Skip to content

code for " Query Your Model with Definitions in FrameNet: An Effective Method for Frame Semantic Role Labeling" at AAAI-2023

Notifications You must be signed in to change notification settings

Zce1112zslx/AGED

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 

Repository files navigation

🎅AGED

Code and Technical Appendix for " Query Your Model with Definitions in FrameNet: An Effective Method for Frame Semantic Role Labeling" at AAAI-2023

Data Download

Preprocessed FrameNet 1.5 and 1.7 data google drive

🤖Technical Details

We use bert-base-uncased as the Pretrained Language Model (PLM) in 🎅🏻AGED. We follow Chen, Zheng, and Chang (2021); Zheng et al. (2022) to first train 🎅🏻AGED on exemplar sentences then train on the train set continually. We search for hyperparameters (learning rate, batch size, and epoch num) with performance in the development set. Performance in the development set is also used to save the best parameters of the models, and we will evaluate 🎅🏻AGED with these parameters in the test set.

Our code is implemented with Pytorch and Huggingface. 🎅AGED is trained on NVIDIA A40 with 40 GB memory and it will take about 4 GPU hours to train 🎅🏻AGED and 0.6 hours when 🎅🏻AGED is trained only with the training dataset.

We show the hyper-parameter settings of 🎅🏻AGED in the following table.

Hyper-parameter Value
BERT version bert-base-uncased
batch size 16/32
learning rate (train only) 1e-5/5e-5
learning rate (pretrain) 5e-5
learning rate (fine-tune) 1e-5/5e-5
lr scheduler Linear decay
warm up ratio 0.1
optimizer BertAdam
epoch num (train only) 15/20/25
epoch num (pretrin) 5/10
epoch num (fine-tune) 10/20

Here, "train only" means 🎅🏻AGED trained only with training dataset. "pretrain" means 🎅🏻AGED w/exemplar trained with exemplar instances. "fine-tune" means 🎅🏻AGED w/exemplar then trained with training dataset.

We refer readers to technical appendix for more details.

About

code for " Query Your Model with Definitions in FrameNet: An Effective Method for Frame Semantic Role Labeling" at AAAI-2023

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages