RAPS

This is the implementation of our paper RAPS: A Novel Few-Shot Relation Extraction Pipeline with Query-Information Guided Attention and Adaptive Prototype Fusion.

Requirements

• python 3.7.13
• PyTorch 1.9.1
• transformers 4.6.0
• numpy 1.21.6

Datasets

We experiment our model on two few-shot relation extraction datasets,

1. FewRel 1.0, FewRel 2.0 training & validation sets
2. FewRel 1.0, FewRel 2.0 test set

Please download data from the official links and put it under ./data/.

Pretrained Language Models

1. Download pretrained bert-base-uncased model from HuggingFace (https://huggingface.co/bert-base-uncased) and put the corresponding config files under ./bert-base-uncased/.
2. Download the CP pretrained model from https://github.com/thunlp/RE-Context-or-Names/tree/master/pretrain and put the checkpoint under ./CP/.

Training

FewRel 1.0 If you want to train a 5-way 1-shot model on FewRel 1.0, run

cd scripts
bash run_train_5_1.sh

In run_train_5_1.sh, you can specify BERT as backend model by export BACKEND="bert", or CP as backend model by export BACKEND="cp".

FewRel 2.0 If you want to train a 5-way 1-shot model on FewRel 2.0, run

cd scripts
bash run_train_5_1_da.sh

In run_train_5_1_da.sh, you can specify BERT as backend model by export BACKEND="bert", or CP as backend model by export BACKEND="cp".

Evaluation

FewRel 1.0 If you want to evaluate a 5-way 1-shot model on FewRel 1.0, run

cd scripts
bash run_eval_5_1.sh

In run_eval_5_1.sh, you can specify BERT as backend model by export BACKEND="bert", or CP as backend model by export BACKEND="cp".

FewRel 2.0 If you want to evaluate a 5-way 1-shot model on FewRel 2.0, run

cd scripts
bash run_eval_5_1_da.sh

In run_eval_5_1_da.sh, you can specify BERT as backend model by export BACKEND="bert", or CP as backend model by export BACKEND="cp".