Code for Analyzing the Influence of Knowledge Graph Information on Relation Extraction

Paper: https://arxiv.org/pdf/2506.16343

Download all the data from: https://drive.google.com/file/d/1mN5OL6UrWDKr0lxyzjNJxWmk6pvJr2zP/view?usp=sharing

Install the ULTRA model from the following repository: https://github.com/DeepGraphLearning/ULTRA. It is the prerequisite for all further experiments.

Zero-shot setting

Steps to reproduce

1. Unpack the following files:

   • wiki.tar.gz
   • fewrel.tar.gz
   • rel_id_title_description_cleaned.jsonl.gz
   • graph_wiki.gt.gz
   • graph_fewrel.gt.gz

2. Run the following command to train the model for each dataset:

   python3 src/train.py ...
   

3. Evaluate on each seed of each dataset by using the following command:

   python3 src/evaluate.py ...
   

Functions

Arguments for train.py:

Argument	Type	Default Value	Description
--dataset_name	str	"fewrel/unseen_5"	Specifies the name of the dataset. This executes the training for all seeds as specified by the --seeds parameter.
--model_type	str	"bert-base-cased"	Specifies the type of model to be used.
--batch_size	int	24	Sets the batch size for training.
--num_workers	int	2	Number of worker processes for data loading.
--accumulate_grad_batches	int	2	Accumulates gradients over a specified number of batches.
--lr	float	5e-5	Learning rate for optimization.
--seeds	int, List	[0, 1, 2, 3, 4]	List of seeds of the dataset to train on.
--include_descriptions	bool	False	Includes descriptions in the textual representation if this flag is present.
--deactivate_graph	bool	False	Deactivate the graph component.
--only_graph	bool	False	Only use the graph component.
--use_filtered_meta_graph	bool	False	Use filtered graph.

Arguments for evaluate.py:

Argument	Type	Default Value	Required	Description
--model_checkpoint	str	-	Yes	Specifies the path to the model checkpoint.
--dataset_name	str	"fewrel/unseen_5_seed_0"	No	Specifies the name of the dataset with the corresponding seed.
--model_type	str	"bert-base-cased"	No	Specifies the type of model to be used.
--batch_size	int	24	No	Sets the batch size for training.
--num_workers	int	2	No	Number of worker processes for data loading.
--accumulate_grad_batches	int	1	No	Accumulates gradients over a specified number of batches.
--other_properties	int	5	No	Specifies the value for some other properties.
--include_descriptions	bool	False	No	Includes descriptions in the textual representation if this flag is present.
--deactivate_graph	bool	False	No	Deactivate the graph component.
--only_graph	bool	False	No	Only use the graph component.
--use_filtered_meta_graph	bool	False	No	Use filtered graph.

Supervised setting

Steps to reproduce

1. Unpack the following files:

   • dwie.tar.gz
   • redocred.tar.gz
   • biorel.tar.gz
   • graph_dwie.gt.gz
   • graph_docred.gt.gz
   • graph_biorel.gt.gz

2. Run the following command to train the model for each dataset:

   python3 src/train.py ...
   

3. Evaluate checkpoint for each dataset by using the following command:

   python3 src/evaluate.py ...
   

Functions

Arguments for train.py:

Here is a table similar to the one you provided, which describes the parameters for your method:

Argument	Type	Default Value	Description
--model_checkpoint	str	None	Specifies the path to a pre-trained model checkpoint.
--batch_size	int	4	Sets the batch size for training and evaluation.
--num_rules	int	1	Number of rules to be used in the model.
--num_hops	int	4	Number of hops for the reasoning process.
--max_length	int	1024	Maximum sequence length for input examples.
--num_epochs	int	30	Number of training epochs.
--gradient_accumulation_steps	int	1	Number of steps to accumulate gradients before updating.
--graph_only	bool	False	Use only the graph component if this flag is present.
--use_hinge_abl	bool	False	Uses hinge loss for abductive reasoning if this flag is present.
--use_at_loss	bool	False	Activates the use of ATLoss if this flag is present.
--lr_encoder	float	3e-5	Learning rate for the encoder component.
--lr_classifier	float	1e-4	Learning rate for the classifier component.
--random_dropout	float	0.2	Dropout rate applied randomly to the model layers during training.
--deactivate_graph	bool	False	Deactivates the graph component if this flag is present.
--short_cut	bool	False	Enables shortcut connectivity if this flag is present.
--use_biorel	bool	False	Utilizes BioRel dataset if this flag is present.
--use_dwie	bool	False	Utilizes DWIE dataset if this is specified.
--use_docred	bool	False	Utilizes DocRED dataset if this flag is present.
--remove_direct_links	bool	False	Removes direct links in the graph or data if this flag is present.
--graph_dim	int	64	Dimensionality of graph embeddings.
--seed	int	42	Seed for random number generation to ensure reproducibility.
--post_prediction	bool	False	Activates post-prediction processing if this flag is present.
Arguments for evaluate.py:

Argument	Type	Default Value	Description
model_checkpoint	str	None	Specifies the path to a pre-trained model checkpoint. This is a required positional argument.
--batch_size	int	4	Sets the batch size for training and evaluation.
--num_rules	int	1	Number of rules to be used in the model.
--num_hops	int	4	Number of hops for the reasoning process.
--max_length	int	1024	Maximum sequence length for input examples.
--gradient_accumulation_steps	int	1	Number of steps to accumulate gradients before updating.
--graph_only	bool	False	Use only the graph component if this flag is present.
--deactivate_graph	bool	False	Deactivates the graph component if this flag is present.
--short_cut	bool	False	Enables shortcut connectivity if this flag is present.
--use_biorel	bool	False	Utilizes BioRel dataset if this flag is present.
--use_dwie	bool	False	Utilizes DWIE dataset if specified.
--remove_direct_links	bool	False	Removes direct links in the graph or data if this flag is present.
--separated	bool	False	Processes inputs as separated components if this flag is present.
--graph_dim	int	64	Dimensionality of graph embeddings.
--post_prediction	bool	False	Activates post-prediction processing if this flag is present.