/SEGMENT

Improving Abstracitve Summarization with Segment-augmented and Position-awareness (ACLING2021)

Primary LanguagePythonMIT LicenseMIT

IMPROVING ABSTRACTIVE SUMMARIZATION WITH SEGMENT-AUGMENTED AND POSITION-AWARENESS (ACLing2021)

Aurhors: Phuc Minh Nguyen and Thao Nhi Tran

Overview Architecture

alt text

Usage

pip install -r requirements.txt

Data

# Download preprocessed data at ./squad/, ./cnndm/ and ./glove/ respectively
wget https://www.dropbox.com/s/dl/0gtz5ckh3ie55oq/emnlp2019focus_redistribute.zip

# Generate train_df.pkl, val_df.pkl, test_df.pkl and vocab.pkl at ./cnndm_out/
python CNNDM_data_loader.py

Details of dataset source are at Dataset_details.md

Train

  1. Abstract Summrization
python train.py --task=SM --model=PG --load_glove=False --data=cnndm \
    --batch_size=16 --eval_batch_size=16 \
    --use_focus=True --n_mixture=1 --decoding=beam  \
    --load_ckpt=-1

Evaluation

  1. Abstract Summrization
python evaluate.py --task=SM --model=PG --load_glove=False --data=cnndm \
    --batch_size=16 --eval_batch_size=64 \
    --use_focus=True --n_mixture=1 --decoding=beam \
    --load_ckpt=3 --eval_only

Result

  • Results on the CNN/DM dataset
Model R1 R2 RL
PG 39.53 17.28 36.38
Bottom-Up 41.22 18.68 38.34
SELECTOR 41.72 18.74 38.79
SEGMENT 42.10 19.24 38.80
SEGMENT + TCN 42.20 19.25 38.86