Controlled Text Reduction

This dataset contains Controlled Text Reduction triplets - document-summary pairs, and the spans in the document that cover the summary. The task input is consists of a document with pre-selected spans in it ("highlights"). The output is a text covering all and only the highlighted content.

The script downloads the data from the original GitHub repository.

Format

The dataset contains the following important features:

doc_text - the input text.
summary_text - the output text.
highlight_spans - the spans in the input text (the doc_text) that lead to the output text (the summary_text).

{'doc_text': 'The motion picture industry\'s most coveted award...with 32.',
  'summary_text': 'The Oscar, created 60 years ago by MGM...awarded person (32).',
  'highlight_spans':'[[0, 48], [50, 55], [57, 81], [184, 247], ..., [953, 975], [1033, 1081]]'}

where for each document-summary pair, we save the spans in the input document that lead to the summary.

Notice that the dataset consists of two subsets:

DUC-2001-2002 - which is further divided into 3 splits (train, validation and test).
CNN-DM - which has a single split.

Citation

If you find the Controlled Text Reduction dataset useful in your research, please cite the following paper:

@misc{https://doi.org/10.48550/arxiv.2210.13449,
  doi = {10.48550/ARXIV.2210.13449},
  url = {https://arxiv.org/abs/2210.13449},
  author = {Slobodkin, Aviv and Roit, Paul and Hirsch, Eran and Ernst, Ori and Dagan, Ido},
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Controlled Text Reduction},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Zero v1.0 Universal}
}