Pytorch implementations of model-centric and data-centric domain-adaption techniques for dependency parsing.
NLPCC-2019依存句法分析领域移植评测任务数据集
源域:平衡语料(BC)
目标域:
- 淘宝产品博客(PB)
- 淘宝产品评论(PC)
- 网络小说《诛仙》(ZX)
nohup python train_ts.py --cuda 1 -lr 1e-3 --bert_lr 2e-5 --bert_layer 4 --batch_size 32 --src_type BC --tgt_type ZX --d_model 768 --arc_size 640 --label_size 128 &> zx_.log &
nohup python train_ts.py --cuda 2 -lr 1e-3 --bert_lr 2e-5 --bert_layer 4 --batch_size 32 --src_type BC --tgt_type PC --d_model 768 --arc_size 640 --label_size 128 &> pc_.log &
nohup python train_ts.py --cuda 3 -lr 1e-3 --bert_lr 2e-5 --bert_layer 4 --batch_size 32 --src_type BC --tgt_type PB --d_model 768 --arc_size 640 --label_size 128 &> pb_.log &
- Domain-Adaptive Pretraining:MLM + Finetuning
- Teacher-Student Learning:pretrained teacher + to-be-trained student
- Multi-task Tri-training:shared BERT + separated Biaffines
- Full Finetuning
- DistanceNet:Dual BERT + domain distance metrics
- Domain Adversarial Training:GRL + domain classification
- VAE Reconstruction:autoencoder