meld-data 1.0.0

MELD: A multilingual and multidomain dataset for named entity recognition (NER)

MELD: Melding Diverse Multilingual and Multi-Domain Datasets for Named Entity Recognition Evaluation

---

MELD is a multilingual and multi-domain dataset for Named Entity Recognition (NER) constructed from 60 existing datasets. Built with reproducibility and extensibility in mind, MELD currently provides gold-standard annotations for 60 languages across up to 14 domains with a total of 601 normalized entity labels. MELD was primarily designed for diverse mutlilingual and multi-domain evaluation but also includes all training and validation sets from its source datasets where available.

Key Features

• Standardized Formats: All datasets are converted to a consistent parquet format, preserving nested and discontinuous annotations, and document boundaries where available.
• Highly Multilingual: Gold-standard annotations for 60 languages and silver-standard annotations derived from Wikipedia for 134 additional languages
• Multi-domain: 14 diverse domains including legal, biomedical, financial, and social media text. Domain diversity is more limited for languages other than English.
• Structural Validation: Several structural issues in source datasets are identified and automatically resolved during processing, such as misaligned span indices and inconsistent IOB labels.
• Reproducible: Fully end-to-end reproducible from published source.
• Extensible: Designed to be extended further through its modular data processing framework. If a data format is already supported, adding new datasets can be as simple as defining a single JSON file.
• Zero-Shot Ready: Provides a normalized entity label mapping specifically designed for zero-shot NER evaluation

Installation

To start working with our dataset, install MELD using pip:

pip install meld-data

For reproducing sentence level tokenization from source, the sentence-segmentation extra needs to be enabled:

pip install 'meld-data[sentence-segmentation]'

For development, we recommend managing your environment with uv:

git clone https://github.com/kgnlp/meld.git
cd meld
uv sync

Listing Available Datasets

To list all datasets available for download:

meld-data list

Download MELD

NOTE: It is recommended to log into a HuggingFace account with huggingface-cli login before downloading datasets to avoid running into API rate limits, particularly when reproducing MELD from source.

To get started, the preprocessed, redistributable subset of MELD can be downloaded using:

meld-data download -v info path/to/download_directory

By default, this downloads the meld:open profile, which includes all dataset available in preprocessed form on the HuggingFace Hub. MELD Open can also be used independently of the meld package. Versions with original kgnlp/meld-open and normalized entity labels kgnlp/meld-open-normalized are available. Datasets not includes in meld:open will be automatically downloaded from their original source and processed locally due to licensing restrictions. To download all datasets including CoNLL-2003:

meld-data download -v info --datasets meld:full path/to/download_directory

NOTE: Currently, the initially downloaded data will contain the original unnormalized entity labels from each dataset. To apply our label normalization, the meld-data hf command can be used.

Notice regarding CoNLL-2003:

Because of copyright restrictions, we cannot redistribute the Reuters Corpus data itself, on which CoNLL-2003 is based. Please refer to the Reuters Corpus licensing information for specific terms and conditions before downloading or using this dataset. This restriction applies only to the CoNLL-2003 dataset and does not affect other datasets in MELD.

MELD Directory Structure

The download_directory passed to meld-data download will contain a downloads subdirectory for source datasets processed by MELD and a meld subdirectory containing the final processed data. The downloads subdirectory can be deleted once data processing is complete. Each dataset in meld contains a meld_metadata.json including additional metadata, statistics, and paths for each subset and split. The NER data itself will be stored in parquet format, optionally in subdirectories for each subset if a dataset includes more than one subset:

download_directory/
├─ downloads/
│  └─ ... # Source datasets processed by MELD
└─ meld/
   ├─ CrossNER/
   │  ├─ meld_metadata.json
   │  ├─ literature/
   │  │  ├─ train.parquet
   │  │  ├─ test.parquet
   │  │  └─ validation.parquet
   │  └─ ...
   ├─ AnatEM/
   │  ├─ meld_metadata.json
   │  ├─ train.parquet
   │  ├─ test.parquet
   │  └─ validation.parquet
   └─ ...

Download Groups of Datasets

Download all redistributable datasets (default):

meld-data download -v info --datasets meld:open path/to/download_directory

Download all non-proprietary datasets that include a test set for evaluation:

meld-data download -v info --datasets meld:non-proprietary-eval path/to/download_directory

Download all non-proprietary datasets including Polyglot-NER:

meld-data download -v info --datasets meld:non-proprietary path/to/download_directory

Download all datasets supported by MELD including CoNLL-2003:

meld-data download -v info --datasets meld:full path/to/download_directory

Profiles and individual dataset names can also be mixed. E.g., for downloading MELD Open and CoNLL-2003:

meld-data download -v info --datasets 'meld:open,CoNLL-2003' path/to/download_directory

Download Specific Datasets

Individual datasets can be downloaded by passing their names as a comma separated list. Dataset names are case-sensitive corresponding to the output of the list command.

meld-data download -v info --datasets "conll-2003,scierc,few-nerd" path/to/download_directory

Reproducing MELD from Source

By default, meld download downloads the already processed version of datasets contained in meld-open to save bandwidth and processing time. To process all datasets from their original source data, add the -r/--reproduce flag. E.g. for reproducing meld:open from source:

meld download -v info -r path/to/download_directory

We use the SAT sentence tokenizer introduced by Frohmann et al. (2024) to tokenize long documents into sentences where no canonical sentence tokenization is available. To avoid slightly different boundaries being generated, e.g., due to GPU non-determinism, sentence boundaries bundled with the MELD package are used by default even when -r/--reproduce is set. To also reproduce the sentence boundaries from scratch, use:

meld download -v info -r --sentence-span-path path/to/new/segmentations path/to/download_directory

Where the directory passed as --sentence-span-path will contain parquet files in the same format as those bundled with the MELD package.

Convert to HuggingFace Datasets Format

The meld-data hf subcommand can be used to convert locally processed MELD data to a format compatible with the HuggingFace datasets library and optionally apply our normalized entity label mapping. For instance, for converting processed datasets belonging to the meld:open subset with normalized entity labels:

meld-data hf -d meld:open --normalize-labels /path/to/processed/meld/data /path/to/converted/data

Note that datasets converted in this way should not be uploaded to the HuggingFace Hub unless the constituent dataset's licensing requirements are fulfilled. See meld-data hf --help for additional options.

Included Datasets

MELD integrates 60 NER datasets spanning 194 languages (60 with gold standard test sets), 14 domains, and 601 normalized entity labels. The table below provides a general overview of the included datasets:

Name	Primary Domain	Languages	Annotation Type	License
AgCNER	Agriculture	zho	gold-standard	CC 0
AgriNER	Agriculture	eng	gold-standard	CC BY-SA 4.0
AnatEM	Biomedical	eng	gold-standard	CC BY-SA 3.0
BC2GM	Biomedical	eng	gold-standard	CC BY 4.0
BC4CHEMD	Biomedical	eng	gold-standard	Unspecified
BC5CDR	Biomedical	eng	gold-standard	Public Domain
BioRED	Biomedical	eng	gold-standard	Public Domain
JNLPBA	Biomedical	eng	gold-standard	GENIA Project License (CC BY 3.0 annotations)
NCBI-Disease	Biomedical	eng	gold-standard	Public Domain
CANTEMIST	Clinical	spa	gold-standard	CC BY 4.0
EBM-NLP	Clinical	eng	gold-standard	Unspecified
RaTE-NER	Clinical	eng	gold-standard, silver-standard	CC BY-NC 4.0
FiNER-139	Finance	eng	gold-standard	CC BY-SA 4.0
TASTEset	Food	eng	gold-standard	MIT
Arabic-Cross-Dialectal-NER	General	apc, ary, arz	gold-standard	Unspecified
Naamapadam	General	asm, ben, guj, hin, kan, mal, mar, ori, pan, tam, tel	gold-standard, silver-standard	CC 0
Thai-NER	General	tha	gold-standard	CC BY 4.0
Turku-NER-corpus	General	fin	gold-standard	CC BY-SA 4.0
TurkuONE	General	fin	gold-standard	CC BY-ND-NC 1.0, CC BY-SA 3.0, CC BY-SA 4.0
NYTK-NerKor	General, Law, Literature, News, Wikipedia	hun	gold-standard	CC BY-SA 4.0
UniversalNER	General, Literature, News, Wikipedia	15 languages	gold-standard	CC BY-SA 4.0
E-NER	Law	eng	gold-standard	CC BY-NC-SA 4.0
German-LER	Law	deu	gold-standard	CC BY 4.0
LegalNERo	Law	ron	gold-standard	CC BY-NC-ND 4.0
Herodotos-Project-NER	Literature	lat	gold-standard	AGPL-3.0 license
CLEANANERCorp	News	ara	gold-standard	GPL 3.0
CoNLL-2003	News	eng	gold-standard	Proprietary text (See Download MELD for details)
EverestNER	News	nep	gold-standard	Non-commercial
FiNER-ORD	News	eng	gold-standard	CC BY-NC 4.0
FoNE	News	fao	gold-standard	CC BY 4.0
idner-news-2k	News	ind	gold-standard	MIT
MasakhaNER-X	News	20 languages	gold-standard	CC BY-NC 4.0
PhoNER-COVID19	News	vie	gold-standard	Research and Education Purposes Only
pioNER	News, Wikipedia	hye	gold-standard, silver-standard	Apache 2.0
FabNER	Science	eng	gold-standard	CC BY 4.0
SciER	Science	eng	gold-standard	GPL 3.0
SCIERC	Science	eng	gold-standard	Unspecified
SciREX	Science	eng	gold-standard	Apache 2.0
SOFC-Exp	Science	eng	gold-standard	CC BY 4.0
SoMeSci	Science	eng	gold-standard	CC BY 4.0
WIESP2022	Science	eng	gold-standard	CC BY 4.0
WLP	Science	eng	gold-standard	MIT
DanfeNER	Social Media	nep	gold-standard	Non-commercial
HarveyNER	Social Media	eng	gold-standard	Unspecified
MIT-Movie	Social Media	eng	gold-standard	Unspecified
MIT-Restaurant	Social Media	eng	gold-standard	Unspecified
Tweebank-NER	Social Media	eng	gold-standard	Apache 2.0
TweetNER7	Social Media	eng	gold-standard	Non-commercial
Weibo-NER	Social Media	zho	gold-standard	CC BY-SA 3.0
WNUT2017	Social Media	eng	gold-standard	CC BY 4.0
StackOverflowNER	Software	eng	gold-standard	MIT
FindVehicle	Transportation	eng	gold-standard	Unspecified
CrossNER	Wikipedia	eng	gold-standard	MIT
Few-NERD	Wikipedia	eng	gold-standard	CC BY-SA 4.0
Japanese-Wikipedia	Wikipedia	jpn	gold-standard	CC BY-SA 3.0
MultiCoNER	Wikipedia	MULTI, ben, deu, eng, fas, fra, hin, ita, por, spa, swe, ukr, zho	silver-standard	CC BY 4.0
MultiNERd	Wikipedia	deu, eng, fra, ita, nld, pol, por, rus, spa, zho	silver-standard	CC BY-NC-SA 4.0
Polyglot-NER	Wikipedia	40 languages	silver-standard	Unspecified
WikiANN	Wikipedia	175 languages	silver-standard	Unspecified
WikiNEuRal	Wikipedia	deu, eng, fra, ita, nld, pol, por, rus, spa	silver-standard	CC BY-NC-SA 4.0

BibTeX Citations

Get citation for MELD:

meld-data cite

Get citations for specific datasets:

meld-data cite conll-2003,scierc

Get citations for all datasets:

meld-data cite --all

Citation

When using MELD, please cite our paper:

@inproceedings{glocker2026meld,
  title = {MELD: Melding Diverse Multilingual and Multi-Domain Datasets for
           Named Entity Recognition Evaluation},
  author = {Glocker, Kevin and Kuhlmann, Marco},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation
               Conference (LREC 2026)},
  month = {May},
  year = {2026},
  pages = {1889--1903},
  address = {Palma, Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  editor = {Piperidis, Stelios and Bel, Núria and van den Heuvel, Henk and Ide,
            Nancy and Krek, Simon and Toral, Antonio},
  doi = {10.63317/32qrd24xac2e},
}

When using the PhoNER COVID19 subset, also cite the following article in accordance with its terms of use:

@inproceedings{PhoNER_COVID19,
  title = {{COVID-19 Named Entity Recognition for Vietnamese}},
  author = {Thinh Hung Truong and Mai Hoang Dao and Dat Quoc Nguyen},
  booktitle = {Proceedings of the 2021 Conference of the North American Chapter
               of the Association for Computational Linguistics: Human Language
               Technologies},
  year = {2021},
}

To retrieve citations for other datasets in MELD, see BibTex Citations.

API Reference

Documentation for the package can be found here.

Contributing

We welcome contributions to expand the dataset! Documentation and guidelines for adding new datasets to MELD are coming soon.

License

This project is licensed under the MIT License. See the LICENSE file for details.