Density-aware Prototypical network(D-Proto)

Codes for our EMNLP 2023 paper: Density-Aware Prototypical Network for Few-Shot Relation Classification.

Environments

• python 3.6.2
• torch 1.7.1
• transformers 4.5.1
• scikit-learn 0.24.2

Datasets

FewRel

You can find the training and validation data here: FewRel.

Training

O-Proto model:

python train.py --dataset fewrel --N 5 --K 5 --Q 3 --batch_size 2 --model oproto --encoder bert --max_length 128 --trainNA 0.5 --optim adamw --hidden_size 768 --seed {}

Bert-Pair model:

python train.py --dataset fewrel --N 5 --K 5 --Q 3 --batch_size 2 --model pair --encoder bert --max_length 128 --trainNA 0.5 --optim adamw --hidden_size 768 --seed {} --pair

MNAV model:

python train.py --dataset fewrel --N 5 --K 5 --Q 3 --batch_size 2 --model mnav --vector_num 20 --encoder bert --max_length 128 --trainNA 0.5 --optim adamw --hidden_size 768 --seed {}

D-Proto model:

python train.py --dataset fewrel --N 5 --K 5 --Q 3 --batch_size 2 --model dproto --gamma 1e-5 --threshold 0.9 --encoder bert --max_length 128 --trainNA 0.5 --optim adamw --hidden_size 768 --seed {}

• N: N in N-way K-shot.
• K: K in N-way K-shot.
• Q: Sample Q query instances in the query set.
• trainNA: NOTA rate in training phase. In testing, test results under 0.15, 0.3, 0.5 NOTA rates are obtained respectively.
• seed: seed. 5/10/15/20/25.