- F1 search
- Each fold model uses the same threshold
- Focalloss
- GHM-C loss
- DiceLoss
- FGM attack
- PGD attack
- K-fold voting (5-7-9 fold)
- Fusion of multiple models
- Pseudo label [reference]
- Optimizers
- Multiple loss: accelerate convergence in the early stage and improve accuracy in the later stage.
- Reverse query-reply
- Bert
- Double bert + LSTM (Double bert model query and reply respectively, then a BiLSTM model the relation between them)
- LCF [link]
- Semantic Role Labeling
- Auxiliary task: Identify query-reply or reply-query
- Bert+GCN
- ERNIE 1.0
- RoBERTa
- BERT-wwm-ext
- Pre-training with test dataset
- Pre-training with QA data in the real estate field
- Pre-training with QA data in other domain [Link1] [Link2]
- Truncation
- Clean
- KFold
- StratifiedKFold
- GroupKFold
- Exchange query-reply pair order
- Delete duplicate queries and emoji
- Cluster analysis for queries & Multi-task learning
- Back translation
- Split reply that exceeds the maximum length
- LaserTagger [link]
- Random Mask
- Prediction results of dev set
- Prediction results of test set
-
0.77308091 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 5-fold voting | BCE loss
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE
-
0.77179287015 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 5-fold voting | GHMC loss
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE --criterion ghmc
-
0.77587103484 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 5-fold voting | BCE loss | datareverse
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE --datareverse
-
0.77975341 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 5-fold voting | BCE loss | FGM
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 4 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE --attack_type fgm --scheduler
-
0.77839228296 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 5-fold voting | BCE loss | PGD
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 4 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE --attack_type pgd --scheduler
-
0.78033390298 | bert_spc | ERNIE | Search_f1 | StratifiedKFold 7-fold voting | BCE loss | FGM
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE --attack_type fgm --scheduler --cross_val_fold 7
-
0.78855493 | bert_spc | ERNIE-TAPT | Search_f1 | StratifiedKFold 7-fold voting | BCE loss | FGM
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-TAPT --attack_type fgm --scheduler --cross_val_fold 7
-
0.78575129534 | bert_spc | ERNIE-TAPT | Search_f1 | StratifiedKFold 7-fold voting | BCE loss | FGM | datareverse
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 1 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-TAPT --attack_type fgm --scheduler --cross_val_fold 7 --datareverse
-
0.77677520596 | bert_spc | ERNIE-TAPT | Search_f1 | BCE loss | FGM | order_predict
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 1 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-TAPT --attack_type fgm --scheduler --cross_val_fold 5 --order_predict
-
0.78326013950 | bert_cap | ERNIE-TAPT | Search_f1 | BCE loss | FGM | batchsize=8 | diff_lr
python train.py --model_name bert_cap --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-TAPT --attack_type fgm --scheduler --train_batch_size 8 --diff_lr
-
0.78496868476 | bert_spc | ERNIE-TAPT | Search_f1 | StratifiedKFold 7-fold voting | BCE loss | FGM | batchsize=32
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 3 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-TAPT --attack_type fgm --scheduler --cross_val_fold 7 --train_batch_size 32
-
0.79014267185 | bert_spc | ERNIE-ALL-TAPT | Search_f1 | GroupKFold 7-fold voting | BCE loss | FGM
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 4 --max_length 100 --cuda 3 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-ALL-TAPT --attack_type fgm --scheduler --cross_val_fold 7 --cv_type GroupKFold
-
0.79081172 | bert_spc | ERNIE-ALL-TAPT | Search_f1 | GroupKFold 7-fold voting | GHMC loss | FGM
python train.py --model_name bert_spc --seed 1000 --bert_lr 2e-5 --num_epoch 4 --max_length 100 --cuda 2 --notsavemodel --log_step 20 --pretrained_bert_name ./pretrain_models/ERNIE-ALL-TAPT --attack_type fgm --scheduler --cross_val_fold 7 --cv_type GroupKFold --criterion ghmc