asset / README.md
albertvillanova's picture
Add paper arXiv URL to dataset card
ca3b122 verified
|
raw
history blame
12 kB
metadata
annotations_creators:
  - machine-generated
language_creators:
  - found
language:
  - en
license:
  - cc-by-sa-4.0
multilinguality:
  - monolingual
size_categories:
  - 1K<n<10K
source_datasets:
  - original
  - extended|other-turkcorpus
task_categories:
  - text-classification
  - text2text-generation
task_ids:
  - text-simplification
paperswithcode_id: asset
pretty_name: ASSET
config_names:
  - ratings
  - simplification
tags:
  - simplification-evaluation
dataset_info:
  - config_name: ratings
    features:
      - name: original
        dtype: string
      - name: simplification
        dtype: string
      - name: original_sentence_id
        dtype: int32
      - name: aspect
        dtype:
          class_label:
            names:
              '0': meaning
              '1': fluency
              '2': simplicity
      - name: worker_id
        dtype: int32
      - name: rating
        dtype: int32
    splits:
      - name: full
        num_bytes: 1036845
        num_examples: 4500
    download_size: 44642
    dataset_size: 1036845
  - config_name: simplification
    features:
      - name: original
        dtype: string
      - name: simplifications
        sequence: string
    splits:
      - name: validation
        num_bytes: 2303484
        num_examples: 2000
      - name: test
        num_bytes: 411019
        num_examples: 359
    download_size: 1055163
    dataset_size: 2714503
configs:
  - config_name: ratings
    data_files:
      - split: full
        path: ratings/full-*
  - config_name: simplification
    data_files:
      - split: validation
        path: simplification/validation-*
      - split: test
        path: simplification/test-*
    default: true

Dataset Card for ASSET

Table of Contents

Dataset Description

Dataset Summary

ASSET (Alva-Manchego et al., 2020) is multi-reference dataset for the evaluation of sentence simplification in English. The dataset uses the same 2,359 sentences from TurkCorpus (Xu et al., 2016) and each sentence is associated with 10 crowdsourced simplifications. Unlike previous simplification datasets, which contain a single transformation (e.g., lexical paraphrasing in TurkCorpus or sentence splitting in HSplit), the simplifications in ASSET encompass a variety of rewriting transformations.

Supported Tasks and Leaderboards

The dataset supports the evaluation of text-simplification systems. Success in this tasks is typically measured using the SARI and FKBLEU metrics described in the paper Optimizing Statistical Machine Translation for Text Simplification.

Languages

The text in this dataset is in English (en).

Dataset Structure

Data Instances

  • simplification configuration: an instance consists in an original sentence and 10 possible reference simplifications.
  • ratings configuration: a data instance consists in an original sentence, a simplification obtained by an automated system, and a judgment of quality along one of three axes by a crowd worker.

Data Fields

  • original: an original sentence from the source datasets
  • simplifications: in the simplification config, a set of reference simplifications produced by crowd workers.
  • simplification: in the ratings config, a simplification of the original obtained by an automated system
  • aspect: in the ratings config, the aspect on which the simplification is evaluated, one of meaning, fluency, simplicity
  • rating: a quality rating between 0 and 100

Data Splits

ASSET does not contain a training set; many models use WikiLarge (Zhang and Lapata, 2017) for training.

Each input sentence has 10 associated reference simplified sentences. The statistics of ASSET are given below.

Dev Test Total
Input Sentences 2000 359 2359
Reference Simplifications 20000 3590 23590

The test and validation sets are the same as those of TurkCorpus. The split was random.

There are 19.04 tokens per reference on average (lower than 21.29 and 25.49 for TurkCorpus and HSplit, respectively). Most (17,245) of the referece sentences do not involve sentence splitting.

Dataset Creation

Curation Rationale

ASSET was created in order to improve the evaluation of sentence simplification. It uses the same input sentences as the TurkCorpus dataset from (Xu et al., 2016). The 2,359 input sentences of TurkCorpus are a sample of "standard" (not simple) sentences from the Parallel Wikipedia Simplification (PWKP) dataset (Zhu et al., 2010), which come from the August 22, 2009 version of Wikipedia. The sentences of TurkCorpus were chosen to be of similar length (Xu et al., 2016). No further information is provided on the sampling strategy.

The TurkCorpus dataset was developed in order to overcome some of the problems with sentence pairs from Standard and Simple Wikipedia: a large fraction of sentences were misaligned, or not actually simpler (Xu et al., 2016). However, TurkCorpus mainly focused on lexical paraphrasing, and so cannot be used to evaluate simplifications involving compression (deletion) or sentence splitting. HSplit (Sulem et al., 2018), on the other hand, can only be used to evaluate sentence splitting. The reference sentences in ASSET include a wider variety of sentence rewriting strategies, combining splitting, compression and paraphrasing. Annotators were given examples of each kind of transformation individually, as well as all three transformations used at once, but were allowed to decide which transformations to use for any given sentence.

An example illustrating the differences between TurkCorpus, HSplit and ASSET is given below:

Original: He settled in London, devoting himself chiefly to practical teaching.

TurkCorpus: He rooted in London, devoting himself mainly to practical teaching.

HSplit: He settled in London. He devoted himself chiefly to practical teaching.

ASSET: He lived in London. He was a teacher.

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

The input sentences are from English Wikipedia (August 22, 2009 version). No demographic information is available for the writers of these sentences. However, most Wikipedia editors are male (Lam, 2011; Graells-Garrido, 2015), which has an impact on the topics covered (see also the Wikipedia page on Wikipedia gender bias). In addition, Wikipedia editors are mostly white, young, and from the Northern Hemisphere (Wikipedia: Systemic bias).

Reference sentences were written by 42 workers on Amazon Mechanical Turk (AMT). The requirements for being an annotator were:

  • Passing a Qualification Test (appropriately simplifying sentences). Out of 100 workers, 42 passed the test.
  • Being a resident of the United States, United Kingdom or Canada.
  • Having a HIT approval rate over 95%, and over 1000 HITs approved.

No other demographic or compensation information is provided in the ASSET paper.

Annotations

Annotation process

The instructions given to the annotators are available here.

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

The dataset may contain some social biases, as the input sentences are based on Wikipedia. Studies have shown that the English Wikipedia contains both gender biases (Schmahl et al., 2020) and racial biases (Adams et al., 2019).

Adams, Julia, Hannah Brückner, and Cambria Naslund. "Who Counts as a Notable Sociologist on Wikipedia? Gender, Race, and the “Professor Test”." Socius 5 (2019): 2378023118823946. Schmahl, Katja Geertruida, et al. "Is Wikipedia succeeding in reducing gender bias? Assessing changes in gender bias in Wikipedia using word embeddings." Proceedings of the Fourth Workshop on Natural Language Processing and Computational Social Science. 2020.

Other Known Limitations

Dataset provided for research purposes only. Please check dataset license for additional information.

Additional Information

Dataset Curators

ASSET was developed by researchers at the University of Sheffield, Inria, Facebook AI Research, and Imperial College London. The work was partly supported by Benoît Sagot's chair in the PRAIRIE institute, funded by the French National Research Agency (ANR) as part of the "Investissements d’avenir" program (reference ANR-19-P3IA-0001).

Licensing Information

Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Citation Information

@inproceedings{alva-manchego-etal-2020-asset,
    title = "{ASSET}: {A} Dataset for Tuning and Evaluation of Sentence Simplification Models with Multiple Rewriting Transformations",
    author = "Alva-Manchego, Fernando  and
      Martin, Louis  and
      Bordes, Antoine  and
      Scarton, Carolina  and
      Sagot, Beno{\^\i}t  and
      Specia, Lucia",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.424",
    pages = "4668--4679",
}

This dataset card uses material written by Juan Diego Rodriguez.

Contributions

Thanks to @yjernite for adding this dataset.