Datasets:

alvations
/

complex-word-id-2018

pretty_name: CWI 2018 — Complex Word Identification Shared Task
license: other
license_name: cwi-2018-shared-task
license_link: https://sites.google.com/view/cwisharedtask2018/
task_categories:
  - token-classification
  - text-classification
language:
  - en
  - de
  - es
  - fr
tags:
  - complex-word-identification
  - cwi
  - lexical-simplification
  - readability
  - bea-2018
  - shared-task
size_categories:
  - 10K<n<100K
configs:
  - config_name: english-news
    data_files:
      - split: train
        path: data/english-news/train.parquet
      - split: validation
        path: data/english-news/validation.parquet
      - split: test
        path: data/english-news/test.parquet
  - config_name: english-wikinews
    data_files:
      - split: train
        path: data/english-wikinews/train.parquet
      - split: validation
        path: data/english-wikinews/validation.parquet
      - split: test
        path: data/english-wikinews/test.parquet
  - config_name: english-wikipedia
    data_files:
      - split: train
        path: data/english-wikipedia/train.parquet
      - split: validation
        path: data/english-wikipedia/validation.parquet
      - split: test
        path: data/english-wikipedia/test.parquet
  - config_name: german
    data_files:
      - split: train
        path: data/german/train.parquet
      - split: validation
        path: data/german/validation.parquet
      - split: test
        path: data/german/test.parquet
  - config_name: spanish
    data_files:
      - split: train
        path: data/spanish/train.parquet
      - split: validation
        path: data/spanish/validation.parquet
      - split: test
        path: data/spanish/test.parquet
  - config_name: french
    data_files:
      - split: test
        path: data/french/test.parquet

CWI 2018 — Complex Word Identification Shared Task

Repackaged release of the Complex Word Identification (CWI) Shared Task 2018 corpus (BEA-13 @ NAACL 2018), converted from the original TSV distribution into parquet with a unified schema across all four languages.

The shared task asks systems to predict whether a target word/phrase in context would be hard to understand for non-native speakers, children, or readers with language disabilities. It supports both a binary classification task (complex vs. simple) and a probabilistic task (proportion of annotators marking the target as difficult).

Configs

Config	Language	Source	Train	Validation	Test
`english-news`	en	Professionally edited news	14,002	1,764	2,095
`english-wikinews`	en	WikiNews articles	7,746	870	1,287
`english-wikipedia`	en	English Wikipedia	5,551	694	870
`german`	de	German Wikipedia	6,151	795	959
`spanish`	es	Spanish Wikipedia	13,750	1,622	2,233
`french`	fr	French Wikipedia	—	—	2,251

French is the cross-lingual zero-shot track: test only, no training data.

Usage

from datasets import load_dataset

ds = load_dataset("alvations/complex-word-id-2018", "english-news")
print(ds["train"][0])
# {'hit_id': '3P7RGTLO6EE07HLUVDKKHS6O7CCKA5',
#  'sentence': 'The barren islands, reefs and coral outcrops ...',
#  'start_offset': 4, 'end_offset': 10, 'target': 'barren',
#  'native_annotators': 10, 'non_native_annotators': 10,
#  'native_complex': 6, 'non_native_complex': 2,
#  'label_binary': 1, 'label_prob': 0.4, 'language': 'en'}

# Cross-lingual eval on French
fr = load_dataset("alvations/complex-word-id-2018", "french", split="test")

Schema

All configs share one schema. Test splits have null label fields (gold labels were never released publicly with the test inputs — only the shared-task organisers held them).

Column	Type	Description
`hit_id`	string	MTurk HIT identifier — sentences sharing a HIT were annotated together in one screen.
`sentence`	string	The full sentence containing the target.
`start_offset`	int32	Character offset where the target begins in `sentence`.
`end_offset`	int32	Character offset where the target ends (exclusive).
`target`	string	The candidate complex word or multiword expression.
`native_annotators`	int32	Number of native-speaker annotators who saw the sentence.
`non_native_annotators`	int32	Number of non-native annotators who saw the sentence.
`native_complex`	int32?	Native annotators who marked the target as difficult. (null in test)
`non_native_complex`	int32?	Non-native annotators who marked the target as difficult. (null in test)
`label_binary`	int32?	0 = simple (no annotator marked it complex); 1 = complex (≥1 annotator marked it). (null in test)
`label_prob`	float32?	Probability label = `(native_complex + non_native_complex) / (native_annotators + non_native_annotators)`. (null in test)
`language`	string	ISO 639-1 language code (`en`, `de`, `es`, `fr`).

Annotation protocol

English: each sentence annotated by 10 native + 10 non-native speakers.
German / Spanish / French: each sentence annotated by 10 annotators (mixed native + non-native). The native_annotators and non_native_annotators columns reflect the actual split per sentence (and do not always sum to 10 across all rows due to the shared-task's annotator-pooling).

Annotators saw paragraph context around each sentence and were asked to mark words likely to be hard for children, non-native speakers, or readers with language disabilities.

Tasks

Binary CWI: predict label_binary. Standard metric: macro-F1.
Probabilistic CWI: predict label_prob ∈ [0, 1]. Standard metric: MAE.
Cross-lingual transfer: train on english-* / german / spanish, evaluate on french (no French training data exists).

Sources

English (news / wikinews / wikipedia): a mix of professionally-edited news articles, WikiNews articles, and English Wikipedia.
German: German Wikipedia.
Spanish: Spanish Wikipedia.
French: French Wikipedia (test-only, for cross-lingual evaluation).

Citation

Primary citation — the shared task report (W18-0507):

@inproceedings{yimam-etal-2018-report,
    title     = "A Report on the Complex Word Identification Shared Task 2018",
    author    = "Yimam, Seid Muhie  and
                 Biemann, Chris  and
                 Malmasi, Shervin  and
                 Paetzold, Gustavo  and
                 Specia, Lucia  and
                 {\v{S}}tajner, Sanja  and
                 Tack, Ana{\"\i}s  and
                 Zampieri, Marcos",
    booktitle = "Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications",
    month     = jun,
    year      = "2018",
    address   = "New Orleans, Louisiana",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/W18-0507/",
    doi       = "10.18653/v1/W18-0507",
    pages     = "66--78",
}

Underlying corpora — please also cite the dataset papers the shared task drew on:

English (News / WikiNews / Wikipedia) — I17-2068:

@inproceedings{yimam-etal-2017-cwig3g2,
    title     = "{CWIG}3{G}2 - Complex Word Identification Task across Three Text Genres and Two User Groups",
    author    = "Yimam, Seid Muhie  and
                 {\v{S}}tajner, Sanja  and
                 Riedl, Martin  and
                 Biemann, Chris",
    booktitle = "Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers)",
    month     = nov,
    year      = "2017",
    address   = "Taipei, Taiwan",
    publisher = "Asian Federation of Natural Language Processing",
    url       = "https://aclanthology.org/I17-2068/",
    pages     = "401--407",
}

German / Spanish / French (multilingual) — RANLP 2017, R17-1104:

@inproceedings{yimam-etal-2017-multilingual,
    title     = "Multilingual and Cross-Lingual Complex Word Identification",
    author    = "Yimam, Seid Muhie  and
                 {\v{S}}tajner, Sanja  and
                 Riedl, Martin  and
                 Biemann, Chris",
    booktitle = "Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017",
    year      = "2017",
    address   = "Varna, Bulgaria",
    url       = "https://aclanthology.org/R17-1104/",
}

Organisers (original shared task)

Sanja Štajner (Mannheim), Chris Biemann (Hamburg), Shervin Malmasi (Harvard Medical School), Gustavo Paetzold (Sheffield), Lucia Specia (Sheffield), Anaïs Tack (UCLouvain / KU Leuven), Seid Muhie Yimam (Hamburg), Marcos Zampieri (Wolverhampton).

Contact: sanja (at) informatik (dot) uni-mannheim (dot) de

Licensing

The CWI 2018 shared task data was distributed by the organisers for research use; no explicit redistribution license was attached to the original release. The source text is drawn from English Wikipedia and WikiNews (CC BY-SA 3.0 / CC BY 2.5), German / Spanish / French Wikipedia (CC BY-SA 3.0), and professionally-edited news articles. Downstream users should respect the licenses of the underlying source material and cite the shared task paper.

Provenance of this release

Source archives: CWI 2018 Training Set.zip and CWI 2018 Test Set.zip as distributed by the shared task organisers.
Conversion: TSV columns mapped 1:1 into parquet with explicit pyarrow schema; row order preserved within each file; no row filtering, deduplication, or text normalisation applied.
Mirror of the raw zips: alvations/stash · cwi-2018/.