Learning to automatically select pull quotes (wikipedia).

This code accompanies the accepted COLING-2020 paper Catching Attention with Automatic Pull Quote Selection.

This project is written in Python3.6.9

The following non-default libraries are used:

• numpy 1.18.2
• sklearn 0.22.2.post1
• seaborn 0.9.0
• matplotlib 3.1.2
• scipy 1.4.1
• keras 2.3.0
• tensorflow 1.14.0
• sumy 0.8.1
• nltk 3.4.5
• textstat 0.6.0
• textblob 0.15.3
• sentence_transformers 0.2.5

To reproduce our dataset:

1. navigate to the datasets/url_lists/ directory and unzip url_lists.zip so that the 4 files are in datasets/url_lists/
2. nagivate to datasets/ and run python3.6 construct_dataset.py source my_save_dir/.
   • source can be one of intercept, ottawa-citizen, cosmo, national-post, or all
   • the samples for a given source will be stored in my_save_dir/source/
   • ⚠️ Update settings.py so that base_pq_directory points to my_save_dir/.
   • ⚠️ This will take a long time.
3. navigate to the root repo folder and run python3.6 calculate_data_stats.py to calculate dataset statistics to compare with our paper.

To reproduce our experimental results, run bash run_experiments.sh (output will be stored in /results).

ℹ️ To first make sure that things work, run bash run_experiments.sh --quick. It should take just a few minutes.

To reproduce the handcrafted feature value distribution figures, run python3.6 view_feature_dists.py

To analyze test articles with a all models, run bash generate_model_samples.sh. The --quick argument can similarly be used to make sure things are working.