Code of IJCAI2016: "Knowledge Representation Learning with Entities, Attributes and Relations".
Evaluation results on entity prediction.
| Model | MeanRank(Raw) | MeanRank(Filter) | Hit@10(Raw) | Hit@10(Filter) |
|---|---|---|---|---|
| TransE | 259 | 200 | 35.8 | 53.0 |
| TransH | 282 | 224 | 33.9 | 50.2 |
| TransR | 260 | 200 | 37.0 | 56.1 |
| KR-EAR(TransE) | 186 | 133 | 38.5 | 54.5 |
| KR-EAR(TransR) | 172 | 118 | 39.5 | 57.3 |
Evaluation results on relation prediction.
| Model | MeanRank(Raw) | MeanRank(Filter) | Hit@10(Raw) | Hit@10(Filter) |
|---|---|---|---|---|
| TransE | 3.1 | 2.8 | 65.9 | 83.8 |
| TransH | 3.4 | 3.1 | 64.9 | 84.1 |
| TransR | 3.4 | 3.1 | 65.2 | 84.5 |
| KR-EAR(TransE) | 2.4 | 2.1 | 67.9 | 86.2 |
| + CRA | 1.8 | 1.6 | 70.9 | 88.7 |
| KR-EAR(TransR) | 2.6 | 2.2 | 66.8 | 89.0 |
| + CRA | 1.9 | 1.6 | 71.5 | 90.4 |
Evaluation results on attribute prediction.
| Model | MeanRank(Raw) | MeanRank(Filter) | Hit@10(Raw) | Hit@10(Filter) |
|---|---|---|---|---|
| TransE | 10.7 | 5.6 | 36.5 | 55.9 |
| TransH | 10.7 | 5.6 | 38.5 | 57.9 |
| TransR | 9.0 | 3.9 | 42.7 | 65.6 |
| KR-EAR(TransE) | 8.3 | 3.2 | 47.2 | 69.0 |
| +AC | 7.5 | 3.0 | 49.4 | 70.4 |
| KR-EAR(TransR) | 8.3 | 3.2 | 47.6 | 69.8 |
We provide FB24k dataset used for the task knowledge base completion in data.zip, using the input format required by our codes.
Datasets are required in the folder data/ in the following format, containing nights files:
-
train-rel.txt: training file of relations, format (e1, e2, rel).
-
test-rel.txt: test file of relations, same format as train-rel.txt.
-
train-attr.txt: training file of attributes, format (e1, val, attar).
-
test-attr.txt: test file of attributes, same format as train-attr.txt.
-
entity2id.txt: all entities and corresponding ids, one per line.
-
relation2id.txt: all relations and corresponding ids, one per line.
-
attribute2id.txt: all attributes and corresponding ids, one per line.
-
val2id.txt: : all values and corresponding ids, one per line.
-
attribute_val.txt: the value set of each attribute
The codes are in the folder KR-EAR(TransE)/, KR-EAR(TransR)/.
Just type make in the folder ./
You need to type the following command in each model folder:
For training:
./main
For testing:
./test
./test_attr
You can also change the parameters when training.
-n : the embedding size of entities, relations
-m : the embedding size of values
-margin: the margin length
If you use the code, please kindly cite the following paper:
Yankai Lin, Zhiyuan Liu, Maosong Sun. Knowledge Representation Learning with Entities, Attributes and Relations. International Joint Conference on Artificial Intelligence (IJCAI 2016).