conll2003 / README.md
mohpez's picture
Update README.md
89ce44b verified
|
Raw History Blame Contribute Delete
5.27 kB
---
language:
- en
pretty_name: CoNLL-2003
task_categories:
- token-classification
task_ids:
- named-entity-recognition
- part-of-speech
size_categories:
- 10K<n<100K
license: mit
tags:
- ner
- pos
- chunking
- conll
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
---
# Dataset Card for CoNLL-2003
## Dataset Summary
CoNLL-2003 is the English dataset from the CoNLL-2003 shared task on language-independent named entity recognition. It contains newswire sentences from the Reuters corpus, annotated with part-of-speech tags, syntactic chunk tags, and named entity tags. This repository stores the data in Parquet format.
## Supported Tasks
- **Named entity recognition (NER):** predict `ner_tags` for each token.
- **Part-of-speech tagging:** predict `pos_tags`.
- **Chunking:** predict `chunk_tags`.
## Languages
English (`en`).
## Dataset Structure
### Data Splits
| Split | Examples |
|------------|----------|
| train | 14,041 |
| validation | 3,250 |
| test | 3,453 |
| **total** | 20,744 |
### Data Fields
- `id` (string): example identifier.
- `tokens` (list of string): the words of the sentence.
- `pos_tags` (list of int): part-of-speech tag per token.
- `chunk_tags` (list of int): syntactic chunk tag per token.
- `ner_tags` (list of `ClassLabel`): named entity tag per token. Values are stored as integers, and the label names are available through the dataset features (see Usage).
All three tag lists have the same length as `tokens`.
### Example
```json
{
"id": "0",
"tokens": ["EU", "rejects", "German", "call", "to", "boycott", "British", "lamb", "."],
"pos_tags": [22, 42, 16, 21, 35, 37, 16, 21, 7],
"chunk_tags": [11, 21, 11, 12, 21, 22, 11, 12, 0],
"ner_tags": [3, 0, 7, 0, 0, 0, 7, 0, 0]
}
```
### Label Mappings
**`ner_tags`** (BIO scheme; entity types are person, organization, location, miscellaneous)
| id | tag | id | tag |
|----|--------|----|--------|
| 0 | O | 5 | B-LOC |
| 1 | B-PER | 6 | I-LOC |
| 2 | I-PER | 7 | B-MISC |
| 3 | B-ORG | 8 | I-MISC |
| 4 | I-ORG | | |
**`chunk_tags`**
0: O, 1: B-ADJP, 2: I-ADJP, 3: B-ADVP, 4: I-ADVP, 5: B-CONJP, 6: I-CONJP, 7: B-INTJ, 8: I-INTJ, 9: B-LST, 10: I-LST, 11: B-NP, 12: I-NP, 13: B-PP, 14: I-PP, 15: B-PRT, 16: I-PRT, 17: B-SBAR, 18: I-SBAR, 19: B-UCP, 20: I-UCP, 21: B-VP, 22: I-VP
**`pos_tags`** (Penn Treebank tags)
0: `"`, 1: `''`, 2: `#`, 3: `$`, 4: `(`, 5: `)`, 6: `,`, 7: `.`, 8: `:`, 9: ` `` `, 10: CC, 11: CD, 12: DT, 13: EX, 14: FW, 15: IN, 16: JJ, 17: JJR, 18: JJS, 19: LS, 20: MD, 21: NN, 22: NNP, 23: NNPS, 24: NNS, 25: NN|SYM, 26: PDT, 27: POS, 28: PRP, 29: PRP$, 30: RB, 31: RBR, 32: RBS, 33: RP, 34: SYM, 35: TO, 36: UH, 37: VB, 38: VBD, 39: VBG, 40: VBN, 41: VBP, 42: VBZ, 43: WDT, 44: WP, 45: WP$, 46: WRB
## Usage
```python
from datasets import load_dataset
ds = load_dataset("moh-pezeshki/conll2003")
print(ds["train"][0])
```
To convert the NER integers to label names:
```python
names = ds["train"].features["ner_tags"].feature.names
row = ds["train"][0]
print([names[i] for i in row["ner_tags"]])
# ['B-ORG', 'O', 'B-MISC', 'O', 'O', 'O', 'B-MISC', 'O', 'O']
```
## Dataset Creation
### Source Data
The text comes from the Reuters Corpus Volume 1 (RCV1), news stories from August 1996 to August 1997. Annotations were produced for the CoNLL-2003 shared task (Tjong Kim Sang and De Meulder, 2003).
### Provenance of This Copy
This is a Parquet copy of [`lhoestq/conll2003`](https://huggingface.co/datasets/lhoestq/conll2003), maintained by [moh-pezeshki](https://huggingface.co/moh-pezeshki). The only change is that the `ner_tags` column was cast to `ClassLabel`, so label names appear in the viewer and are available via `ds.features["ner_tags"].feature.names`. The stored values are still integers and are identical to the source. The split sizes match the source (14,041 / 3,250 / 3,453).
## Considerations
- The data is newswire from 1996-1997, so entity types, topics, and writing style may not transfer well to modern or non-news text.
- Some sentences are headlines, datelines (e.g. "LONDON 1996-08-22"), or table-like fragments rather than ordinary prose.
- The text is derived from Reuters RCV1, which has its own usage terms. Check those terms before redistributing or using the data commercially.
## Licensing Information
The `license: other` field in the metadata is a placeholder. The underlying text comes from Reuters RCV1, so users should review the RCV1 and CoNLL-2003 terms before use.
## Citation
```bibtex
@inproceedings{tjong-kim-sang-de-meulder-2003-introduction,
title = "Introduction to the {C}o{NLL}-2003 Shared Task: Language-Independent Named Entity Recognition",
author = "Tjong Kim Sang, Erik F. and De Meulder, Fien",
booktitle = "Proceedings of the Seventh Conference on Natural Language Learning at {HLT}-{NAACL} 2003",
year = "2003",
pages = "142--147",
url = "https://www.aclweb.org/anthology/W03-0419"
}
```
## Contact
Maintained by [moh-pezeshki](https://huggingface.co/moh-pezeshki). For issues with this copy, use the Community tab.