This README file describes the data used in Woodsend and Lapata (ACL,
2010). There are two archives:

training.tgz and test.tgz

The training.tgz archive contains CNN articles and their corresponding
highlights. Each document and its highlight have the same prefix but
different suffixes. So for example the highlights for the article
tiger.conservation2008.doc.txt are tiger.conservation2008.hlights.txt.
Highlights and documents are in different directories (training/hlights and
training/docs, respectively).

The training files have been annotated as follows. In the beginning of
each line there is a number, ranging from 1, 2, and 3. The numbers
denote whether the document sentence corresponds to a highlight.
Specifically, label (1) means that the sentence must be in the
highlights, label (2) that the sentence could be in the highlights and
label (3) that the sentence is not in the highlights.

The structure of the test.tgz archive is analogous to
training.tgz. Again it has two directories (test/doc and
test/hlights), each containing files corresponding to the documents
and their highlights. The test files have no alignment annotations.

The files are all one sentence per line, and tokenized. If you use the
data please cite:


@InProceedings{woodsend-lapata:2010:ACL,
  author    = {Woodsend, Kristian  and  Lapata, Mirella},
  title     = {Automatic Generation of Story Highlights},
  booktitle = {Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics},
  month     = {July},
  year      = {2010},
  address   = {Uppsala, Sweden},
  publisher = {Association for Computational Linguistics},
  pages     = {565--574},
  url       = {http://www.aclweb.org/anthology/P10-1058}
}