MELS-Benchmark

---
language:
  - de
  - en
  - fr
  - nl
  - sv
license: mit
task_categories:
  - text-retrieval
  - sentence-similarity
tags:
  - entity-linking
  - skills
  - multilingual
  - ranking
  - information-retrieval
  - ESCO
configs:
- config_name: bel_q_fr_c_en
  data_files:
  - split: queries
    path: "bel_q_fr_c_en/queries-00000-of-00001.parquet"
  - split: corpus
    path: "bel_q_fr_c_en/corpus-00000-of-00001.parquet"
- config_name: bel_q_fr_c_fr
  data_files:
  - split: queries
    path: "bel_q_fr_c_fr/queries-00000-of-00001.parquet"
  - split: corpus
    path: "bel_q_fr_c_fr/corpus-00000-of-00001.parquet"
- config_name: bel_q_nl_c_en
  data_files:
  - split: queries
    path: "bel_q_nl_c_en/queries-00000-of-00001.parquet"
  - split: corpus
    path: "bel_q_nl_c_en/corpus-00000-of-00001.parquet"
- config_name: bel_q_nl_c_nl
  data_files:
  - split: queries
    path: "bel_q_nl_c_nl/queries-00000-of-00001.parquet"
  - split: corpus
    path: "bel_q_nl_c_nl/corpus-00000-of-00001.parquet"
- config_name: deu_q_de_c_de
  data_files:
  - split: queries
    path: "deu_q_de_c_de/queries-00000-of-00001.parquet"
  - split: corpus
    path: "deu_q_de_c_de/corpus-00000-of-00001.parquet"
- config_name: deu_q_de_c_en
  data_files:
  - split: queries
    path: "deu_q_de_c_en/queries-00000-of-00001.parquet"
  - split: corpus
    path: "deu_q_de_c_en/corpus-00000-of-00001.parquet"
- config_name: swe_q_sv_c_en
  data_files:
  - split: queries
    path: "swe_q_sv_c_en/queries-00000-of-00001.parquet"
  - split: corpus
    path: "swe_q_sv_c_en/corpus-00000-of-00001.parquet"
- config_name: swe_q_sv_c_sv
  data_files:
  - split: queries
    path: "swe_q_sv_c_sv/queries-00000-of-00001.parquet"
  - split: corpus
    path: "swe_q_sv_c_sv/corpus-00000-of-00001.parquet"
---

# MELS: Multilingual Entity Linking of Skills

MELS is a collection of 8 datasets for evaluating the linking of skill mentions to the
ESCO Skills taxonomy. It covers 3 countries and 4 languages.

## Background

MELS is a sibling dataset to [MELO (Multilingual Entity Linking of Occupations)](https://huggingface.co/datasets/federetyk/MELO-Benchmark).
Both datasets were built using the same methodology and the same type of source data:
crosswalks between national taxonomies and ESCO, published by official labor-related
organizations from EU member states.

The difference is the entity type~~:~~
- **MELO** links occupation mentions (job titles) to ESCO Occupations
- **MELS** links skill mentions to ESCO Skills

MELS covers fewer countries than MELO because fewer EU member states have published
ESCO skill crosswalks. While MELO includes crosswalks from 21+ countries, only 3
countries (Belgium, Germany, Sweden) have published skill crosswalks that could be
used for MELS. This limited scope is why MELS was not published as a standalone
benchmark, but the data remains useful for skill entity linking evaluation.

**2026-01-01 Update**: Austria, Czechia, and Estonia have recently uploaded crosswalks for skills
as well [[*](https://esco.ec.europa.eu/en/use-esco/eures-countries-mapping-tables)]. We plan to
include these in a future version of MELS.

## Dataset Structure

Each subset (configuration) contains two splits:

- **`queries`**: Skill mentions from national taxonomies, with indices of matching ESCO skills
- **`corpus`**: ESCO skill labels

### Schema

**queries split:**
| Column | Type | Description |
|--------|------|-------------|
| `text` | `string` | The skill mention (surface form) |
| `labels` | `list[int]` | Indices of relevant corpus elements |

**corpus split:**
| Column | Type | Description |
|--------|------|-------------|
| `text` | `string` | The ESCO skill label (surface form) |

## Available Subsets

The subset names follow the pattern: `{country}_q_{query_lang}_c_{corpus_lang}`

| Subset | Country | Query Lang | Corpus Lang | # Queries | # Corpus |
|--------|---------|------------|-------------|-----------|----------|
| `bel_q_fr_c_fr` | Belgium | fr | fr | 2,247 | 17,312 |
| `bel_q_fr_c_en` | Belgium | fr | en | 2,247 | 97,520 |
| `bel_q_nl_c_nl` | Belgium | nl | nl | 2,247 | 25,748 |
| `bel_q_nl_c_en` | Belgium | nl | en | 2,247 | 97,520 |
| `deu_q_de_c_de` | Germany | de | de | 1,722 | 19,466 |
| `deu_q_de_c_en` | Germany | de | en | 1,722 | 97,520 |
| `swe_q_sv_c_sv` | Sweden | sv | sv | 4,381 | 19,251 |
| `swe_q_sv_c_en` | Sweden | sv | en | 4,381 | 100,273 |

### Subset Naming Convention

- `{country}`: ISO 3166-1 alpha-3 country code (e.g., `deu` for Germany)
- `q_{lang}`: Query language (ISO 639-1 code)
- `c_{lang}`: Corpus language (ISO 639-1 code)

## Usage

```python
from datasets import load_dataset

# Load a specific subset
ds = load_dataset("Avature/MELS-Benchmark", "deu_q_de_c_de")

# Access the data
query_surface_forms = ds["queries"]["text"]
corpus_surface_forms = ds["corpus"]["text"]
label_lists = ds["queries"]["labels"]

# Example: Get relevant corpus texts for the first query
query_idx = 0
print(f"Query: {query_surface_forms[query_idx]}")
print(f"Relevant ESCO skills:")
for corpus_idx in label_lists[query_idx]:
    print(f"  - {corpus_surface_forms[corpus_idx]}")
```

## Relation to MELO

MELS uses the same methodology as MELO. For details on how the datasets were
constructed from ESCO crosswalks, see the MELO paper and repository:

- **Paper:** [MELO: An Evaluation Benchmark for Multilingual Entity Linking of Occupations](https://recsyshr.aau.dk/wp-content/uploads/2024/10/RecSysHR2024-paper_2.pdf)
- **Repository:** [github.com/Avature/melo-benchmark](https://github.com/Avature/melo-benchmark)
- **HuggingFace:** [Avature/MELO-Benchmark](https://huggingface.co/datasets/federetyk/MELO-Benchmark)

## Citation

If you use this dataset, please cite the MELO paper (which describes the methodology
used to construct both MELO and MELS):

```bibtex
@inproceedings{retyk2024melo,
  title        = {{MELO: An Evaluation Benchmark for Multilingual Entity Linking of Occupations}},
  author       = {Federico Retyk and Luis Gasco and Casimiro Pio Carrino and Daniel Deniz and Rabih Zbib},
  booktitle    = {Proceedings of the 4th Workshop on Recommender Systems for Human Resources
                  (RecSys in {HR} 2024), in conjunction with the 18th {ACM} Conference on
                  Recommender Systems},
  year         = {2024},
  url          = {https://recsyshr.aau.dk/wp-content/uploads/2024/10/RecSysHR2024-paper_2.pdf},
}
```

## License

This dataset is licensed under the MIT License. See the [LICENSE](https://github.com/Avature/melo-benchmark/blob/main/LICENSE) file for more information.