Datasets:

meihant
/

DocEE

+---
+language:
+- en
+pretty_name: DocEE English — Normal Setting (Schema Complete)
+task_categories:
+- text-classification
+- token-classification
+size_categories:
+- 10K<n<100K
+tags:
+- event-extraction
+- document-level-event-extraction
+- docee
+configs:
+- config_name: normal
+  default: true
+  data_files:
+  - split: train
+    path: data/train-*.jsonl
+  - split: validation
+    path: data/validation-*.jsonl
+  - split: test
+    path: data/test-*.jsonl
+---
+# DocEE English: normal setting, complete argument schema
+This dataset reformats the official English [DocEE dataset](https://github.com/tongmeihan1995/DocEE), 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. Chinese data and the cross-domain setting are not included.
+| Split | Source rows | Removed malformed rows | Released rows |
+|---|---:|---:|---:|
+| train | 21,966 | 6 | 21,960 |
+| validation | 2,748 | 0 | 2,748 |
+| test | 2,771 | 1 | 2,770 |
+| Total | 27,485 | 7 | 27,478 |
+## Fields
+Exactly five fields are provided per record:
+| Field | Type | Meaning |
+|---|---|---|
+| `title` | string | Original source title, unchanged. |
+| `document` | string | Original full document, unchanged. |
+| `event_type` | string | Original event classification label. |
+| `schema` | list of strings | All roles for this event in the official schema, in schema order. |
+| `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. |
+JSON `null` becomes Python `None`. Missing roles are never omitted and never represented by an empty list or the string `"None"`. Every schema role appears exactly once. The list of role/value objects keeps the nested feature structure consistent across event types without adding unrelated roles from other event schemas.
+All text annotations are retained in source order within each role, **including repeated strings and repeated annotations**. No trimming, translation, deduplication or span reconstruction is applied. Offsets are validated against the source document but omitted from this text-list output format. `schemas.json` contains the complete official 59-event schema.
+```python
+from datasets import load_dataset
+ds = load_dataset("meihant/DocEE-normal", "normal")
+row = ds["train"][0]
+arguments_by_role = {entry["role"]: entry["values"] for entry in row["arguments"]}
+assert list(arguments_by_role) == row["schema"]
+# 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=...)`.
+## Cleaning and validation
+Malformed **whole records are dropped, not repaired**. Required title/document/label fields must be nonempty strings. Labels and argument roles must belong to the official schema. Annotations must have nonblank string text, valid integer character offsets, and exact agreement between the text and `document[start:end]`.
+Seven rows fail these checks: train source index 19587 has a `NaN` title; train indices 5369, 5677, 5690, 15683, 15922 and test index 261 contain blank argument text. Indices are zero-based in the original split files. All retained annotation offsets match their document text.
+Each serialized output row is read back and compared with its retained source row. The check covers the title, document, event label, ordered complete role schema, and exact ordered text list for every role. The release retains 353,896 argument annotations. Split counts, removal reasons, file hashes, event counts, and annotation statistics are in `validation_report.json`.
+The source has duplicate documents and cross-split document overlap; these are preserved to retain the official normal setting. After malformed-row removal, duplicate document occurrences beyond the first are train 34, validation 0, test 27. Exact document-text hash intersections contain 52 distinct documents between train/validation, 12 between train/test, and 25 between validation/test. Account for this when interpreting evaluation results. Repeated argument strings are particularly common in the source training data and are preserved as requested.
+## Provenance
+- [Official normal setting folder](https://drive.google.com/drive/folders/1fUuFM0Dm59KUKA1EFP3rk8GONJU1s-ud).
+- [Official event schema, pinned version](https://github.com/tongmeihan1995/DocEE/blob/0a3f43f2f5bfba0459c485d1bff9e5ad4d3eb6fe/Event_Schema.md).
+- Retrieved on 2026-09-07. Source file identifiers and SHA-256 hashes are in `source_manifest.json` and `validation_report.json`.
+- Conversion code is provided under `scripts/prepare_dataset.py`; it uses Python 3.11's standard library.
+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
+Tong et al. (2022), *DocEE: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction*, NAACL 2022. See the [official repository](https://github.com/tongmeihan1995/DocEE) for the authors' citation instructions and paper link.

data/test-00000.jsonl ADDED Viewed