data

Data

We used the Blogs Authorship Corpus, IMDb62, Amazon 5-core Reviews datasets to construct our benchmark. The description of the how the data was selected, processed, and split into Train, Dev, and Test splits is described in our paper. The data we used is available in this release.

Benchmark Construction:

Reproducing our benchmark can be obtained by using the create_data.sh script, which encapsulates the following steps:

1. Downloading each dataset separately and applying a preprocessing step if needed.
2. Annotating the data examples with fine-grained linguistic features by using the utils/annotate_data.py script.
3. Obtaining the RST relations by using rstfinder. This can be done by 1) writing every data example to a separate file and 2) invoking the rstfinder/parse_data.sh script.
4. Spliting and descritizing the data to create Train, Dev, and Test splits using the utils/split_and_discretize.py script.