README.md

RAP for PRGC

The code is basically borrowed from here, with slight modification.

Requirements

The main requirements are:

• python==3.7.9
• pytorch==1.6.0
• transformers==3.2.0
• tqdm

Split Data

Run the following commands to randomly split the datasets.

cd data
bash run_sample.bash

For each datasets, we get the corresponding low resource folders nyt_retrieved_ratio and webnlg_retrieved_ratio with 5 different random seeds, taking nyt_retrieved_ratio as example, we can get folders as follow:

data/nyt_retrieved_ratio
├── seed2
│   ├── 0.1
│   ├── 0.01
│   └── 0.05
├── seed3
│   └── ...
├── seed5
│   └── ...
├── seed7
│   └── ...
└── seed11
    └── ...

Usage

1. Get pre-trained BERT model for PyTorch

Download BERT-Base-Cased which contains pytroch_model.bin, vocab.txt and config.json. Put these under ./pretrain_models.

2. Train

Just run the script in ./script by sh train.sh.

Here's an example:

#!/bin/bash
export seed=5
export data="nyt_retrieved"
python ../train.py \
--ex_index=${data} \
--epoch_num=100 \
--device_id=3 \
--corpus_type=${data} \
--ensure_corres \
--ensure_rel

3. Evaluate

Just run the script in ./script by sh evaluate.sh.

Here's an example:

#!/bin/bash
export seed=5
export data="nyt_retrieved"
python ../evaluate.py \
--ex_index=${data} \
--device_id=1 \
--mode=test \
--corpus_type=${data} \
--ensure_corres \
--ensure_rel \
--corres_threshold=0.5 \
--rel_threshold=0.1 \
--restore_file=best