Source Code for "Frame Semantic Role Labeling Using Arbitrary-Order Conditional Random Fields" at AAAI-2024.
Our code is based on AGED, thanks for their great work!
The preprocessed FrameNet 1.5 and 1.7 data are the same as AGED.
We show the hyper-parameter settings in the following table.
Hyper-parameters | Value |
---|---|
Pretrained Language Model | bert-base-uncased |
BERT embedding dimension | 768 |
batch size | 32 |
optimizer | BertAdam |
scheduler | linear warmup |
warmup ratio (train only) | 0.05 |
warmup ratio (pretrain) | 0.01 |
warmup ratio (fine-tune) | 0.05 |
learning rate (train only) | 5e-5 |
learning rate (pretrain) | 5e-5 |
learning rate (fine-tune) | 2.5e-5 |
gradient clipping | 5.0 |
MLP layers | 1 |
MLP activation function | ReLU |
MLP dimension | 768 |
rank dimension | 512 |
mean-field inference iterations | 3 |
epoch num (train only) | 20 |
epoch num (pretrain) | 5 |
epoch num (finetune) | 10 |
bash first.sh
bash first_pretrain.sh
bash arbitrary.sh
bash arbitrary_pretrain.sh