An implementation of our model with PyTorch.
We provide FB15k and FB15k-237 datasets used in our experiment, the data set can be downloaded here
Each Dataset is organized in the following format (using FB15k as example):
fb15k
├─train.txt: training data
├─valid.txt: valid data
├─test.txt: test data
└─encode
├─entity_id.txt: entity to id mapping
├─relation_id.txt: relation to id mapping
├─train_encode.txt: encoded training data
├─valid_encode.txt: encoded valid data
└─test_encode.txt:encoded test datadump.rdb is a dump of Redis containing relational paths of each dataset. In the dump, db0 and db1 correspond to FB15k and FB15k-237 respectively. In each db, key is the composition of entity pair and number of steps, while value is a mapping from the relational path to its frequency in the form of JSON. For example, to query about the 2-step relational paths between entity 0 and entity 1020, the key is "0,1020:2", and the returned value is
{
"[481, -767]": 1,
"[285, -767]": 1,
"[525, -767]": 1,
"[84, -767]": 1
}