bishan, wen-tau, xiaodong, jianfeng, lideng, cornell and msr, ICLR15
Problem:
- Lack of careful comparison with TranE/NTN/RESCAL on design choices that affect the learning results.
- Evaluated on link prediction task, hard to explain what relational properties are being captured and to what extent they are captured during the embedding process.
Contributions:
- Present a general framework for multirelational learning that unifies most multi-relational embedding models developed in the past, including NTN, TranE.
- Empirically evaluate different choices of entity representations and relation representations under this framework on the canonical link predictin task and show that a simple bilinear formulation achieves new state-of-the-art results for the task.
- Propse and evaluate a novel approach that utilized the learned embeddings to mine logical Horn-clause. Demonstrate that out blabla outperforms a SOTA rule mining system AMIE on mining rules that involes compositional reasoning.
Content:
NTN/TransE differ in different parametrization of relation operators.