SciBERTSUM - A deep learning model with LongScientificFormer for scientific paper summarization

For encoding a text longer than 512 tokens, for example 800. Set max_pos to 800 during both preprocessing and training.

Some codes are borrowed from ONMT(https://github.com/OpenNMT/OpenNMT-py)

Data Preparation

Option 1: download the processed data

Put all files into raw_data directory. The structure of raw_data folder is

root
├── logs
├── models
├── raw_data
│   ├── 0
│   ├── 1
│   ├── 2
│   ├── 3
│   ├── 4
│   ├── ...
├── results
├── src
├── stanford-corenlp-4.2.2
├── README.md
├── requirements.txt

Step 2. Download Stanford CoreNLP

We will need Stanford CoreNLP to tokenize the data. Download it here and unzip it. Then add the following command to your bash_profile:

export CLASSPATH=/path/to/stanford-corenlp-4.2.2/stanford-corenlp-4.2.2.jar

replacing /path/to/ with the path to where you saved the stanford-corenlp-4.2.2 directory.

step 3. extracting sections from GROBID XML files

python preprocess.py -mode extract_pdf_sections -log_file ../logs/extract_section.log

step 4. extracting text from TIKA XML files

python preprocess.py -mode get_text_clean_tika -log_file ../logs/extract_tika_text.log

step 5. Tokenize texts from papers and slides using stanfordCoreNLP

python preprocess.py -mode tokenize  -save_path ../temp -log_file ../logs/tokenize_by_corenlp.log

Step 6. Extract source, section, and target from tokenized files

python preprocess.py -mode clean_paper_jsons -save_path ../json_data/  -n_cpus 10 -log_file ../logs/build_json.log

Step 7. Generate BERT `.pt` files from source, sections and targets

python preprocess.py -mode format_to_bert -raw_path ../json_data/ -save_path ../bert_data  -lower -n_cpus 40 -log_file ../logs/build_bert_files.log

Model Training

First run: For the first time, you should use single-GPU, so the code can download the BERT model. Use -visible_gpus -1, after downloading, you could kill the process and rerun the code with multi-GPUs.

Train

python train.py  -ext_dropout 0.1 -lr 2e-3  -visible_gpus 1,2,3 -report_every 200 -save_checkpoint_steps 1000 -batch_size 1 -train_steps 100000 -accum_count 2  -log_file ../logs/ext_bert -use_interval true -warmup_steps 10000

To continue training from a checkpoint

python train.py  -ext_dropout 0.1 -lr 2e-3  -train_from ../models/model_step_99000.pt -visible_gpus 1,2,3 -report_every 200 -save_checkpoint_steps 1000 -batch_size 1 -train_steps 100000 -accum_count 2  -log_file ../logs/ext_bert -use_interval true -warmup_steps 10000

Test

python train.py -mode test  -test_batch_size 1 -bert_data_path ../bert_data -log_file ../logs/ext_bert_test -test_from ../models/model_step_99000.pt -model_path ../models -sep_optim true -use_interval true -visible_gpus 1,2,3 -alpha 0.95 -result_path ../results/ext

Deployment

The model is delpoyed in a Streamlit app.

First, at the root folder, install Grobid for pdf extraction. You can run the following command in ther terminal to install Grobid:

git clone https://github.com/kermitt2/grobid.git
cd grobid
./gradlew clean install

Then, you run the GROBID Service API with the command:

./gradlew run

Next, to run the Streamlit application, please open a new terminal and run the following command from root directory:

cd src
streamlit run app.py

trminhnam/cs420-SciBERTSUm