This article is published at the Scientific Scholarly Processing (SDP) 2021 workshop.

1. Download the original papers and slides from here and put it in the slide_generator_data directory
2. data dir structure:

all_test.txt: list of test files
all_train.txt: list of train files
all_val.txt:  list of validation files
train train  val: directories containing the data
*.sents.txt: sentences from the paper extracted by GROBID. Each line represents one sentence.
*.windowed_summarunner_scores.txt: Each line represents the label for the corresponding sentence

3. You need to download and set the glove word vectors

To train the model run:

$CUDA_VISIBLE_DEVICES=0 python3 batch_train.py

If the data_files/vocab.txt and lookup.pkl do not exist it will create the vocab file and lookup table.

Note: To change the training mode from summarunner to windowed_summarunner: change the batch_data_utils.py -> get_article_labels

1. Change the cpt (checkpoint) variable in batch_test.py.
2. Set the name of the output file e.g. predicted_windowed_summarunner.
3. Run $CUDA_VISIBLE_DEVICES=2 python3 batch_test.py

run python3 evaluateROUGE155.py