Multilingual Job Title Matching

Updated 6 July 2026

Multilingual job title matching is a method that ranks, retrieves, and normalizes job titles across languages to identify equivalent occupational roles.
It employs techniques like shared-embedding retrieval and hierarchical classification to overcome challenges such as abbreviation, ambiguity, and noise in job title data.
This approach supports applications including career path prediction, talent sourcing, and labor market analysis by integrating structured taxonomies and industrial training data.

Multilingual job title matching is the task of ranking, retrieving, or normalizing job titles across languages so that heterogeneous surface forms refer to the same or highly similar occupational role. In recent work, the task is formulated as multilingual retrieval over structured title corpora and knowledge bases, shared-embedding matching between queries and candidate titles, or hierarchical classification into occupational taxonomies such as ESCO, ISCO, KZiS, and KldB. It supports automatic candidate matching, career path prediction, job market analysis, talent sourcing, recommendation, and labour-market analytics. The task is technically difficult because titles are short, ambiguous, noisy, frequently abbreviated, often gender-marked, and commonly appear inside broader CV or vacancy text rather than as clean canonical labels (Lavi, 2021, Gasco et al., 17 Jul 2025).

1. Problem formulation and semantic scope

Multilingual job title matching is not restricted to exact synonym detection. In the TalentCLEF 2025 formulation, Task A is a retrieval problem: given a query job title, the system must return a ranked list of similar job titles drawn from a corpus, with relevance operationalized as semantic similarity rather than exact synonymy. Annotators judged whether a candidate title plausibly refers to the same or highly similar role, accounting for near-synonyms, abbreviations, duties, seniority, industry context, and specialization or generalization. This makes the task inherently many-to-many: multiple surface forms can be relevant to one query, and the label space is the set of corpus titles rather than a small closed ontology (Gasco et al., 17 Jul 2025).

Industrial matching systems frequently embed titles in a broader document-matching setting. Randstad’s work explicitly models full-text CVs parsed from PDFs and structured vacancy descriptions, arguing that CV and vacancy texts “complement one another like pieces in a puzzle” rather than behaving as paraphrases. In this view, titles are core signals within heterogeneous text, but their meaning is conditioned by surrounding discourse, such as seniority, function, and domain cues. The same bi-encoder method can therefore support title-to-title mapping and title-to-description retrieval in a shared multilingual space (Lavi, 2021).

Recent cross-lingual title models make the equivalence criterion more explicit. JobBERT-V3 defines a match in terms of semantic equivalence after accounting for local naming conventions, seniority markers, and specializations. This includes cases such as English “Senior Software Engineer, Backend” and German “Senior Backend-Entwickler.” The formulation also distinguishes applications that require strict seniority consistency from applications that match at the role-family level (Decorte et al., 29 Jul 2025).

A recurring technical theme is that classical lexical overlap is insufficient. The multilingual barrier is not only translational; it also involves vocabulary gaps, discourse shifts, morphological variation, and domain-specific abbreviations. Examples reported across the literature include “logistics” versus “logistiek,” “instructor” versus “tutor,” and “Warehouse Operator” versus “Logistiek Medewerker.” This motivates methods that learn language-agnostic or code-anchored semantic spaces instead of relying on direct word overlap or inference-time translation (Lavi, 2021, Lavi et al., 2021).

2. Data resources, supervision signals, and taxonomic anchors

The field combines three major forms of supervision: historical matching outcomes, public multilingual benchmarks, and occupational taxonomies. Historical placements provide high-value industrial labels; benchmarks provide comparable public evaluation; taxonomies provide canonical multilingual anchors, aliases, descriptions, and hierarchical structure.

Resource	Languages	Role in matching
Randstad placement history	Multiple markets; examples include Dutch and English	Positive CV–vacancy pairs from actual placements
TalentCLEF 2025 Task A	English, Spanish, German, Chinese	Public multilingual retrieval benchmark
ESCO v1.0.9	27 languages	Occupation and skill taxonomy with preferred and alternative labels
BAfA / KldB data	German	Ontology-aligned title supervision for semantic search
JobBERT-V3 training set	English, German, Spanish, Chinese	Balanced multilingual contrastive corpus of 21,123,868 titles

Randstad constructs training data from its internal history of candidate placements, treating successful CV–vacancy matches as positives and using a cosine similarity log loss that typically implies in-batch negatives. The paper does not report the number of pairs or full language coverage, but it explicitly situates the problem in 38 markets worldwide and highlights multilinguality and cross-language similarity challenges (Lavi, 2021).

TalentCLEF 2025 is the first evaluation campaign focused on skill and job title intelligence. For Task A, the NLPnorth system description reports 15,000 labeled pairs of related job titles per language for English, Spanish, and German training data, 100 manually annotated development samples per language, and 5,000 test job titles with evaluation based on a gold-standard subset of 100 titles per language; development and test materials also include Chinese. The overview paper complements this with corpus-level statistics for manually curated dev and test sets derived from real job pipelines, including preserved noise, abbreviations, typos, incomplete entries, and gender-marked Spanish and German forms (Zhang et al., 23 Jun 2025, Gasco et al., 17 Jul 2025).

ESCO is a central multilingual knowledge source. ESCO v1.0.9 covers 3,008 occupations and 13,890 skills or competences across 27 languages, with language-agnostic concept codes, preferred labels, alternative labels, major groups, descriptions, and explicit occupation–skill relations. This structure supports concept-level normalization: titles in different languages can be mapped to the same ESCO code, and aliases or linked skills can serve as positives during pretraining or downstream retrieval (Zhang et al., 2023).

Several systems instead anchor titles to national or statistical taxonomies. The ontology-aligned embeddings work uses 525,207 German occupational terms from the Bundesagentur für Arbeit, with KldB 2010 as the target hierarchy and a rule-based alignment to ISCED 2011. The Polish vacancy-classification work uses KZiS, a 6-digit hierarchy aligned to ISCO-08, with multilingual expansion to 24 languages via translation and a curated KZiS–ESCO mapping for exact matches (Hihn et al., 5 Sep 2025, Beręsewicz et al., 2024).

Synthetic multilingual supervision has also become prominent. JobBERT-V3 starts from 5,280,967 filtered English title–skill tuples and translates each title into German, Spanish, and Simplified Chinese using prompt-based machine translation with OpenAI gpt-4.1-nano, producing a balanced multilingual corpus of 21,123,868 titles. The cross-lingual signal comes from shared ESCO skill sets rather than task-specific parallel title labels (Decorte et al., 29 Jul 2025).

3. Model families and training objectives

The dominant model family is the multilingual bi-encoder or siamese encoder. Randstad fine-tunes multilingual BERT with a bi-encoder structure and a cosine similarity log loss layer, producing semantically aligned CV and vacancy embeddings. The scoring function is cosine similarity,

$s(x, y) = \frac{x \cdot y}{\|x\| \|y\|}$

and the training objective is described as a cosine similarity log loss with in-batch negatives, canonically instantiated as an InfoNCE-style softmax over matched and mismatched pairs. This design makes nearest-neighbour retrieval tractable and directly transfers to title-to-title or title-to-description matching (Lavi, 2021).

The same general architecture appears in conSultantBERT, which fine-tunes a siamese Sentence-BERT over 274,407 resume–vacancy pairs labeled by staffing consultants. Its best-performing objective is cosine-similarity regression using mean squared error between cosine similarity and a gold label in $[0,1]$ . The resulting multilingual sentence space was shown qualitatively to bridge English–Dutch content without translation, a property that is directly relevant to title matching (Lavi et al., 2021).

Public benchmark systems span at least three paradigms. NLPnorth compares a discriminative classification setup, a contrastive dual-encoder setup, and prompt-based zero-shot embedding retrieval. The classification model places a single linear layer on top of a multilingual encoder and trains with cross-entropy over related versus unrelated pairs. The contrastive model constructs positives such as preferred title $\rightarrow$ description and preferred title $\rightarrow$ alternative title from ESCO, and trains with temperature-free InfoNCE over cosine similarity and in-batch negatives. The prompt-based method uses instruction-tuned LLM embedders with fixed prefixes such as “Given a job title, find the most relevant job titles.” and ranks by cosine similarity without fine-tuning (Zhang et al., 23 Jun 2025).

Structured two-tower retrieval extends this paradigm from short titles to richer documents. In the freelancer-project alignment model, both sides are represented as sets of sections, encoded by a frozen multilingual E5 backbone with learned categorical section-type embeddings, a document-level transformer head, and hierarchical pooling with inverse-length weights. The architecture supports title-only queries, title-plus-skills queries, and full job descriptions while preserving the relative importance of short sections such as titles (Jouanneau et al., 2024).

Ontology-aware language modeling provides a complementary route. ESCOXLM-R performs domain-adaptive pretraining on ESCO descriptions in 27 languages with dynamic masked language modeling and an ESCO Relation Prediction objective. ERP classifies relations among concept pairs as Random, Linked, or Grouped, forcing the encoder to internalize multilingual taxonomic structure. This makes the model especially suited to short occupation and skill spans, and the authors report state-of-the-art results on 6 out of 9 evaluated downstream datasets (Zhang et al., 2023).

A separate line of work casts the task as hierarchical multi-class classification. In multilingual job-ad classification for vacancy statistics, the posterior over a leaf code is factorized as

$P(y_6 \mid x) = \prod_{d=1}^{6} P(y_d \mid x, y_{<d}),$

allowing prediction to proceed from broad occupational groups to specific codes. In the 2021 multi-label description-to-title model, the task is instead cast as multilabel prediction over a standardized taxonomy of 68 normalized job titles, with multilingual BERT producing the highest instance-level F1 among compared models (Beręsewicz et al., 2024, Tran et al., 2021).

4. Retrieval pipelines, indexing, and production deployment

The practical advantage of bi-encoder title matching is that candidate titles, vacancies, profiles, or ontology nodes can be pre-embedded offline and searched with approximate nearest-neighbour infrastructure. Randstad emphasizes that the bi-encoder yields a maintainable and scalable pipeline, unlike cross-encoders that require pairwise concatenation at inference. The same logic underlies title catalogs, ESCO occupation indices, and multilingual corpora of historical job titles (Lavi, 2021).

ANN infrastructure is now standard in deployment-oriented systems. The structured two-tower retriever stores freelancer embeddings in Qdrant, updates them daily, encodes incoming job queries online, and applies optional hard filters such as geography before re-ranking. The deployment achieved 95th percentile latency $\leq 3$ seconds, versus tens of seconds previously, and an A/B test reported a 5.63% improvement in conversion for effective matches. The paper also details how empty sections are encoded as [CLS] and [END], which stabilizes computation graphs while marking missing content (Jouanneau et al., 2024).

conSultantBERT describes the same serving pattern in title-centric terms: embed a query title, search an ANN index such as FAISS or HNSW over a multilingual title catalog, apply a cosine threshold, and optionally aggregate variants into canonical parents. The model uses 768-dimensional embeddings from bert-base-multilingual-cased; mean pooling and cosine scoring are sufficient for scalable semantic retrieval (Lavi et al., 2021).

Ontology-aligned systems phrase classification itself as semantic search. The German KldB work inserts enriched title embeddings into an HNSW similarity graph, retains top- $k$ edges above a similarity threshold, and assigns KldB or ISCED labels by majority or weighted vote among neighbours. Because classification is reframed as nearest-neighbour search, new titles and even new ontologies can be appended incrementally without retraining the entire system (Hihn et al., 5 Sep 2025).

Efficiency remains a central argument for dense retrieval. Randstad’s discussion of SBERT notes the standard comparison in which finding the most similar pair in 10,000 sentences drops from roughly 65 hours with cross-encoder BERT or RoBERTa to roughly 5 seconds with SBERT. JobBERT-V3 continues this efficiency-first design: it uses a multilingual MPNet encoder with a lightweight projection head, produces 1024-dimensional title embeddings for matching, and is roughly half the size of multilingual-e5-large-instruct while remaining competitive on multilingual ranking (Lavi, 2021, Decorte et al., 29 Jul 2025).

5. Benchmarks, metrics, and empirical performance

Mean Average Precision is the dominant public metric for multilingual title retrieval. In the NLPnorth TalentCLEF 2025 submission, prompting was the strongest Task A test-time method, with English 0.537, German 0.496, Spanish 0.442, Chinese 0.495, and average 0.493; across English, Spanish, and German, the abstract reports average MAP $\approx 0.492$ . The best contrastive model reached average 0.480, and the best classification model 0.440. On validation, contrastive and prompt-based systems were essentially tied, suggesting that ESCO-derived contrastive supervision and large instruction-tuned embedders are both competitive for multilingual title retrieval (Zhang et al., 23 Jun 2025).

The TalentCLEF overview broadens the comparison to the full leaderboard. In official monolingual results, AlexU-NLP achieved average MAP 0.534, TechWolf 0.517, pjmathematician 0.515, NLPnorth 0.492, and the baseline 0.360. In cross-lingual evaluation over English queries against Spanish and German corpora, pjmathematician and AlexU-NLP tied at average MAP 0.514. The Chinese track, despite having no provided Chinese training data, reached 0.516 MAP for pjmathematician, while AlexU-NLP obtained the highest overall average including Chinese at 0.527. The overview further concludes that training strategy matters more than sheer model size, since carefully fine-tuned encoder systems often matched or outperformed much larger decoder-based embedding models (Gasco et al., 17 Jul 2025).

JobBERT-V3 reports stronger validation numbers on the same benchmark setting. Its validation MAP is 0.6302 for English, 0.4562 for German, 0.5090 for Spanish, and 0.5845 for Chinese, all above the multilingual MPNet base; versus multilingual-e5-large-instruct, it reports higher MAP and nDCG in all four languages while using roughly half the parameters. On blind test data, it reaches en–en 0.533, es–es 0.519, de–de 0.500, zh–zh 0.510, en–es 0.510, en–de 0.498, and en–zh 0.515, indicating consistent monolingual and cross-lingual behavior without task-specific parallel supervision (Decorte et al., 29 Jul 2025).

Older industrial evidence is aligned with these benchmark results, although measured on broader document-matching tasks. conSultantBERT reports a best ROC-AUC of 0.8459 for its regression objective plus cosine similarity on resume–vacancy matching, improving substantially over TF-IDF and untuned BERT baselines. The same paper reports precision 0.7714, recall 0.7664, and F1 0.7677 for the best cosine-based run, and presents qualitative cross-lingual English–Dutch heatmaps showing successful alignment without translation (Lavi et al., 2021).

Classification-oriented title normalization remains viable when the label set is fixed. In job-title prediction from multilingual job descriptions, multilingual BERT achieved 62.20% instance-level F1 on the development set and 47.44% on the test set, outperforming XLM-R, PhoBERT, DistilBERT, and multiple word-embedding baselines. This result concerns description-to-title prediction rather than open-ended retrieval, but it shows that multilingual title semantics can also be learned effectively as supervised multilabel classification (Tran et al., 2021).

6. Fairness, failure modes, and research directions

Fairness has become a first-class concern because language and grammatical gender can act as proxy variables in labour-market systems. Randstad explicitly notes that the language of a CV correlates with nationality and can act as a proxy discriminator, and observes that multilingual matching may reduce the language barrier so candidates are matched regardless of CV language unless the vacancy requires a specific language. The same work situates HR in the context of high-risk AI under the European Commission’s Artificial Intelligence Act discussion (Lavi, 2021).

TalentCLEF incorporated an explicit fairness track for Spanish and German using Rank-Biased Overlap on gendered query pairs. The overview reports that NLPnorth and Ixa both achieved perfect $\mathrm{RBO}=1.000$ in both languages, indicating rankings invariant to gendered forms; Ixa achieved this by translating titles into English to remove grammatical gender. High-performing systems such as AlexU-NLP and TechWolf also showed very low sensitivity without explicit debiasing, with average RBO values of 0.976 and 0.971 respectively (Gasco et al., 17 Jul 2025).

The dedicated gender-bias study shows, however, that out-of-the-box multilingual sentence encoders still exhibit varying degrees of gender bias. It evaluates Spanish, German, French, and Portuguese by comparing rankings generated from masculine and feminine query forms over the same corpus. The paper uses a finite full-list average-overlap formulation,

$\mathrm{RBO} = \frac{1}{k}\sum_{d=1}^{k}\frac{|S_{:d}\cap T_{:d}|}{d},$

with $[0,1]$ 0 equal to the full list length. MAP and parity do not move together: on masculine corpora, multilingual E5 large achieves the best average MAP at 0.3901, while MiniLM yields the strongest parity in Spanish, French, and Portuguese. A qualitative Spanish example shows that abogada produces more junior or assistant-like results than abogado, while the masculine form ranks more lawyer specializations (García-Sardiña et al., 17 Sep 2025).

Failure modes are recurrent across studies. Reported issues include low-resource languages and dialects, rare or emerging titles, abbreviation-heavy strings such as “PM,” polysemy in titles such as “Consultant,” noisy PDF parsing, train–test distribution shift, synthetic-translation artifacts, and mismatch between clean taxonomy labels and organically noisy resume or job-offer strings (Lavi, 2021, Zhang et al., 23 Jun 2025). Ontology-based systems add another difficulty: national taxonomies and international ontologies are often many-to-many aligned, so ambiguity should be surfaced rather than silently collapsed (Hihn et al., 5 Sep 2025).

Several future directions are already explicit in the literature. These include more languages, richer hierarchical labels, stronger use of ESCO or other ontologies, hybrid lexical–neural retrieval, deeper fairness analysis beyond RBO, human-in-the-loop evaluation, active learning for taxonomy drift, and cross-lingual code-anchored alignment using ontology supervision rather than direct translation. A plausible implication is that multilingual job title matching is converging toward an overview of dense retrieval, structured occupational knowledge, and fairness-aware evaluation, rather than toward a single universal architecture (Gasco et al., 17 Jul 2025, Hihn et al., 5 Sep 2025, Zhang et al., 2023).