DocEE: Document-level Event Extraction

Dataset overview

DocEE is a benchmark for extracting events and their arguments from full news documents, introduced by Tong et al. at NAACL 2022. Its annotations describe the main event in each document and the people, organizations, locations, times, and other event-specific information associated with it. Arguments can occur across multiple sentences, making the dataset suitable for studying document-level understanding. See the paper for the dataset construction process and original experiments.

The English ontology contains 59 event types and 356 distinct argument types. The event types comprise 31 hard-news types and 28 soft-news types, with a dedicated argument schema for each event type.

This dataset reformats the official English DocEE dataset, using its normal_setting train/dev/test files. It contains one main event per document and preserves the original split membership and retained row order. The upstream dev split is named validation here. The cross-domain setting is not included.

Chinese counterpart: meihant/DocEE-zh provides Chinese news event extraction data, based on Liu et al., Findings of EMNLP 2024. It contains 34,536 records, 58 event types, and 344 argument types, with full data and reproducible train/validation/test splits. Both datasets use the same four-field format.

This Hugging Face release provides 27,478 documents, with all roles from the corresponding event schema included in every record.

How the event schema is designed

DocEE draws on the distinction between hard news and soft news in journalism:

Event group	Number of event types	Focus	Examples
Hard news	31	Urgent events with immediate public consequences	Earthquakes, road crashes, armed conflicts
Soft news	28	Human-interest, celebrity, sports, and entertainment stories	Celebrity marriages, sports competitions, awards ceremonies

These are broad groups used to design the ontology; each document receives one of the 59 specific event labels. See the original repository and paper, Section 3.1.

For each event type, the authors first examined 20 Wikipedia event pages and used shared infobox keys as candidate argument roles. They then collected 20 New York Times stories of that type. Five journalism students identified additional facts readers would want to know; their suggestions were merged, trial annotation was conducted, and rarely occurring roles were filtered out. This produces both common roles, such as dates, and event-specific roles, such as earthquake magnitude or flood water level. Schema construction details.

An event schema therefore specifies which kinds of information to extract for a given event type. The ontology has 356 distinct role names overall; each event uses only its own subset. For example, the released Earthquakes schema has 15 roles, including Epicenter, Magnitude, Affected Area, and Casualties and Losses. The complete role lists are available in schemas.json.

Supported tasks

• Event classification: predict the main event type from the event trigger and document.
• Event argument extraction: extract the annotated text values for each role in an event schema.
• Joint event extraction: predict the event type and its arguments together.

The event_type field and arguments are gold annotations. For event classification or joint extraction, use the event trigger and document as input and these annotations as supervision. The event ontology is available separately in schemas.json.

Dataset statistics

Property	Value
Language	English
Setting	Normal
Documents / main events	27,478
Event types	59 (31 hard news + 28 soft news)
Argument types (distinct role names)	356
Argument text values	228,692

Argument values are counted once per distinct text within each document and role.

Split	Documents	Argument text values
Train	21,960	190,511
Validation	2,748	19,000
Test	2,770	19,181
Total	27,478	228,692

The official normal-setting split membership is retained; seven malformed records were excluded.

Fields

Exactly four fields are provided per record:

Field	Type	Meaning
event_trigger	string	Original news title, used as event-trigger context (not an annotated trigger span).
document	string	Original full document, unchanged.
event_type	string	Original event classification label.
arguments	list of objects	One {role, values} entry for every schema role. values is a list of all annotated texts for that role, or JSON null if unannotated.

Every schema role appears exactly once. Unannotated roles have JSON null values, which become Python None. Annotated roles contain a list of distinct text values in first-occurrence order: identical strings within the same document and role are included only once. schemas.json provides the complete event ontology.

Example: an earthquake event

The following is a complete record from the test split, about the 1933 Diexi earthquake. Its document text and annotations are reproduced as stored in this release. It demonstrates the four fields and all 15 roles in the Earthquakes schema.

{
  "event_trigger": "1933 Diexi earthquake",
  "document": "The 1933 Diexi earthquake occurred in Diexi, Mao County, Szechwan, Republic of China on August 25 with a moment magnitude of 7.3 and a maximum Mercalli intensity of X (Extreme). With up to 9,300 killed, this was the deadliest earthquake of 1933. This earthquake destroyed the town of Diexi and surrounding villages, and caused many landslides, and killed about 9,000 people. [4] The old town of Diexi sank into the landslide dam-created Diexi Lake. ",
  "event_type": "Earthquakes",
  "arguments": [
    {"role": "Number of Rebuilding House", "values": null},
    {"role": "Number of Destroyed Building", "values": null},
    {"role": "Economic Loss", "values": null},
    {"role": "Epicenter", "values": null},
    {"role": "Number of Trapped People", "values": null},
    {"role": "Number of Aftershocks", "values": null},
    {"role": "Aid Agency", "values": null},
    {"role": "Number of Evacuated People", "values": null},
    {"role": "Casualties and Losses", "values": ["up to 9,300 killed"]},
    {"role": "Affected Area", "values": ["Diexi, Mao County, Szechwan, Republic of China"]},
    {"role": "Temporary Settlement", "values": null},
    {"role": "Date", "values": ["1933", "August 25"]},
    {"role": "Number of Rescued People", "values": null},
    {"role": "Magnitude", "values": ["7.3"]},
    {"role": "Aid Supplies/Amount", "values": null}
  ]
}

Here, event_trigger is the source title and event_type is the specific class Earthquakes, a hard-news event type. Magnitude has one annotated value, while Date has two. Roles without an annotation, such as Economic Loss, remain present with values: null (Python None). Null means that this role has no gold annotation; it does not establish that the corresponding fact is absent in the real world. Values are distinct text strings within each document and role.

The record does not contain a separate schema field: the role names are listed in arguments, and the reusable ontology is provided in schemas.json.

Usage

from datasets import load_dataset

ds = load_dataset("meihant/DocEE", "normal")
row = ds["train"][0]
arguments_by_role = {entry["role"]: entry["values"] for entry in row["arguments"]}
# arguments_by_role[role] is None or a list of all annotated strings.

The JSONL files can also be read directly with Python's standard json module. For local use with datasets, supply the three split-specific globs under data_files to load_dataset("json", data_files=...).

For example, an unannotated role is represented as {"role": "Date", "values": null} in JSON. An annotated role is represented as {"role": "Date", "values": ["Friday evening", "Saturday"]}. These are format illustrations rather than a dataset record.

Resources

• Original dataset repository
• Paper (NAACL 2022)
• Official normal-setting data
• Event schemas

Conversion code, source information, and detailed validation statistics are included in scripts/prepare_dataset.py, source_manifest.json, and validation_report.json. The source split membership, including document overlap, is preserved.

License

No explicit data license was provided in the inspected upstream repository or download files. This reformat does not assign a new license; original dataset and article rights remain with their respective owners.

Citation

If you use this dataset, please cite the original DocEE paper:

MeiHan Tong, Bin Xu, Shuai Wang, Meihuan Han, Yixin Cao, Jiangqi Zhu, Siyu Chen, Lei Hou, and Juanzi Li. 2022. DocEE: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3970–3982, Seattle, United States. Association for Computational Linguistics. Paper and citation. DOI: 10.18653/v1/2022.naacl-main.291.

@inproceedings{tong-etal-2022-docee,
    title = "{D}oc{EE}: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction",
    author = "Tong, MeiHan and Xu, Bin and Wang, Shuai and Han, Meihuan and Cao, Yixin and Zhu, Jiangqi and Chen, Siyu and Hou, Lei and Li, Juanzi",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.291/",
    doi = "10.18653/v1/2022.naacl-main.291",
    pages = "3970--3982"
}

To identify this particular preprocessing, also reference meihant/DocEE and record the dataset revision used in your experiments.