Pre-BiafParser

Performance

Pretrained Model	Model	Dataset	UAS	LAS
---	SOTA	PTB	97.42	96.26
---	Deep Biaffine	PTB	95.87	94.22
roberta-large(supar)	Deep Biaffine	PTB	97.33	95.86
roberta-base(Our)	Deep Biaffine	PTB	96.60	95.21
bert-base-cased(Our)	Deep Biaffine	PTB	96.80	95.20
bert-large-cased(Our)	Deep Biaffine	PTB	96.95	95.29
---	---	---	---	---
---	Deep Biaffine	CTB-5	89.30	88.23
Bert(Fixed-8)	Deep Biaffine	CTB-5	92.96	91.80
electra-large(supar)	Deep Biaffine	CTB-5	92.45	89.55
electra-base(Our)	Deep Biaffine	CTB-5	90.66	87.07
bert-base-multi(Our)	Deep Biaffine	CTB-5	91.26	89.96

Roberta learing-rate： 2e-5， batch_size: 48, epoch: 20

Bert Base EN pretr_lr： 2e-5，other_lr: 5e-4, adatrans：3-layers, heads：8, batch_size: 48, epoch: 100 -----------------96.74 | 95.13

Bert Base pretr_lr： 1e-5，other_lr: 2e-4, adatrans：3-layers, heads：8, batch_size: 48, epoch: 120 -----------------96.80 | 95.21

Bert Large EN pretr_lr： 1e-5，other_lr: 2e-4, adatrans：3-layers, heads：8, batch_size: 30, epoch: 100

Bert CN pretr_lr： 2e-5，other_lr: 5e-4, adatrans：3-layers, heads：8, batch_size: 48, epoch: 120

Bert Base EN frozen pretr_lr： 0，other_lr: 2e-4, adatrans：3-layers, heads：8, batch_size: 48, epoch: 100 -----------------96.63 | 94.92

Related Repo

Bert(Fixed-8) https://github.com/LindgeW/BiaffineParser

roberta-large(supar) https://github.com/yzhangcs/parser

File: giga.100.txt

https://github.com/yzhangcs/parser/releases/download/v1.1.0/giga.100.zip