README.md

RAP for RelationPrompt

The original code of RelationPrompt is here. And for our experiments, we just simply implement the Extractor part of RelationPrompt.

Requirements

General

• Python (verified on 3.7)

Python Packages

• see requirements.txt

conda create -n relationprompt python=3.7
conda activate relationprompt
pip install -r requirements.txt

Split Data

Run the following commands to randomly split the datasets.

cd outputs/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. Train

Run the following command for training, seed here indicate the different random seed used when splitting datasets.

python train.py --seed 2 --ratio 0.01 --data_name webnlg_retrieved -lr 5e-4 --eval_steps 20

2. Evaluate

Run the following command for evaluation (The evaluation is also conducted after the training) :

python predict.py --seed 2 --ratio 0.01 --data_name webnlg_retrieved -lr 5e-4