The accurate alignment of occupational titles across multiple languages poses a key challenge in modern talent management systems. Linguistic diferences, semantic ambiguity, and the absence of structured references hinder the automatic identification of equivalent roles in diverse cultural contexts. Although language models have improved significantly, their performance remains limited. To address this, we propose a multilingual system for the semantic matching of job titles in English, Spanish, and German. The approach combines a pre-trained model with a fine-tuning process, using positive and negative examples generated from occupational family relationships defined by the ESCO taxonomy. The goal is to represent job titles within a shared semantic space that enables meaningful cross-lingual comparisons. The results demonstrate high performance in binary classification (F1 score: 0.9573) and information retrieval tasks. English shows the strongest performance (MAP: 0.5057), followed by Spanish (MAP: 0.4071) and German (MAP: 0.3160), confirming the system's ability to operate efectively in multilingual settings. These findings support the use of semantic representations to enhance linguistic interoperability in human resource applications.
The contemporary labor market is undergoing a profound transformation driven by technological
advances and new labor dynamics. Technologies such as automation and artificial intelligence (AI) are
reshaping traditional roles. At the same time, the expansion of remote work has radically changed hiring
and performance conditions [
Within this scenario, Large Language Models (LLMs) have gained strategic relevance in talent
management for both remote and on-site jobs. Generative AI tools, such as ChatGPT, are revolutionizing
key human resources (HR) functions by optimizing processes like recruitment, continuous training, and
career development planning. Recent studies emphasize that LLMs enhance operational eficiency in
Human Capital Management (HCM) by automating repetitive tasks and accelerating decision-making
processes [
However, significant challenges arise with the widespread use of NLP and LLMs in human resource
management. One major challenge is multilingualism: while English dominates much of AI development,
global organizations require support in multiple languages. Current LLMs often fail in languages
other than English, especially in low-resource languages [
In response to these challenges, international initiatives ofer both conceptual and practical support.
ESCO (European Skills, Competences, Qualifications, and Occupations) stands out as a standardized
and multilingual taxonomy at the European level. ESCO functions as a shared dictionary of skills,
qualifications and occupations, available in the 27 oficial EU languages [
Finally, advances in semantic representation are empowering truly multilingual and intelligent HCM
systems. In particular, the use of pre-trained language models enables the generation of vector
representations of texts, job ofers, profiles, and skills that capture their meaning uniformly across languages. It
has enabled automated recommendation and matching functions in multiple languages. Recent studies
have developed approaches in which a multilingual encoder is pre-trained using unsupervised learning
on large datasets of job postings and skills, followed by contrastive fine-tuning using pairs of similar
and dissimilar job ofers based on the ESCO taxonomy. This two-phase training strategy significantly
improves the quality of the resulting representations, more closely aligning the semantic vectors of
job posts with related meanings. As a result, the same model can map job ofers and candidates from
diferent languages into a shared semantic space, enabling direct comparisons[
Research in HCM and HR using NLP techniques has grown significantly over the past five years. Among
the most prominent approaches is the normalization and recommendation of job titles in multilingual
and noisy environments, addressed through contrastive learning strategies that leverage occupational
relationships in taxonomies such as ESCO Figure 1, thus improving semantic alignment [
In addition, other lines of work apply these techniques to tasks such as job vacancy fraud detection [18], demonstrating their versatility. Semantic representations generated by pre-trained models such as
BERT or SBERT, fine-tuned to the occupational domain through skill co-occurrence or textual matching, have also been highlighted [15, 16, 19, 17] . Table 1 summarizes the contributions and techniques employed in these studies. Thus, the present article is situated within this line of research, adopting the approach of semantic representations generated by pre-trained models.
The corpus used in Task A Multilingual Job Title Matching consists of job titles in three main languages: English, Spanish, and German. These titles originate from various occupational domains and professional sectors and have been collected and processed to facilitate the identification and comparison of equivalent occupations across languages. The organization responsible for the task provided the corpus used for the task.
The training set uses public terminologies, ensuring that the included job titles are representative of
a wide range of occupational areas and aligned with standard market nomenclature. In contrast, The
validation and test sets were manually annotated by domain experts appointed by the organizers of the
shared task. As described in the TalentCLEF overview paper [
The training data is presented in tabular format, with three separate files, one for each language. Each entry consists of an occupational family identifier, an occupation identifier, and a pair of related job titles, both associated with the same occupational family. In total, there are three training files: Spanish with 28,880 entries, German with 23,023, and English with 20,724. Altogether, 72,627 job title pairs were grouped by family, where each pair reflects a potential semantic relationship or equivalence between professional roles. This structure allows for the analysis of both textual overlaps and semantic proximity between associated terms.
To illustrate the lexical characteristics of the training data, Figure 5 combines three complementary perspectives: Jaccard similarity (word overlap) Figure 2, Levenshtein similarity (character-level variation) Figure 3, and cosine similarity using TF-IDF (semantic proximity) Figure 4. Most job pairs exhibit low Jaccard similarity, indicating little direct vocabulary overlap. However, Levenshtein similarity reveals shared orthographic patterns, such as gender variations or syntactic changes. In contrast, cosine similarity highlights thematic connections between occupations, even when explicit vocabulary is not shared.
It should be noted that this analysis was conducted on a combined multilingual corpus, which may introduce noise into the lexical metrics. Equivalent titles in diferent languages tend to obtain low scores in Jaccard or Levenshtein because of the lack of textual matching.
An exploratory analysis of the corpus was also conducted, focusing on job title length and lexical distribution by language. On average, Spanish titles tend to be longer in both word and character count, which may indicate greater specificity or redundancy in descriptive expressions. German titles are more concise in terms of token count but often feature complex compound words. English titles fall between the two in both length and lexical complexity.
The corresponding word clouds illustrate the most frequent terms in job titles for each language. In Spanish, technical and operational terms predominate, with explicit gender marking Figure 6. In English, there is a noticeable tendency toward technical and managerial roles, often expressed using gender-neutral language Figure 7. In German, despite some encoding issues afecting specific characters, similar patterns can be observed, including terms such as director and engineer appearing in both their masculine and feminine forms 8.
The validation set is structured into three files per language: queries, corpus elements, and qrels. This setup simulates an information retrieval scenario in which the system’s ability to identify relevant occupations based on a given title is evaluated.
The queries file contains unique identifiers linked to job titles used as queries. The corpus elements ifle includes identifiers for occupations that form the pool of potentially relevant candidates. Finally, the qrels file defines the binary relevance relationship between each query and the corpus elements, assigning a value of 1 when a relevant relationship exists and 0 otherwise.
This structure enables precise evaluation of retrieval models, as it allows for a clear distinction between relevant and non-relevant predictions based on expert-labeled data.
Finally, the test set follows a structure similar to the validation set, with separate files for queries and corpus elements, organized by language. Unlike the validation set, no relevance file is provided in this phase, as participants are required to generate their own predictions in TREC format and submit them for external evaluation.
This phased structure, supervised training, labeled validation, and blind testing supports the progressive development of robust models capable of generalizing efectively in multilingual and semantically ambiguous contexts, such as job title alignment.
Figure 9 illustrates the complete pipeline developed for the training, evaluation, and deployment of a multilingual job title matching model. This process is structured into multiple stages that cover the entire workflow from initial data ingestion to final export of results in TREC format. The key components of this architecture are detailed below.
The process begins with the loading of annotated files in three languages: Spanish (es), English (en), and German (de). Each file contains pairs of job titles along with their corresponding family identifiers, which represent semantically related occupational groupings. Language labels are added and an initial text normalization is performed, such as removing extra whitespace and converting all text to lowercase, to ensure consistency across the data set.
Once individually processed, the three datasets are concatenated into a single DataFrame. This step enables multilingual unified training, which is essential for building a model capable of functioning efectively regardless of the language of the job title.
Two types of job title pairs are generated. Positive examples are composed of titles belonging to the same family identifier, indicating high semantic similarity, and are labeled with 1.0. Negative examples are created through a "hard-negative sampling" strategy, where a job title is paired with another randomly selected from a diferent family, and assigned a label of 0.0. These pairs are transformed into Input Example objects, encapsulating the pair of texts and their respective labels.
The pre-trained model from Sentence-BERT Transformers [20] was fine-tuned using a contrastive learning approach with CosineSimilarityLoss. This loss function optimizes the cosine similarity between sentence embeddings by increasing the similarity of positive job-title pairs while decreasing it for negative ones. The training process included periodic evaluations by language (es, en, de) using a custom callback to monitor multilingual performance.
Model evaluation is conducted on two complementary levels. First, binary classification assesses the model’s ability to accurately distinguish between pairs of similar and dissimilar job titles using standard evaluation metrics such as precision, precision, recall, F1 score, and AUC. Second, an Information Retrieval (IR) evaluation measures the model’s practical utility in search and recommendation tasks. This evaluation employs metrics such as Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), and Precision at K (P@K), computed per language using the queries, corpus, and qrels datasets that simulate real-world search scenarios. The fine-tuned model is saved permanently, allowing it to be reused without retraining, which is crucial for future inference tasks or production deployment.
In the final step, ranking files are generated in TREC format to evaluate the performance of the model in monolingual and cross-lingual retrieval scenarios. All possible combinations between query and corpus languages (e.g., en-en, es-es, de-de, en-es, and en-de) are processed, and cosine similarity scores are used to produce ranked outputs. These results are saved in .trec files for standardized evaluation.
During the training process, the multilingual model was fine-tuned using a cosine similarity loss, resulting in progressive improvements across all evaluation metrics. As shown in Table 2, the training loss steadily decreased from 0.0697 to 0.0339 in the final step, indicating stable convergence of the model. Simultaneously, both precision and F1-score increased, ultimately surpassing 95%, with a final F1-score of 0.9573 and a precision of 0.9574, confirming the model’s ability to efectively distinguish between similar job title pairs.
In terms of information retrieval (IR), the results reveal a notable diference in performance across languages. According to Table 3, the model achieved higher performance in English (EN), with a MAP value of 0.5057 and MRR of 0.7856, whereas Spanish (ESP) and German (GE) yielded MAP values of 0.4071 and 0.3160, respectively. The P@k metrics reinforce this trend: English achieved a P@1 of 0.6286, compared to 0.1297 in Spanish and 0.2956 in German. This suggests that, while the model is robust, its performance is influenced by linguistic characteristics and potential diferences in data quality across languages.
Furthermore, the results of the system evaluation presented in Table 4 show that the average MAP in the three languages (en, es, de) was 0.36, which is consistent with the values obtained during training, particularly when adjusting for variability between pairs of monolingual and cross-lingual. Notably, better performance was observed in English-English matching (0.408) compared to SpanishSpanish (0.348) and German-German (0.324), while cross-lingual matches also demonstrated competitive performance, such as in the English-German case (0.344).
Overall, these results validate the efectiveness of the proposed model in both binary classification and information retrieval tasks. The metrics obtained demonstrate that the system is capable of identifying semantic similarities between professional titles in diferent languages, with outstanding performance across the evaluated languages. These findings support the usefulness of the adopted approach for multilingual occupational alignment tasks, while also highlighting opportunities for improvement in the representation of lower-performing languages.
Considering the results obtained, there remain several opportunities to further enhance the proposed system. One potential direction involves optimizing the model’s performance in lower-performing languages, such as German. This could be addressed through the use of supportive machine translation techniques, synthetic data augmentation, or language-specific fine-tuning, which would allow for better adaptation to the linguistic and semantic particularities of each language.
Additionally, it is worth exploring more advanced methods for generating negative examples, such as those based on dynamic embeddings or supervised contrastive learning strategies, to increase semantic discrimination between non-equivalent titles. This may contribute to improvements in both binary classification and retrieval metrics.
Equally important is the integration of complementary structured information, such as task descriptions, required skills, or hierarchical relationships from the ESCO system. Incorporating these signals could enrich the semantic representation of job titles and enhance the model’s generalization capabilities in real-world scenarios.
By virtue of the results obtained, it is inferred that the proposed system evidences a robust ability to identify semantic similarities between multilingual job titles, with outstanding performance in English. However, areas of opportunity have been identified in languages with lower performance, such as German, suggesting the need to implement language-specific strategies, including data augmentation or targeted optimization techniques.
At the research level, it has been concluded that the incorporation of more refined negative examples, as well as the integration of additional structured data, such as task descriptions or occupational hierarchies, can enrich the semantic representation and optimize the generalization of the model in real contexts. As can be seen from the above, future directions reinforce the value of the adopted approach and point to its evolution towards more accurate, equitable, and adaptive systems for multilingual occupational alignment.
During the preparation of this investigation, ChatGPT (OpenAI) was used for the revision of translations into English, as well as for grammatical and spelling correction. After using this tool, the content was reviewed and edited as necessary, and full responsibility for the content of the publication is assumed.
The authors express their gratitude to the Call 933 “Training in National Doctorates with a Territorial, Ethnic and Gender Focus in the Framework of the Mission Policy — 2023” of the Ministry of Science, Technology and Innovation (Minciencia). In addition, we thank the team of the Artificial Intelligence Laboratory VerbaNex https://github.com/VerbaNexAI, afiliated with the UTB, for their contributions to this project. [11] M. Zhang, R. van der Goot, B. Plank, Escoxlm-r: Multilingual taxonomy-driven pre-training for the job market domain, Proceedings of the Annual Meeting of the Association for Computational Linguistics 1 (2023) 11871–11890. URL: https://arxiv.org/pdf/2305.12092. doi:10.18653/v1/2023. acl-long.662. [12] J. Rosenberger, L. Wolfrum, S. Weinzierl, M. Kraus, P. Zschech, Careerbert: Matching resumes to esco jobs in a shared embedding space for generic job recommendations, Expert Systems with Applications 275 (2025) 127043. URL: https://www.sciencedirect.com/science/article/pii/ S0957417425006657?utm_source=chatgpt.com. doi:10.1016/J.ESWA.2025.127043. [13] M. Beręsewicz, M. Wydmuch, H. Cherniaiev, R. Pater, Multilingual hierarchical classification of job advertisements for job vacancy statistics, 2024. URL: https://arxiv.org/abs/2411.03779. arXiv:2411.03779. [14] P. Safikhani, H. Avetisyan, D. Föste-Eggers, D. Broneske, Discover artificial intelligence automated occupation coding with hierarchical features: a data-centric approach to classification with pretrained language models, Discover Artificial Intelligence 3 (123) 6. URL: https://doi.org/10.1007/ s44163-023-00050-y. doi:10.1007/s44163-023-00050-y. [15] J.-J. Decorte, J. V. Hautte, T. Demeester, C. Develder, Jobbert: Understanding job titles through skills, 2021. URL: https://arxiv.org/abs/2109.09605. arXiv:2109.09605. [16] M. Y. Bocharova, E. V. Malakhov, V. I. Mezhuyev, Vacancysbert: the approach for representation of titles and skills for semantic similarity search in the recruitment domain vitaliy i. mezhuyev 2), Online) Applied Aspects of Information Technology 6 (2023) 52–59. URL: https://doi.org/10.15276/ aait.06.2023.4. doi:10.15276/aait.06.2023.4. [17] D. Lavi, V. Medentsiy, D. Graus, consultantbert: Fine-tuned siamese sentence-bert for matching jobs and job seekers, CEUR Workshop Proceedings 2967 (2021). URL: https://arxiv.org/pdf/2109.06501. [18] K. Taneja, J. Vashishtha, S. Ratnoo, Fraud-bert: transformer based context aware online recruitment fraud detection, Discover Computing 28 (2025) 1–16. URL: https://link.springer.com/article/10. 1007/s10791-025-09502-8. doi:10.1007/S10791-025-09502-8/TABLES/6. [19] M. Yamashita, J. T. Shen, T. Tran, H. Ekhtiari, D. Lee, James: Normalizing job titles with multi-aspect graph embeddings and reasoning, 2023. URL: https://arxiv.org/abs/2202.10739. arXiv:2202.10739. [20] N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2019. URL: http://arxiv.org/abs/1908.10084.