VerbaNex at TalentCLEF 2026: A Hybrid Multilingual Retrieval and Approach for Contextualized Job-Person Matching

EasyChair Preprint 16023

Abstract

This paper presents the participation of the VerbaNex AI team in TalentCLEF 2026 Task A: Contextualized Job-Person Matching, a shared task focused on retrieving and ranking the most relevant candidates for job vacancies in multilingual environments. To address this challenge, we propose a hybrid retrieval architecture that combines dense semantic retrieval based on multilingual embeddings with lexical retrieval using BM25, followed by a neural re-ranking stage employing a multilingual cross-encoder.

The proposed framework includes document preprocessing, job posting enrichment, hybrid candidate retrieval, and contextualized reranking. Candidate profiles and job descriptions are represented using the BGE-M3 multilingual bi-encoder, enabling semantic matching across English and Spanish. Retrieval scores obtained from dense and lexical approaches are normalized and combined through weighted fusion, while a cross-encoder refines the final ranking by modeling deeper interactions between job requirements and candidate qualifications.

Experiments were conducted using the official TalentCLEF 2026 Task A dataset under monolingual and cross-lingual retrieval settings. The proposed system achieved an average MAP of 0.560, obtaining MAP scores of 0.558 for English-English retrieval and 0.554 for Spanish-Spanish retrieval. Additionally, the system reached MRR values of 0.859 and 0.840 in the respective monolingual scenarios. Results demonstrate the effectiveness of combining semantic and lexical retrieval strategies for contextualized talent matching, while also highlighting the remaining challenges associated with cross-lingual candidate retrieval and ranking.

Keyphrases: Candidate Ranking, Information Retrieval, Job-Person Matching, Multilingual NLP

@booklet{EasyChair:16023,
  author    = {Melissa Moreno and Juan Carlos Martínez-Santos and Edwin Puertas},
  title     = {VerbaNex at TalentCLEF 2026: A Hybrid Multilingual Retrieval and Approach for Contextualized Job-Person Matching},
  howpublished = {EasyChair Preprint 16023},
  year      = {EasyChair, 2026}}