Skip to content

Latest commit

 

History

History
95 lines (62 loc) · 3.9 KB

README.rst

File metadata and controls

95 lines (62 loc) · 3.9 KB
Raw

Helper Codes

CNN-Daily Mail

First Option: Download the processed data

@JafferWilson provided the processed data, which you can download from here.

Second Option: Process dataset using @abisee code

Check out the pre-processing in here.

Third Option: Process dataset using our helper code

Download Raw Data

Through our helper codes, we first download all the articles in both CNN and Daily Mail dataset along with the title of each article and their highlights. All you need to do is to download the "*.story" files for each dataset from here. As per @abisee, these files contain 114 examples for which the article text is missing, we store list of these files in the filter_data.txt file and during our pre-processing we remove them from the dataset.

Download article, headline, and highlights from raw data

Then, run following to start downloading the articles and headline of each article. Please note that, this dataset doesn't have headline information, therefore we have to download the headline from each url by re-downloading the whole article.

python cnn_dm_downloader.py ~/data/cnn_dm/[cnn,dailymail]/ ~/data/cnn_dm/[cnn,dailymail]/processed/ article

And run this to download the highlights:

python cnn_dm_downloader.py ~/data/cnn_dm/[cnn,dailymail]/ ~/data/cnn_dm/[cnn,dailymail]/processed/ highlight

Pre-process the data

The main pre-processing that we do on this dataset is the following:

  1. Use Spacy to tokenize and process the articles and its highlights.
  2. We first sentence tokenize each article and then use word tokenization and re-join everything to get the full sentence.
  3. Each sentences is terminated by an "EOS" token rather than "." and each article (highlight, headline) has the following format:
"<d> <s> first sentence ... </s> <s> second sentence ... </s> </d>".
  1. Replacing all Named-Entities with "_" delimited version of them. For instance, given "Virgini Tech", we get "virginia_tech"
  2. Collect POS tag and Named-Entities of each token in the text and store them in a separate file.

We can run our pre-processing using the following command:

python cnn_dm_data_maker.py ~/data/cnn_dm/[cnn,dailymail]/ ~/data/cnn_dm/[cnn,dailymail]/[article,title,highlight] [article,title/highlight]

Create Train, Dev, Test

Finally, we use the following to create our train, dev, test datasets:

python cnn_dm_data_merger.py ~/data/cnn_dm/ cnn_dm.txt

Newsroom Dataset

Download Raw Data

You can follow this link to download and extract this dataset. After downloading you should have "[train,test,dev].data" files.

Run Pre-processing

This is completely similar to the CNN/Daily Mail pre-processing. The input directory must contain "[train,test,dev].data" files.

python summaries_data_maker.py ~/data/newsroom/