Triplet loss text classification with BERT

Data

unzip data.zip

Usage

Triplet Loss Experiments (no hard negative mining)

For experiments, you should comment out the config files for the experiments you want to run:

1. No augmentation

python multi_seed_triplet_ap_vanilla.py

2. Standard EDA augmentation

python multi_seed_triplet_ap_eda_alpha.py

3. Curriculum two-stage augmentation

python multi_seed_triplet_ap_eda_twostep.py

4. Curriculum gradual augmentation

python multi_seed_triplet_ap_eda_gradual.py

Triplet Loss Experiments (with hard negative mining)

1. No augmentation

python multi_seed_triplet_ap_vanilla_mine.py

2. Standard EDA augmentation

python multi_seed_triplet_ap_eda_mine_alpha.py

3. Curriculum two-stage augmentation

python multi_seed_triplet_ap_eda_mine_twostep.py

4. Curriculum gradual augmentation

python multi_seed_triplet_ap_eda_mine_gradual.py

Other augmentation methods in standard vs two-stage curriculum

Token Substitution

python triplet_ap_sr_alpha.py
python triplet_ap_sr_twostep.py

Word Dropout

python triplet_ap_rd_alpha.py
python triplet_ap_rd_twostep.py

SwitchOut

python triplet_ap_so_alpha.py
python triplet_ap_so_twostep.py

Back-translation

python triplet_ap_bt_alpha.py
python triplet_ap_bt_twostep.py

Very Simple Baselines

Run LR/MLP baselines for classification of BERT-avgpool encodings:

python multi_seed_mlp.py

Run k-NN baseline for classification of BERT-avgpool encodings (not used in paper):

python knn_ap_vanilla.py