Multilingual Skill Extraction: Hard and Soft Skills

This repository centralizes our research on skill extraction and classification from job offers and resumes in a multilingual, multidomain setting. The work was carried out through a multidisciplinary, industry-grounded collaboration among HR researchers, NLP researchers and HR-tech professionals, covering six languages, English, French, German, Italian, Spanish, and Portuguese, for both hard-skill extraction and soft-skill classification.

By Laura Vásquez-Rodríguez, Bertrand Audrin, Samuel Michel, Samuele Galli, Julneth Rogenhofer, Jacopo Negro Cusa and Lonneke van der Plas.

This README instead acts as a single entry point to the publications and the public datasets used to keep our benchmarks reproducible (see Data Availability).

Publications

Skill Extraction from Resumes and Job Offers across Six Languages

📄 Paper · accepted at the 11th Swiss Text Analytics Conference (SwissText), 2026.

TL;DR — We comprehensively evaluate rule-based, semantic, and supervised skill-extraction methods on 1,200 annotated job offers and resumes across six languages and diverse domains. Supervised models reach F1 scores up to 0.6, while rule-based methods offer better interpretability. We find that skills are formulated very differently in resumes vs. job offers, with resumes notably understudied in prior academic work.

Soft Skills in the Wild: Challenges in Multilingual Classification

📄 Paper · accepted at the 10th Swiss Text Analytics Conference (SwissText), 2025.

TL;DR — Using a multilingual BERT-based classifier for soft skills, we compare soft skills, hard skills, and occupations along surface-level and semantic properties. Even when constrained to established taxonomy categories, soft skills show far greater variability in their textual expression than other entity types, a key challenge for recruitment algorithms.

Data Availability

Due to the proprietary nature of the job offers and resumes provided by our industrial partner, the release of the primary dataset is restricted. To ensure reproducibility while respecting these constraints, we adopt two strategies:

1. Experiments on public data. Alongside the proprietary benchmarks, we report results on publicly available job offers and taxonomies (see Public Datasets below).
2. Methodological transparency. Our papers document the methodology in full so the benchmarks remain as reproducible as possible, providing a transparent framework for our real-world skill-extraction approach.

Public Datasets

Dataset	Language(s)	Skill type	Link
Green	English	Hard / ICT skills in job ads	https://huggingface.co/datasets/jjzha/green
ESCO taxonomy	Multilingual (EU)	Skill & occupation taxonomy	https://esco.ec.europa.eu/en/use-esco/download
Fijo	French	Soft skills (insurance job ads)	https://huggingface.co/datasets/jjzha/fijo
Sayfullina	English	Soft skills	https://huggingface.co/datasets/jjzha/sayfullina

Citation

If you use this work, please cite the relevant paper(s):

@inproceedings{vasquez-rodriguez-etal-2026-skill,
    title = "Skill Extraction from Resumes and Job Offers across Six Languages",
    author = "V{\'a}squez-Rodr{\'i}guez, Laura  and
      Audrin, Bertrand  and
      Michel, Samuel  and
      Galli, Samuele  and
      Rogenhofer, Julneth  and
      Cusa, Jacopo Negro  and
      van der Plas, Lonneke",
    booktitle = "Proceedings of the 11th edition of the Swiss Text Analytics Conference",
    year = "2026"
}

@inproceedings{vasquez-rodriguez-etal-2025-soft,
    title = "Soft Skills in the Wild: Challenges in Multilingual Classification",
    author = "V{\'a}squez-Rodr{\'i}guez, Laura  and
      Audrin, Bertrand  and
      Michel, Samuel  and
      Galli, Samuele  and
      Rogenhofer, Julneth  and
      Cusa, Jacopo Negro  and
      van der Plas, Lonneke",
    editor = {Gerber, Jonathan  and
      Cieliebak, Mark  and
      Tuggener, Don  and
      H{\"u}rlimann, Manuela},
    booktitle = "Proceedings of the 10th edition of the Swiss Text Analytics Conference",
    month = may,
    year = "2025",
    address = "Winterthur, Switzerland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.swisstext-1.11/",
    pages = "108--114"
}

Acknowledgments

This research was conducted at Idiap Research Institute as part of the SEM24 project, funded by Innosuisse, the Swiss Innovation Agency, and carried out in collaboration with our industrial partner Arca24, whose job offers and resumes made this multilingual, real-world study possible.