Various tests on ACL 2019 paper: Learning Attention-based Embeddings for Relation Prediction in Knowledge Graphs
Baseline code taken from relationPrediction.
Direct reproduction of baselines is present in the baseline folder on Freebase and Wordnet. Scripts to reproduce the results are fb15.sh and wordnet.sh, and the corresponding outputs are fb15_out.txt and wordnet_out.txt, respectively.
-
Using only GAT trained model to evaluate the results (without ConvKB)
Results present in folder
baseline/- file istest_gat.pyand script isgatonly.sh -
DISTMULT scoring function instead of TransE
Results present in folder
tests/distmult/ -
Random Initialization of Entity and Relational Embeddings instead of TransE
Results present in folder
tests/transe/