Hugging Face's logo

Datasets:
microsoft
/
msr_text_compression

Tasks:
Summarization
Languages:
English
Size:
1K<n<10K
License:
Dataset card Files Community
2
msr_text_compression / README.md
albertvillanova's picture
albertvillanova HF staff
Reorder split names (#1)
c906a40
preview code
|
raw
history blame contribute delete
7.67 kB
metadata
annotations_creators:
  - crowdsourced
language_creators:
  - found
language:
  - en
license:
  - other
license_details: Microsoft Research Data License Agreement
multilinguality:
  - monolingual
size_categories:
  - 1K<n<10K
source_datasets:
  - extended|other-Open-American-National-Corpus-(OANC1)
task_categories:
  - summarization
task_ids: []
pretty_name: MsrTextCompression
dataset_info:
  features:
    - name: source_id
      dtype: string
    - name: domain
      dtype: string
    - name: source_text
      dtype: string
    - name: targets
      sequence:
        - name: compressed_text
          dtype: string
        - name: judge_id
          dtype: string
        - name: num_ratings
          dtype: int64
        - name: ratings
          sequence: int64
  splits:
    - name: train
      num_bytes: 5001312
      num_examples: 4936
    - name: validation
      num_bytes: 449691
      num_examples: 447
    - name: test
      num_bytes: 804536
      num_examples: 785
  download_size: 0
  dataset_size: 6255539

Dataset Card for [Dataset Name]

Table of Contents

Dataset Description

Dataset Summary

This dataset contains sentences and short paragraphs with corresponding shorter (compressed) versions. There are up to five compressions for each input text, together with quality judgements of their meaning preservation and grammaticality. The dataset is derived using source texts from the Open American National Corpus (ww.anc.org) and crowd-sourcing.

Supported Tasks and Leaderboards

Text Summarization

Languages

English

Dataset Structure

Data Instances

It contains approximately 6,000 source texts with multiple compressions (about 26,000 pairs of source and compressed texts), representing business letters, newswire, journals, and technical documents sampled from the Open American National Corpus (OANC1).

  • Each source text is accompanied by up to five crowd-sourced rewrites constrained to a preset compression ratio and annotated with quality judgments. Multiple rewrites permit study of the impact of operations on human compression quality and facilitate automatic evaluation.
  • This dataset is the first to provide compressions at the multi-sentence (two-sentence paragraph) level, which may present a stepping stone to whole document summarization.
  • Many of these two-sentence paragraphs are compressed both as paragraphs and separately sentence-bysentence, offering data that may yield insights into the impact of multi-sentence operations on human compression quality.
Description Source Target Average CPS Meaning Quality Grammar Quality
1-Sentence 3764 15523 4.12 2.78 2.81
2-Sentence 2405 10900 4.53 2.78 2.83

Note: Average CPS = Average Compressions per Source Text

Data Fields 

{'domain': 'Newswire',
 'source_id': '106',
 'source_text': '" Except for this small vocal minority, we have just not gotten a lot of groundswell against this from members, " says APA president Philip G. Zimbardo of Stanford University.',
 'targets': {'compressed_text': ['"Except for this small vocal minority, we have not gotten a lot of groundswell against this," says APA president Zimbardo.',
   '"Except for a vocal minority, we haven\'t gotten much groundswell from members, " says Philip G. Zimbardo of Stanford University.',
   'APA president of Stanford has stated that except for a vocal minority they have not gotten a lot of pushback from members.',
   'APA president Philip G. Zimbardo of Stanford says they have not had much opposition against this.'],
  'judge_id': ['2', '22', '10', '0'],
  'num_ratings': [3, 3, 3, 3],
  'ratings': [[6, 6, 6], [11, 6, 6], [6, 11, 6], [6, 11, 11]]}}
  • source_id: index of article per original dataset
  • source_text: uncompressed original text
  • domain: source of the article
  • targets:
    • compressed_text: compressed version of source_text
    • judge_id: anonymized ids of crowdworkers who proposed compression
    • num_ratings: number of ratings available for each proposed compression
    • ratings: see table below

Ratings system (excerpted from authors' README):

  • 6 = Most important meaning Flawless language (3 on meaning and 3 on grammar as per the paper's terminology)
  • 7 = Most important meaning Minor errors (3 on meaning and 2 on grammar)
  • 9 = Most important meaning Disfluent or incomprehensible (3 on meaning and 1 on grammar)
  • 11 = Much meaning Flawless language (2 on meaning and 3 on grammar)
  • 12 = Much meaning Minor errors (2 on meaning and 2 on grammar)
  • 14 = Much meaning Disfluent or incomprehensible (2 on meaning and 1 on grammar)
  • 21 = Little or none meaning Flawless language (1 on meaning and 3 on grammar)
  • 22 = Little or none meaning Minor errors (1 on meaning and 2 on grammar)
  • 24 = Little or none meaning Disfluent or incomprehensible (1 on meaning and 1 on grammar)

See README.txt from data archive for additional details.

Data Splits

There are 4,936 source texts in the training, 448 in the development, and 785 in the test set.

Dataset Creation

Annotations

Annotation process

Compressions were created using UHRS, an inhouse crowd-sourcing system similar to Amazon’s Mechanical Turk, in two annotation rounds, one for shortening and a second to rate compression quality:

  1. In the first round, five workers were tasked with abridging each source text by at least 25%, while remaining grammatical and fluent, and retaining the meaning of the original.
  2. In the second round, 3-5 judges (raters) were asked to evaluate the grammaticality of each compression on a scale from 1 (major errors, disfluent) through 3 (fluent), and again analogously for meaning preservation on a scale from 1 (orthogonal) through 3 (most important meaning-preserving).

Additional Information

Licensing Information

Microsoft Research Data License Agreement

Citation Information

@inproceedings{Toutanova2016ADA, title={A Dataset and Evaluation Metrics for Abstractive Compression of Sentences and Short Paragraphs}, author={Kristina Toutanova and Chris Brockett and Ke M. Tran and Saleema Amershi}, booktitle={EMNLP}, year={2016} }

Contributions

Thanks to @jeromeku for adding this dataset.