RBERT for Relation Extraction task for KLUE

Project Description

Relation Extraction task is one of the task of Korean Language Understanding Evaluation(KLUE) Benchmark.
Relation extraction can be defined as multiclass classification task for relationship between subject entity and object entity.
Classes are such as no_relation, per:employee_of, org:founded_by... totaling 30 labels.
This repo contains custom fine-tuning method utilizing monologg's R-BERT Implementation.
Custom punctuations with Pororo NER has been added to the dataset prior to the model's training.
If you want to refer to the experimentation note such as punctuation method of the entity, please refer to the blog post

RBERT structure can also be used on Code Clone Detection Task and Natural Language Inference Task.

Argument	type	Default	Explanation
batch_size	int	32	batch size for training and inferece
num_folds	int	5	number of fold for Stratified KFold
num_train_epochs	int	5	number of epochs for training
loss	str	focalloss	loss function
gamma	float	1.0	focalloss's gamma value
optimizer	str	adamp	optimizer for training
scheduler	str	get_cosine_schedule_with_warmup	learning rate scheduler
learning_rate	float	0.00005	initial learning rate
weight_decay	float	0.01	Loss function's weight decay, preventing overfit
warmup_step	int	500
debug	bool	false	debug with CPU device for better error representation
dropout_rate	float	0.1
save_steps	int	100	number of steps for saving the model
evaluation_steps	int	100	number of step until the evaluation
metric_for_best_model	str	eval/loss	the metric for determining which is the best model
load_best_model_at_end	bool	True

Name		Name	Last commit message	Last commit date
Latest commit History 27 Commits
data		data
.gitignore		.gitignore
README.md		README.md
config.yaml		config.yaml
dataset.py		dataset.py
inference.py		inference.py
loss.py		loss.py
metrics.py		metrics.py
models.py		models.py
requirements.txt		requirements.txt
train.py		train.py