literal-embeddings

Neural Regression with Embeddings for Numeric Attribute ( Literals ) Prediction in Knowledge Graphs

Installation and Requirements

To execute the Literal Embedding model, please install the dice-embeddings framework

Configurable Arguments

The following arguments can be used to configure the script. Default values are provided, and they will be used unless explicitly modified on the main script.

Default Arguments

• --dataset_dir: KGs / FamilyT
• --lit_dataset_dir: KGs/FamilyL
• --batch_size: 1024
• --num_epochs: 100
• --embedding_dim: 128
• --lr: 0.05
• --optimize_with_literals: True
• --lit_lr: 0.0001
• --lit_epochs: 500
• --save_embeddings_as_csv: False
• --save_experiment: False
• --pretrained_kge: False
• --pretrained_kge_path: None

--optimize_with_literals

• If set to True, both the Knowledge Graph Embedding (KGE) model and the Literal Embedding model are trained together in a combined manner. If set to False (default), only the KGE model is trained.

--pretrained_kge

If pretrained_kge is set to true, provide a valid path to a folder that contains a pre-trained KGE model. This will bypass all the combined training procedures and only run an instance of LiteralEmbedding model.

Usage

Set the desired args within the main script and run

python main.py