Datasets:
annotations_creators:
- expert-generated
language:
- et
language_creators:
- found
license:
- cc-by-nc-4.0
multilinguality:
- monolingual
paperswithcode_id: noisyner
pretty_name: NoisyNER
size_categories:
- 10K<n<100K
source_datasets:
- original
tags:
- newspapers
- 1997-2009
task_categories:
- token-classification
task_ids:
- named-entity-recognition
Dataset Card for NoisyNER
Table of Contents
- Table of Contents
- Dataset Description
- Dataset Structure
- Dataset Creation
- Considerations for Using the Data
- Additional Information
Dataset Description
- Repository: Estonian NER corpus, NoisyNER dataset
- Paper: Named Entity Recognition in Estonian, Analysing the Noise Model Error for Realistic Noisy Label Data
- Dataset: NoisyNER
- Domain: News
Dataset Summary
NoisyNER is a dataset for the evaluation of methods to handle noisy labels when training machine learning models.
- Entity Types:
PER,ORG,LOC
It is from the NLP/Information Extraction domain and was created through a realistic distant supervision technique. Some highlights and interesting aspects of the data are:
- Seven sets of labels with differing noise patterns to evaluate different noise levels on the same instances
- Full parallel clean labels available to compute upper performance bounds or study scenarios where a small amount of gold-standard data can be leveraged
- Skewed label distribution (typical for Named Entity Recognition tasks)
- For some label sets: noise level higher than the true label probability
- Sequential dependencies between the labels
For more details on the dataset and its creation process, please refer to the original author's publication https://ojs.aaai.org/index.php/AAAI/article/view/16938 (published at AAAI'21).
This dataset is based on the Estonian NER corpus. For more details see https://aclanthology.org/W13-2412/
Supported Tasks and Leaderboards
Languages
The language data in NoisyNER is in Estonian (BCP-47 et)
Dataset Structure
Data Instances
An example of 'train' looks as follows.
{
'id': '0',
'tokens': ['Tallinna', 'õhusaaste', 'suureneb', '.'],
'lemmas': ['Tallinn+0', 'õhu_saaste+0', 'suurene+b', '.'],
'grammar': ['_H_ sg g', '_S_ sg n', '_V_ b', '_Z_'],
'ner_tags': [5, 0, 0, 0]
}
Data Fields
The data fields are the same among all splits.
id: astringfeature.tokens: alistofstringfeatures.lemmas: alistofstringfeatures.grammar: alistofstringfeatures.ner_tags: alistof classification labels (int). Full tagset with indices:
{'O': 0, 'B-PER': 1, 'I-PER': 2, 'B-ORG': 3, 'I-ORG': 4, 'B-LOC': 5, 'I-LOC': 6}
Data Splits
The splits are the same across all configurations.
| train | validation | test |
|---|---|---|
| 11365 | 1480 | 1433 |
Dataset Creation
Curation Rationale
Source Data
Initial Data Collection and Normalization
Tkachenko et al (2013) collected 572 news stories published in the local online newspapers Delfi and Postimees between 1997 and 2009. Selected articles cover both local and international news on a range of topics including politics, economics and sports. The raw text was preprocessed using the morphological disambiguator t3mesta (Kaalep and Vaino, 1998) provided by Filosoft. The processing steps involve tokenization, lemmatization, part-of-speech tagging, grammatical and morphological analysis.
Who are the source language producers?
Annotations
Annotation process
According to Tkachenko et al (2013) one of the authors manually tagged the corpus and the other author examined the tags, after which conflicting cases were resolved. The total size of the corpus is 184,638 tokens. Tkachenko et al (2013) provide the following number of named entities in the corpus:
| PER | LOC | ORG | Total | |
|---|---|---|---|---|
| All | 5762 | 5711 | 3938 | 15411 |
| Unique | 3588 | 1589 | 1987 | 7164 |
Hedderich et al (2021) obtained the noisy labels through a distant supervision/automatic annotation approach. They extracted lists of named entities from Wikidata and matched them against words in the text via the ANEA tool (Hedderich, Lange, and Klakow 2021). They also used heuristic functions to correct errors caused by non-complete lists of entities, grammatical complexities of Estonian that do not allow simple string matching or entity lists in conflict with each other. For instance, they normalized the grammatical form of a word or excluded certain high false-positive words. They provide seven sets of labels that differ in the noise process. This results in 8 different configurations, when added to the original split with clean labels.
Who are the annotators?
Personal and Sensitive Information
Considerations for Using the Data
Social Impact of Dataset
Discussion of Biases
Other Known Limitations
Additional Information
Dataset Curators
Licensing Information
Citation Information
@inproceedings{tkachenko-etal-2013-named,
title = "Named Entity Recognition in {E}stonian",
author = "Tkachenko, Alexander and
Petmanson, Timo and
Laur, Sven",
booktitle = "Proceedings of the 4th Biennial International Workshop on {B}alto-{S}lavic Natural Language Processing",
month = aug,
year = "2013",
address = "Sofia, Bulgaria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/W13-2412",
pages = "78--83",
}
@article{Hedderich_Zhu_Klakow_2021,
title={Analysing the Noise Model Error for Realistic Noisy Label Data},
author={Hedderich, Michael A. and Zhu, Dawei and Klakow, Dietrich},
volume={35},
url={https://ojs.aaai.org/index.php/AAAI/article/view/16938},
number={9},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2021},
month={May},
pages={7675-7684},
}
Contributions
Thanks to @phucdev for adding this dataset.