KGCompletion / TransL

TransL

The source code of paper "Translating Embedding with Local Connection for Knowledge Graph Completion".

Link prediction on FB15k-237:

	Raw MRR	Filter MRR	Hits@1	Hits@3	Hits@10
unif	0.227	0.342	0.244	0.379	0.535
bern	0.248	0.355	0.260	0.389	0.551

Triplet classification on WN11 and FB13:

	WN11	FB13
unif	0.861	0.838
bern	0.866	0.856

Data

We provide FB15k-237, FB13 and WN11 datasets used for the tasks of link prediction and triplet classification.
Each dataset in the following format, containing five files:

• entity2id.txt: all entities and corresponding ids, format (entity, id)
• relation2id.txt: all relations and corresponding ids, format (relation, id)
• test.txt: testing file, format (head_entity, relation, tail_entity, label)
• train.txt: training file, format (head_entity, relation, tail_entity)
• valid.txt: validation file, format (head_entity, relation, tail_entity, label)

Training

Usage:

python code/train.py

You can change the hyper-parameters.
-dim: entity and relation sharing embedding dimension
-margin_pos: margin of positive triplets
-margin_neg: margin of negative triplets
-rate: learning rate
-batch: batch size
-epoch: number of training epoch
-method: stratege of constructing negative triplets, options: unif, bern
-data: dataset of the model, options: FB15k-237, WN11, FB13

Testing

Usage: Link prediction:

python code/test-lp.py

Triplet classification:

python code/test-tc.py

You can change the hyper-parameters.
-dim: entity and relation sharing embedding dimension
-margin_pos: margin of positive triplets
-margin_neg: margin of negative triplets
-rate: learning rate
-batch: batch size
-epoch: number of training epoch
-method: stratege of constructing negative triplets, options: unif, bern
-data: dataset of the model, options: FB15k-237, WN11, FB13

It will evaluate on test.txt and report the results.