The way of obtaining the embeddings of the knowledge graph objects through modeling with binary classification method from the level of triple structure is coarser in granularity for the existing knowledge representat...
详细信息
The way of obtaining the embeddings of the knowledge graph objects through modeling with binary classification method from the level of triple structure is coarser in granularity for the existing knowledge representation learning models based on the probability, and the space-time efficiency of negativesampling is lower for the most of the knowledge representation learning models at present. To solve these problems, this paper proposes a knowledge representation learning model KRL_Match, which carries out the knowledge graph objects matching centered on a certain kind of knowledge graph objects (head entity, tail entity, relation), and executes multi-classification learning to determine the true matching and dynamic implicit negative sampling. Specifically, first, we make two classes of the knowledge graph objects of target and source in the same kind of knowledge graph objects matched mutually by their matrix multiplication operation in a knowledge graph batch sample space, which is constructed by random sampling from the universe set of the knowledge graph instance, and the knowledge graph objects matching sample spaces will be implicitly generated meanwhile;then, we measure the matching degree of each matching of the knowledge graph objects by softmax regression multi-classification method in each implicit sample space;finally, we fit the real probability with the matching degree by optimizing the cross-entropy loss based on the local closed world assumption. We conduct the knowledge graph objects matching for the knowledge representation learning inspired by the attention mechanism and firstly create the dynamic implicit negative sampling method in the knowledge representation learning. Experiments show that the KRL_Match model has achieved better performances compared with the baselines: Hits@10 (filter) has increased by 12.2% and 6.1% on benchmarks FB15K and FB15K237 respectively for the entity prediction task, and accuracy has increased by 12.6% on benchmark
暂无评论