Skip to content


Switch branches/tags


Failed to load latest commit information.
Latest commit message
Commit time


Code for the paper: BillSum: A Corpus for Automatic Summarization of US Legislation (Kornilova and Eidelman, 2019)

This paper was be presented at EMNLP 2019 Workshop on New Frontiers in Summarization. Link to slides from workshop

Accessing the Dataset: This dataset was updated on 12/3/2019, if you accessed the dataset prior to this date, please redownload it.

Link to Google Drive

TensorFlow Datasets - does not contain the "clean" versions of the texts

If you do something cool with the data, share on our Kaggle Page!

Information on how the dataset was collected is under

Data Structure The data is stored in a jsonlines format, with one bill per line.

  • text: bill text

  • clean_text: a preprocessed version of the text that was used to train the models in the paper

  • summary: (human-written) bill summary

  • title: bill title (can be used for generating a summary)

  • bill_id: An identified for the bill - in US data it is SESSION_BILL-ID, for CA BILL-ID


  1. Install python dependencies (If using conda, use env.lst. If using pip, use requirements.txt)
  2. Set the env BILLSUM_PREFIX to the base directory for all the data. (Download from link above)
  3. Set PYTHONPATH=. to run code from this directory.
  4. Install packages from environment.lst (we used conda, but you should be able to use pip


The results for the intermediate steps (explained below) can be found here

For all the experiments described in the paper, the texts were first cleaned using the script billsum/data_prep/ Results will be saved into the BILLSUM_PREFIX/clean_final directory.

Sumy baselines

  1. Clone sumy and checkout the branch ak_fork (This is a minor modification on the original sumy library that allows it to work with my sentence selection logic).
  2. In that directory run pip install -e .
  3. From this directory, run bill_sum/

Supervised Experiments

Preparing the data

  1. Run billsum/data_prep/ to clean up the whitespace formatting in the dataset. Outputs new jsonlines files with 'clean_text' field + original fields to BILLSUM_PREFIX/clean_data

  2. Run billsum/data_prep/ to create labeled dataset.

This script takes each document, splits it into sentences, processes them with Spacy to get useful syntactic features and calculates the Rouge Score relative to the summary.

Outputs for each dataset part will be a pickle file with a dict of (bill_id, sentence data) pairs. (Stored under PREFIX/sent_data/) directory

Bill_id --> [
	('The monthly limitation for each coverage month during the taxable year is an amount equal to the lesser of 50 percent of the amount paid for qualified health insurance for such month, or an amount equal to 112 of in the case of self-only coverage, $1,320, and in the case of family coverage, $3,480. ',
	  [('The ', 186, 'the', '', 'O', 'DET', 'det', 188),
	   ('monthly ', 187, 'monthly', 'DATE', 'B', 'ADJ', 'amod', 188),
	   ('limitation ', 188, 'limitation', '', 'O', 'NOUN', 'nsubj', 197),
	  {'rouge-1': {'f': 0.2545454500809918,
	    'p': 0.3783783783783784,
	    'r': 0.1917808219178082},
	   'rouge-2': {'f': 0.09459459021183367, 'p': 0.14583333333333334, 'r': 0.07},
	   'rouge-l': {'f': 0.16757568176139123,
	    'p': 0.2972972972972973,
	    'r': 0.1506849315068493}}),

Running Bert Models

  1. Clone Replace the file with billsum/bert_helpers/ (adds custom code to read data in and out of files). Install dependencies as described in this repo.

  2. Create train.tsv / test.tsv files with billsum/bert_helpers/ These will be stored under PREFIX/bert_data (set $BERT_DATA_DIR to point here)

  3. Download the Bert-Large, Uncased model.

  4. Set $BERT_BASE_DIR environment variable to point to directory where you downloaded the model

  5. Pretrain the Bert Model (run from the cloned bert repo)

python \
  --input_file=$BERT_DATA_DIR/all_texts_us_train.txt \
  --output_file=$BERT_DATA_DIR/all_texts_us_train.tfrecord \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --do_lower_case=True \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --masked_lm_prob=0.15 \
  --random_seed=12345 \

Set $BERT_MODEL_DIR to the directory where you want to store your pretrained model.

python \
  --output_dir=$BERT_MODEL_DIR \
  --do_train=True \
  --do_eval=True \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --train_batch_size=32 \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --num_train_steps=20000 \
  --num_warmup_steps=10 \

This will take a while to run.

  1. To train the classifier model run (from bert repo):

Change BERT_CLASSIFIER_DIR to the directory where you want to store the classifier - should be different from pretraining directory. This script will create a model in the BERT_CLASSIFIER_DIR and store the sentence predictions in BERT_CLASSIFIER_DIR/ dir.

For clarity:

  • BERT_BASE_DIR: directory of the original downloaded model (same as for step 3)
  • BERT_MODEL_DIR: directory where the output of the pretraining was stored
  • BERT_DATA_DIR: directory with all train/test examples
  • BERT_CLASSIFIER_DIR: directory where new model should

After this procedure is run, two files will be generated in the BERT_CLASSIFIER_DIR: test_results.tsv / ca_test_results.tsv -- this contain sentence level predictions for each test sentence. Rename the test_results.tsv file to us_test_results.tsv. Then copy both of them over to the bert_data folder.

  1. Evaluate results using bill_sum/bert_helpers/ Change the prefix variable to point to BERT_CLASSIFIER_DIR from above.

Results will be stored under BILLSUM_PREFIX/score_data/

Running feature classifier + ensemble

Run bill_sum/ Results will be stored under BILLSUM_PREFIX/score_data/

To get computations for the ensemble method run billsum/

Final Result aggregation

The PrintFinalScores.ipynb will compute the summary statistics for each method + generate the Oracle scores.


No releases published


No packages published