To run this repo, you need to install pytorch>=1.4, transformers, learn2learn.
First, set the corresponding files inconfig and scripts.
For scripts, select the mode, see Table 1 for more details. For config, it's enough to adjust model most times.
set model of `ace.json` in `config` to "Main model"
set mode of `train.sh` in `scripts` to "train"
run the command "bash scripts/train.sh"Train base model first, then train routing, head expert and tail expert.
Train routing:
set model of `ace.json` in `config` to "Selector"
set mode of `train.sh` in `scripts` to "train"
run the command "bash scripts/train.sh"Train head expert:
set model of `ace.json` in `config` to "Head"
set mode of `train.sh` in `scripts` to "train"
run the command "bash scripts/train.sh"Train tail expert:
# meta-train first
set model of `ace.json` in `config` to "Meta"
set mode of `train.sh` in `scripts` to "meta"
run the command "bash scripts/train.sh"
# fine-tune
set model of `ace.json` in `config` to "FewRole"
set mode of `train.sh` in `scripts` to "train"
run the command "bash scripts/train.sh"Evaluate RBDEF:
# no training, only evaluation
set model of `ace.json` in `config` to "Fuse"
set mode of `train.sh` in `scripts` to "threshold"
run the command "bash scripts/train.sh"The result can be seen in logs.
For more details, see code.
Table 1
| mode | sub_mode | argument | explanation |
|---|---|---|---|
| preprocess | - | - | preprocess the data from sentence-level to entity-level |
| train | - | load the saved model or not | training, set the configuration in config
|
| evaluate | - | load the saved model or not | evaluation, set the configuration in config
|
| statistic | - | - | compute and save role2entity.json and role2event.json |
| indicator | filename of saved model | - | test on devset, run this before rank mode |
| rank | filename of the indicator's result | - | rank roles by F1, run this after indicator mode |
| important | - | - | set the flag whether or not the sample belong to |
| meta | - | - | meta-training, set the configuration in config
|
| threshold | - | - | evaluate the RBDEF |
| save | filename of saved model | - | save the classifier from the saved model |
| fewshot | - | - | fine-tuning the initializations search by different meta learning algorithms, see sec 4.4 in paper |
| group | - | - | divide dataset into different groups for training Base+Fairness |
| parameter | - | - | compute the number of parameters of models |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent