Datasets:
rcds
/

ArXiv:
Tags:
legal
License:

The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

Dataset Card for MultiLegalNeg

Dataset Summary

This dataset consists of German, French, and Italian court documents annotated for negation cues and negation scopes. It also includes a reformated version of ConanDoyle-neg ( Morante and Blanco. 2012), SFU Review (Konstantinova et al. 2012), BioScope (Szarvas et al. 2008) and Dalloux (Dalloux et al. 2020).

Languages

Language Subset Number of sentences Negated sentences
French fr 1059 382
Italian it 1001 418
German(Germany) de(DE) 1068 1098
German (Switzerland) de(CH) 206 208
English SFU Review 17672 3528
English BioScope 14700 2095
English ConanDoyle-neg 5714 5714
French Dalloux 11032 1817

Dataset Structure

Data Fields

  • text (string): full sentence
  • spans (list): list of annotated cues and scopes
    • start (int): offset of the beginning of the annotation
    • end (int): offset of the end of the annotation
    • token_start(int): id of the first token in the annotation
    • token_end(int): id of the last token in the annotation
    • label (string): CUE or SCOPE
  • tokens (list): list of tokens in the sentence
    • text (string): token text
    • start (int): offset of the first character
    • end (int): offset of the last character
    • id (int): token id
    • ws (boolean): indicates if the token is followed by a white space

Data Splits

For each subset a train (70%), test (20%), and validation (10%) split is available.

How to use this dataset

To load all data use 'all_all', or specify which dataset to load as the second argument. The available configurations are 'de', 'fr', 'it', 'swiss', 'fr_dalloux', 'fr_all', 'en_bioscope', 'en_sherlock', 'en_sfu', 'en_all', 'all_all'

from datasets import load_dataset

dataset = load_dataset("rcds/MultiLegalNeg", "all_all")

dataset
DatasetDict({
    train: Dataset({
        features: ['text', 'spans', 'tokens'],
        num_rows: 26440
    })
    test: Dataset({
        features: ['text', 'spans', 'tokens'],
        num_rows: 7593
    })
    validation: Dataset({
        features: ['text', 'spans', 'tokens'],
        num_rows: 4053
    })
})

Source Data

Annotations

The data is annotated for negation cues and their scopes. Annotation guidelines are available here

Annotation process

Each language was annotated by one native speaking annotator and follows strict annotation guidelines

Citation Information

Please cite the following preprint:

@misc{christen2023resolving,
      title={Resolving Legalese: A Multilingual Exploration of Negation Scope Resolution in Legal Documents}, 
      author={Ramona Christen and Anastassia Shaitarova and Matthias Stürmer and Joel Niklaus},
      year={2023},
      eprint={2309.08695},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
Downloads last month
349

Models trained or fine-tuned on rcds/MultiLegalNeg