are treated as core skills and skills marked “optional” as contextual skills; graded relevance sets core to 2 and contextual to 1, with graded NDCG as the primary metric. On that benchmark, the top three systems reached NDCG(graded) values of 0.8068, 0.7913, and 0.7793. Organizers report that contrastive fine-tuning of dense embeddings yields a 5–8 point lift in NDCG over out-of-the-box embeddings, prompt-based query expansion via HyDE adds approximately 2 points, listwise neural reranking adds 1–3 points, and ensembles often outperform the best individual retriever by 3–5 points (Gasco et al., 30 Jun 2026).

6. Limitations, disagreements, and research directions

The literature does not support a single universally dominant modeling strategy. One line of work argues that contrastive fine-tuning on domain-specific title-skill pairs is the strongest single ingredient in benchmark performance (Gasco et al., 17 Jul 2025), while another reports that, for the finer-grained job-to-skill mapping in its experimental setup, explicit discriminative supervision outperformed contrastive learning and that a fine-tuned mdeberta-v3-base classifier achieved the best test MAP of 0.290 (Zhang et al., 23 Jun 2025). A plausible implication is that performance depends strongly on the supervision source, candidate inventory, and evaluation protocol rather than on architecture class alone.

Another persistent limitation is incomplete context. The LaBSE approach explicitly notes that its weak labels rely solely on job-title similarity; if similar titles are under-represented or inconsistent, the supervision is noisy. The same work also states that it ignores the full job description, which may contain clue words for rare or highly specialized skills, and that fine-tuning on English slightly degrades performance on other languages even though zero-shot multilingual behavior remains reasonable (Anand et al., 2022). SkillRec exposes a related issue from a different angle: exact string matching against job descriptions can produce noisy labels, especially in sparser job categories, and the authors connect residual errors to the limited and unevenly distributed 6 K posting dataset (Ong et al., 2023).

Benchmark design has also become part of the research problem. TalentCLEF 2026 frames the task as not only relevance retrieval but also skill-type classification, and recommends capping the number of skills per occupation, normalizing skill names, maintaining a clear distinction between “essential” and “optional” abilities, and using multi-stage retrieval $J$ 1 fusion $J$ 2 reranking pipelines (Gasco et al., 30 Jun 2026). This suggests a shift away from purely flat binary relevance toward graded, taxonomy-aware ranking.

The broader research trajectory points toward systems that combine structured taxonomies, labor-market evidence, multilingual encoders, and ranking-oriented training. Knowledge-graph enrichment with live vacancies addresses the dynamic nature of the labor market (Groot et al., 2021); multilingual shared encoders enable title-only inference across languages without task-specific supervision (Decorte et al., 29 Jul 2025); and shared benchmarks make it possible to compare classification, contrastive, prompt-based, and reranking pipelines under common metrics (Gasco et al., 17 Jul 2025). The remaining open problem is not whether job title-based skill prediction can be performed from titles alone, but under which supervision, representation, and evaluation assumptions it remains reliable, specialized, and transferable.