README.md

GM_GEN

The code of the paper "Graph-based Model Generation for Few-Shot Relation Extraction". This paper has been accepted to EMNLP2022. You can find the main results (username is GM_GEN) in the paper on FewRel 1.0 competition on CodaLab competition websit: FewRel 1.0 Competition

We will release the paper link after camera ready.

Environments

• python 3.7.10
• PyTorch 1.7.1
• transformers 4.7.0

Datasets and Models

You can find the training and validation data here: FewRel 1.0 data. For the test data, you can easily download from FewRel 1.0 competition website: https://competitions.codalab.org/competitions/27980

Code

Put all data in the data folder, CP pretrained model in the CP_model folder (you can download CP model from https://github.com/thunlp/RE-Context-or-Names/tree/master/pretrain or Google Drive)

Train

You can train a 5-way-1-shot model by:

python train_demo.py --trainN 5 --N 5 --K 1 --Q 1 --pretrain_ckpt ../Bert4SemiRE/bert/BERT-BASE-UNCASED --cat_entity_rep --model gm_gen --dot

or a 5-way-5-shot model by:

python train_demo.py --trainN 5 --N 5 --K 5 --Q 1 --pretrain_ckpt ../Bert4SemiRE/bert/BERT-BASE-UNCASED --cat_entity_rep --model gm_gen

Evaluation

The learning rate of the generated model will affect the model performance, so we recommend that you select an appropriate learning rate through the valiadation set before setting the learning rate by:

python train_demo.py --trainN 5 --N 5 --K 5 --Q 1 --pretrain_ckpt ../Bert4SemiRE/bert/BERT-BASE-UNCASED --cat_entity_rep --model gm_gen --load_ckpt ./checkpoint/gm_gen-bert-train_wiki-val_wiki-5-1-dot-catentity.pth.tar --param_lr 0.0001 --test_params --only_test

Then, an online test can be performed by:

python train_demo.py --trainN 5 --N 5 --K 5 --Q 1 --pretrain_ckpt ../Bert4SemiRE/bert/BERT-BASE-UNCASED --cat_entity_rep --model gm_gen --test_online --load_ckpt ./checkpoint/gm_gen-bert-train_wiki-val_wiki-5-1-dot-catentity.pth.tar --test_lr 0.0001 --only_test

Some explanations of the parameters in the script:

--load_ckpt
	the path of the trained model
--pretrain_ckpt
	the path of the pretrained model
--param_lr
	the learning rate of the generated models on validation set
--test_lr
	the learning rate of the generated models on test set

Results

BERT on FewRel 1.0

	5-way-1-shot	5-way-5-shot	10-way-1-shot	10-way-5-shot
Val	92.65	95.62	86.81	91.27
Test	94.89	96.96	91.23	94.30

CP on FewRel 1.0

	5-way-1-shot	5-way-5-shot	10-way-1-shot	10-way-5-shot
Val	96.97	98.32	93.97	96.58
Test	97.03	98.34	94.99	96.91