Adversarial Knowledge Representation Learning Without External Model

Knowledge representation learning, which embeds entities and relations of knowledge graph into low-dimensional vectors, is efficient for predicting missing facts. Knowledge graph datasets only store positive triplets. Nevertheless, negative cases are similarly crucial in knowledge representation learning. Conventionally, corrupted triplets are uniformly generated as negative cases, but actually, these corrupted triplets are heterogeneous. The majority of corrupted triplets are trivial, and they have limited influence on learning. Regarding the large number of corrupted triplet candidates, it is not efficient to train the model by uniformly generated corrupted triplets. Generative adversarial network (GAN)-inspired approaches are proposed to remit easily discriminated negative training examples, enabling faster and better convergence of the embedding models. Pre-trained external sampling models are required in these approaches. In this paper, we introduce a simple but strong negative sampling approach for adversarial knowledge representation learning, named loss adaptive sampling mechanism, which is efficient without an external sampling model. Furthermore, false negative cases are always over-trained in the training stage with efficient negative sampling approaches. We propose a push-up mechanism and verify whether it is feasible to alleviate these over-trained false negative cases. The experimental results show that our adversarial knowledge representation learning approach outperforms the GAN-based sampling method—KBGAN.

Abstract licence: CC BY-NC-ND

39 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.