This paper describes our approach for TalentCLEF 2025 Task A, focusing on multilingual and cross-lingual job title matching. The core challenge lies in the inherent brevity and ambiguity of job titles across diferent languages (English, Spanish, and German) and professional sectors. Our methodology employs a curriculum learning strategy to fine-tune an embedding model, gradually exposing it to more complex data involving job titles and their descriptions. We further enhance retrieval performance through a hybrid system combining semantic search with BM25 keyword matching, followed by a multilingual cross-encoder reranker. Experimental results on the validation set demonstrate the efectiveness of our phased training approach and hybrid retrieval, achieving a top average mAP of 56% on the validation set and an average mAP of 53% on the test set.
The modern workplace has undergone a profound transformation in recent years, driven by technological innovation, globalization, and shifting social dynamics. Technological advancements—particularly in artificial intelligence and natural language processing—are reshaping how companies source, assess, and manage human capital. At the same time, the globalization of the workforce, enabled by remote hiring and digital collaboration tools, has introduced new complexities in matching candidates to job roles across linguistic and cultural boundaries. These developments require intelligent systems capable of handling large-scale, multilingual data while maintaining the semantic integrity of role descriptions and candidate profiles.
A central challenge in this space is the variability and ambiguity of job titles. Job titles are often brief, under-specified, and highly context-dependent.
Furthermore, diferent organizations frequently use distinct terms to describe similar roles. For instance, the positions of “Software Engineer,” “Backend Developer,” and “Platform Engineer” may share significant overlap in responsibilities, yet difer in naming conventions based on organizational or regional preferences. This terminological inconsistency becomes even more pronounced in multilingual contexts, where translation, cultural nuance, and domain-specific jargon further complicate the task of semantic alignment.
Task A of TalentCLEF 2025 [
In this notebook, we present AlexU-NLP’s approach to Task A, which combines curriculum learning, hybrid retrieval, and re-ranking to address the multifaceted challenges of job title normalization and retrieval. Our methodology employs a curriculum-based fine-tuning strategy for a multilingual embedding model, progressively introducing more complex data—from isolated job titles to rich titledescription pairs—thereby enhancing the model’s capacity to learn meaningful semantic representations. To maximize retrieval performance, we adopt a hybrid strategy that integrates dense semantic search with BM25-based sparse retrieval, followed by a multilingual cross-encoder re-ranker.
This system not only achieves strong performance on the TalentCLEF 2025 validation and test sets but also demonstrates practical relevance in real-world Human Capital Management (HCM) scenarios. By improving the robustness and adaptability of job title matching systems, our approach supports more accurate talent identification and enhances the alignment between workforce capabilities and organizational needs in a multilingual, cross-sector labor market.
In the evolving landscape of job recommendation systems, recent research has focused on enhancing
the semantic understanding of job titles and descriptions to improve candidate-job matching. Zbib
et al.[
Complementing this, Laosaengpha et al. [
Furthermore, Deniz et al. [
The corpus for Task A [
• Training Data: Provided as 15,000 pairs of related job titles per language (English, Spanish,
German). • Validation Data: Structured into three distinct files per language: queries, corpus elements, and qrels (query relevance assessments). This set contains 100 query job titles per language, each with a list of related job titles from the corpus. A knowledge base of 2,500 unique job titles per language serves as the corpus for retrieval tasks within the validation set. • Test Data: A background set comprising 5,000 job titles. The evaluation is conducted on a subset of the background set, that will be a gold standard corpus of 100 job titles in each language.
To enhance the contextual understanding of job titles during model fine-tuning and inference, we
implemented a two-pronged data augmentation strategy. First, for enriching the training dataset, we
used the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy [
We selected the multilingual-E5-large-instruct model [
We implemented a curriculum learning approach that incrementally increases the complexity of the training data. This allows the model to first establish a robust understanding of job title semantics before integrating the richer contextual information from job descriptions. A key aspect of our methodology is its emphasis on cross-lingual learning. For instance, for each English job title, our training data included not only pairs with its relevant English title and description but also pairs with translations of both the title and its description into German and Spanish. This systematic exposure to semantically equivalent information across languages is designed to encourage the alignment of embeddings in the multilingual space, thereby mitigating language-specific clustering. The curriculum was structured into several stages. These are briefly described in the following subsections, while full details regarding their specific configurations can be found in Section 5.4.
The curriculum commenced with training on symmetric (job title, job title) pairs. These initial stages prioritized monolingual data, with the model being trained on pairs within each language (en-en, es-es, de-de) before progressing to cross-lingual title pairs (en-es, en-de). The primary objective of these early stages was to preserve and refine the pretrained model’s intrinsic ability to align job titles based purely on their semantics, forming a strong foundation for subsequent learning.
Following the initial alignment, the curriculum gradually introduced asymmetric pairs of (job title, job title + job description). This was done in a controlled manner, balancing these richer contextual pairs with the title-only examples from the preceding stages. Similar to the early stages, this phase also began with same-language pairs before incorporating cross-lingual pairs. This part of the curriculum acted as a form of soft domain adaptation, injecting more extensive contextual cues to help disambiguate job titles and discourage overfitting to short, potentially ambiguous title tokens, while simultaneously reinforcing the symmetric retrieval structure learned earlier.
The concluding stages of the curriculum placed a strong emphasis on the asymmetric (job title, job title + job description) format. This was intended to allow the model to fully adapt to the anticipated real-world inference conditions, where job descriptions are expected to provide significant contextual information.
Our retrieval architecture employs a two-stage process, consisting of an initial hybrid retrieval phase followed by a neural reranking mechanism.
In the first stage, we perform a hybrid search by combining signals from dense and sparse retrieval methods.
• For dense retrieval, query job titles and corpus entries (each comprising a job title and its corresponding description) are encoded into vector representations using our fine-tuned 5, model; relevance is then scored using cosine similarity. • Concurrently, for sparse retrieval, we utilize the BM25 algorithm to compute lexical similarity scores between the query (title and description) and each corpus entry (title and description).
The relevance scores from these two retrieval components are first normalized to a common range. These normalized scores, denoted as 25, for BM25 and , for the vector-based semantic similarity, are then integrated using a weighted linear fusion. The final fused score, , for each candidate document is computed as:
= (25 · 25,) + ( · ,) Based on empirical evaluation on our validation set, we determined the optimal weights to be 25 = 0.15 for the BM25 component and = 0.85 for the semantic vector component.
The resulting candidate list, ranked by , is subsequently passed to the second stage, where
we used a fine-tuned version of jina-reranker-v2-base-multilingual cross-encoder model
released by Jina AI [
We established a baseline using the paraphrase-multilingual-MiniLM-L12-v2 model [
We evaluated several pretrained multilingual embedding models in a
zeroshot setting on the validation data. The models tested included BGE-M3 [
The fine-tuning experiments utilized Multiple Negatives Ranking Loss (MNRL). The positive instances for this loss consisted of pairs structured as either (job title, job title) or (job title, job title + description), where the first element consistently served as the anchor and the second as the positive example. We conducted preliminary fine-tuning experiments: • Fine-tuning multilingual 5, on (job title, job title) pairs resulted in overfitting. • Fine-tuning multilingual 5, on (job title, job title + description) pairs showed substantial improvement over the title-only approach and zero-shot performance.
We also experimented with ESCOXLM-R [
The cross-lingual curriculum learning strategy for the multilingual 5, model was implemented in six stages. Throughout each curriculum stage, we also employed MNRL. The data composition for each stage was as follows: • Stage 1: 100% monolingual (job title, job title) pairs (en-en, es-es, de-de). • Stage 2: 60% monolingual (job title, job title) pairs; 40% cross-lingual (job title, job title) pairs (en-es, en-de). • Stage 3: 60% monolingual (job title, job title); 20% cross-lingual (job title, job title); 20% monolingual (job title, job title + description). • Stage 4: 30% monolingual (job title, job title); 20% cross-lingual (job title, job title); 30% monolingual (job title, job title + description); 20% cross-lingual (job title, job title + description). • Stage 5: 10% monolingual (job title, job title); 10% cross-lingual (job title, job title); 60% monolingual (job title, job title + description); 20% cross-lingual (job title, job title + description). • Stage 6: 70% monolingual (job title, job title + description); 30% cross-lingual (job title, job title + description).
For the second-stage refinement of our retrieval pipeline, we employed the jina-reranker-v2-base-multilingual cross-encoder model. This model was specifically ifne-tuned for the task using a Binary Cross-Entropy (BCE) loss function. The training data was formulated from (query, candidate document, label) tuples. Each candidate document consisted of a job title concatenated with its corresponding description. Positive instances were created using known relevant query-document pairs ( = 1), while hard negative instances ( = 0) were incorporated to improve the model’s discriminative power. These hard negatives were mined from the corpus by utilizing the jina-embedding-v3 model [14] to retrieve documents (job titles and their descriptions) that were highly ranked for a given query but were non-relevant.
We also explored using a Large Language Model for reranking. The top 10 documents retrieved by our dense retrieval (curriculum-trained 5,) were presented to the Gemma3-27B model [15]. The model was prompted to reorder these candidates based on relevance to the query. The specific prompt template utilized for this task is detailed in Listing 1:
Listing 1: LLM Reranking Prompt Template Below is a query and a list of 10 candidate job titles.
Rank them from most relevant (rank=1) to least relevant (rank=10).
Query: {query} Candidates: {candidates} Return only the ordering as comma-separated numbers (e.g. 3,1,2,...): In the prompt above, {query} is replaced with the actual query job title, and {candidates} is replaced with a numbered list of the top-10 candidate job titles and their descriptions retrieved by the dense model.
All results reported in this section are Mean Average Precision (mAP) scores obtained on the oficial validation set, unless otherwise specified for test set evaluations.
To establish a baseline and select a strong foundation model, we evaluated several pretrained multilingual embedding models in a zero-shot setting. Table 1 presents these results. The 5, model achieved the highest average mAP of 47.27%, outperforming the oficial task baseline by approximately 9%. This superior performance justified its selection as our base model for fine-tuning.
We then investigated the impact of fine-tuning on performance, initially without employing the full curriculum learning strategy. Table 2 shows the results for 5, and ESCOXLM-R (with various pooling strategies) when fine-tuned on pairs of (job title, job title + job description). Finetuning 5, improved its average mAP from 47.27% (zero-shot) to 53.82%. This underscores the value of incorporating job descriptions, which provide essential context for disambiguating job titles. The 5, model also consistently outperformed the ESCOXLM-R variants, despite the latter being pretrained specifically on ESCO data. We hypothesize that the base architecture and the instruction-tuning of the E5 model make it more amenable for fine-tuning on this specific task structure.
Table 3 illustrates the performance of our system, which incorporates the full cross-lingual curriculum learning strategy for 5,, and subsequently evaluates the impact of diferent reranking 5, (LLM reranking) 5, (BM25 & cross-encoder reranking) Language mAP (dev)
mAP (test) English German Spanish Average English German Spanish Average English German Spanish Average approaches. The application of our cross-lingual curriculum learning strategy boosted the average mAP of 5, on the development set to 55.76%, a substantial improvement from the 53.82% achieved with direct fine-tuning (Table 2). Further enhancements were observed with reranking. • The hybrid retrieval approach yielded the highest average mAP on the validation set (55.97% mAP). • LLM-based reranking with Gemma3-27B also showed a slight improvement over the curriculumonly model on the validation set (55.82% mAP).
On the test set, both the curriculum-trained model without further reranking and our system (curriculum + BM25 + cross-encoder reranker) achieved an average mAP of 53%. This indicates good generalization for the curriculum-trained model, though the slight gains from reranking observed on the validation set did not fully translate to the test set for all configurations.
A core component of our methodology is the enrichment of job titles with full descriptions, either sourced from ESCO or generated by an LLM. To explicitly validate the necessity of this step during inference, we conducted an ablation study on the validation set. We compared the performance of our ifnal curriculum-trained model under two conditions: 1) the standard approach, where query titles are used to retrieve from a corpus of job title + description, and 2) a title-only approach, where query titles retrieve from a corpus containing only job titles, with descriptions omitted.
The results, presented in Table 4, confirm the significant benefit of including descriptions. The average mAP dropped from 55.76% to 53.18% when descriptions were removed from the inference corpus. This performance decrease underscores that the contextual information provided by the descriptions is crucial for the model to disambiguate similar or ambiguous job titles, justifying our data augmentation strategy for the inference phase. Curriculum E5 (Title + Description) 0.5226 Curriculum E5 (Title Only) 0.4992
English
All experiments were conducted on a single NVIDIA A100 GPU. We analyzed the trade-of between retrieval quality and computational cost for our primary system configurations. The baseline curriculumtrained E5-large-instruct model ofers the lowest latency, requiring only a single embedding pass for the query followed by a fast vector search. Our final system, which adds BM25 (negligible overhead) and a cross-encoder reranker, introduces higher latency by requiring 10 additional forward passes per query. This cost yielded a marginal mAP improvement (55.76% to 55.97% on validation), making the system suitable for applications where accuracy is paramount. The LLM-based reranking using Gemma3-27B was the most resource-intensive approach, incurring substantial latency and memory usage for a minimal performance gain on the validation set and a drop on the test set, proving impractical for this task.
Despite the promising results, this study has certain limitations that should be acknowledged. Firstly, our data augmentation strategy relies on LLM-generated descriptions for inference, and while eforts were made to ensure contextual plausibility, the synthetic data may not fully capture the stylistic diversity or factual nuances of authentic ESCO or real-world job descriptions. Discrepancies in quality or representativeness could subtly influence model performance on job titles reliant on these synthetic contexts. Secondly, the system’s evaluation is based on the English, Spanish, and German languages within the TalentCLEF dataset. Its generalization to entirely diferent languages or highly specialized job domains not well-represented in the training data requires further investigation; the observed lower performance for German titles across several models may hint at existing cross-lingual representation challenges. Thirdly, the current work did not incorporate specific mechanisms to assess or mitigate potential performance disparities across job titles that may be stereotypically associated with diferent gender groups; addressing this is an important consideration for ensuring fairness in practical applications. Finally, the optimal configuration of various components, such as curriculum learning parameters, hybrid retrieval weights, or the selected LLM prompts, was determined based on validation set performance. As is common in such iterative development, this process may have led to some degree of adaptation to the validation data’s specific characteristics, as suggested by the slight variations in performance gains between the validation and test sets for some of our reranking configurations.
Our participation in TalentCLEF 2025 Task A demonstrated the eficacy of a carefully designed curriculum learning strategy combined with data augmentation, hybrid retrieval, and cross-encoder reranking for multilingual job title matching. The phased introduction of complexity and cross-lingual signals enabled our model to achieve strong performance. For future work, we plan to explore more sophisticated hard negative mining techniques for both the embedding model and the reranker.
Further investigation and experimenting with larger, more capable LLMs for reranking (perhaps with more elaborate prompting) could also yield improvements. Additionally, we intend to improve the synthetic data generation process by ensuring that the descriptions generated by the LLM more accurately reflect the linguistic style, content scope, and structural properties inherent in authentic ESCO descriptions.
During the preparation of this work, the authors used Gemini 2.5 Pro in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.