DocEE-zh: Chinese Document-level Event Extraction

Dataset overview

DocEE-zh is a Chinese news benchmark for document-level event extraction, introduced by Liu et al. in Findings of EMNLP 2024. Each document is annotated with one main event and fine-grained arguments, such as people, dates, locations, causes, and event-specific facts. Evidence can span multiple sentences. The benchmark supports event classification and event argument extraction. See the DocEE-zh paper and official repository.

The dataset contains 34,536 records, covering 58 event types and 344 argument types, with 193,351 annotated argument values.

English counterpart: meihant/DocEE provides the English DocEE normal setting, with 59 event types and 356 argument types, based on Tong et al. (NAACL 2022). Both releases use the same four-field format. Their argument vocabularies and split provenance differ; they are not presented as parallel translations.

Event schemas

DocEE-zh extends the DocEE schema framework, covering hard-news and soft-news events. Hard news includes events such as earthquakes and transport accidents; soft news includes celebrity, sports, and entertainment events. These groups guide ontology design; the prediction target is the specific event type.

An event schema specifies the argument roles relevant to an event. Common roles describe dates and locations, while specialized roles describe facts such as earthquake magnitude and focal depth. The role lists for each event type are available in schemas.json.

Data format

Field	Meaning
event_trigger	Original Chinese news headline; a document-level cue, not a token-level trigger annotation
document	Full Chinese news text
event_type	Original event classification label
arguments	List of all roles for the event type; each has role and values

values is a list of all distinct annotated text strings for that role, or JSON null (Python None) when unannotated. Null denotes absence of an annotation, not a claim that the fact does not exist. Every schema role appears exactly once. For event classification, use event_trigger and document as input. For argument extraction conditioned on the event type, additionally supply event_type and the corresponding role list; arguments is the target.

Earthquake example

The following is a complete retained source record, with its full document and all earthquake roles. Data values remain in their original language. For example, 震级 means magnitude and 震源深度 means focal depth. A role with null has no annotated value in this document.

{
  "event_trigger": "希腊克里特岛发生6.5级地震",
  "document": "　　新华社快讯：据欧洲-地中海地震中心消息，希腊克里特岛27日发生6.5级地震。",
  "event_type": "Earthqueakes",
  "arguments": [
    {
      "role": "临时安置点",
      "values": null
    },
    {
      "role": "人员伤亡",
      "values": null
    },
    {
      "role": "余震数量",
      "values": null
    },
    {
      "role": "受困人员数量",
      "values": null
    },
    {
      "role": "受影响地区",
      "values": [
        "希腊克里特岛"
      ]
    },
    {
      "role": "受损建筑",
      "values": null
    },
    {
      "role": "援助物资/数量",
      "values": null
    },
    {
      "role": "支援机构",
      "values": null
    },
    {
      "role": "日期",
      "values": [
        "27日"
      ]
    },
    {
      "role": "疏散数量",
      "values": null
    },
    {
      "role": "经济损失",
      "values": null
    },
    {
      "role": "获救人数",
      "values": null
    },
    {
      "role": "重建房屋数量",
      "values": null
    },
    {
      "role": "震源深度",
      "values": null
    },
    {
      "role": "震级",
      "values": [
        "6.5级"
      ]
    }
  ]
}

Loading

The full configuration contains all 34,536 records. The stratified configuration provides 27,588 training, 3,480 validation, and 3,468 test records, generated with seed 42 and approximate 80/10/10 proportions. Identical document texts are grouped in the same split.

from datasets import load_dataset

splits = load_dataset("meihant/DocEE-zh", "stratified")
full = load_dataset("meihant/DocEE-zh", "full", split="full")
print(splits["train"][0])

Choose either the full configuration or its partitioned counterpart; concatenating both duplicates records.

Citation

Please cite the Chinese dataset paper when using DocEE-zh:

@inproceedings{liu-etal-2024-docee,
    title = "{D}oc{EE}-zh: A Fine-grained Benchmark for {C}hinese Document-level Event Extraction",
    author = "Liu, Minghui and Tong, MeiHan and Peng, Yangda and Hou, Lei and Li, Juanzi and Xu, Bin",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.35/",
    doi = "10.18653/v1/2024.findings-emnlp.35",
    pages = "637--649"
}