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.