HUMAID-NER

HUMAID-NER is an English disaster-tweet dataset for joint humanitarian-category classification and span-level named entity recognition. It extends the HumAID benchmark with character-offset entity annotations across ten operational entity types.

Supported tasks

• Token classification or span-based named entity recognition
• Multiclass humanitarian information classification
• Joint and multitask learning
• Crisis-information extraction and retrieval research

Dataset structure

Each Parquet row contains:

Field	Type	Description
record_id	string	Stable identifier derived from the exact text.
text	string	English disaster-related social-media text.
humanitarian_category	string	Document-level humanitarian category slug.
entities	list of structs	Character-offset entity annotations.
entity_count	integer	Number of annotated entities in the record.

Each entity struct contains start_char, end_char, text, entity_type, and annotation_source. Offsets are zero-based Unicode code-point positions and end_char is exclusive.

Data splits

Split	Records
Train	53,531
Validation	7,788
Test	15,159

The release contains 245,180 entity spans. Record identifiers are unique across the release, and the three splits are disjoint.

Humanitarian categories

• caution_and_advice
• displaced_people_and_evacuations
• infrastructure_and_utility_damage
• injured_or_dead_people
• missing_or_found_people
• not_humanitarian
• other_relevant_information
• requests_or_urgent_needs
• rescue_volunteering_or_donation_effort
• sympathy_and_support

Entity types

• LOCATION
• DISASTER_TYPE
• CASUALTY
• DISPLACED
• REQUEST
• RESOURCE
• RESCUE
• ORGANIZATION
• PERSON
• NUMBER

Definitions and the legacy BIO label mapping are provided in metadata/label_definitions.json.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset("AijazAli7/humaid-ner")
print(dataset["train"][0])

For custom local loading:

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={
        "train": "data/train.parquet",
        "validation": "data/validation.parquet",
        "test": "data/test.parquet",
    },
)

Annotation process

Entity annotations were generated with a reproducible hybrid pipeline combining transformer-based spaCy recognition, disaster-domain EntityRuler patterns, structured regular expressions, and priority-based overlap resolution. The annotation source for every span is preserved as spacy or regex.

Intended use

The dataset is intended for non-commercial research on crisis informatics, disaster-response NLP, named entity recognition, humanitarian text classification, and multitask learning. Appropriate uses include model development, benchmarking, error analysis, and reproducibility studies.

Limitations and responsible use

• Entity annotations are generated by a hybrid automated pipeline and may contain false positives, missed entities, or boundary errors.
• Social-media language, locations, disaster types, and communication practices may shift over time.
• Class and entity distributions are imbalanced; report per-class metrics as well as aggregate scores.
• Text may contain sensitive, distressing, or personally identifying content posted during emergencies. Do not use it to identify, profile, or contact individuals.
• Models trained on this dataset should not be used as the sole basis for emergency-response decisions.
• Users remain responsible for following the source platform's terms, the dataset license, and applicable law.

Dataset distributions

• Hugging Face — direct loading with the Hugging Face datasets library and split-based Parquet files.
• Kaggle — Parquet splits, normalized CSV tables, metadata, and supporting documentation.

License

HUMAID-NER is an adaptation of the HumAID dataset published by the Qatar Computing Research Institute under CC BY-NC-SA 4.0.

HUMAID-NER adds disaster-specific span-level entity annotations, standardized labels, validated character offsets, and research-ready data formats. This adapted release is distributed under CC BY-NC-SA 4.0. The original authors and QCRI do not necessarily endorse this adaptation.

• License: https://creativecommons.org/licenses/by-nc-sa/4.0/
• Upstream dataset: https://huggingface.co/datasets/QCRI/HumAID-all
• HUMAID-NER publication: https://doi.org/10.62019/zabvxd97
• Original HumAID publication: https://doi.org/10.1609/icwsm.v15i1.18116

Citation

If you use this dataset, cite both HUMAID-NER and the original HumAID dataset.

@article{ali2026humaidner,
  title   = {HUMAID-NER: A Disaster Tweet Dataset for Joint Named Entity Recognition and Event Classification via Uncertainty-Weighted Multitask Learning},
  author  = {Ali, Aijaz and Basir, Nazish and Nawaz, Sarfaraz and Arain, Danish Nazir and Ali, Haris},
  journal = {The Asian Bulletin of Big Data Management},
  volume  = {6},
  number  = {1},
  pages   = {138--152},
  year    = {2026},
  doi     = {10.62019/zabvxd97}
}

@article{alam2021humaid,
  title   = {HumAID: Human-Annotated Disaster Incidents Data from Twitter with Deep Learning Benchmarks},
  author  = {Alam, Firoj and Qazi, Umair and Imran, Muhammad and Ofli, Ferda},
  journal = {Proceedings of the International AAAI Conference on Web and Social Media},
  volume  = {15},
  number  = {1},
  pages   = {933--942},
  year    = {2021},
  doi     = {10.1609/icwsm.v15i1.18116}
}

References

• HUMAID-NER paper: https://arxiv.org/abs/2609.16964
• Published article DOI: https://doi.org/10.62019/zabvxd97
• Official QCRI HumAID distribution: https://huggingface.co/datasets/QCRI/HumAID-all
• Original HumAID paper: https://doi.org/10.1609/icwsm.v15i1.18116