Power Entity Alignment Active Learning Framework Based on Semantic and Structure Sampling Strategy

Entity alignment (EA), which aims to match equivalent entities in different knowledge graphs, is essential in knowledge fusion. The current mainstream method is the neural network EA model, which relies on seed alignment training data and has a very high labeling cost. In this paper, we designed an active learning framework for a neural network EA model based on semantic and structure-aware sampling strategy. The framework aims to create highly informative data to produce more efficient EA models with lower labeling costs. Our framework addresses the main challenge of applying active learning techniques to EA models: how to leverage dependencies between entities and semantic information about the entities themselves in an active learning strategy. Most active learning strategies assume that the data instances to be sampled are independent and uniformly distributed. However, the entities in the knowledge graph are related. To solve this problem, we compare the effects of four sampling strategies on three EA models on power standard datasets. The experimental results show that the proposed semantic and structure-aware active learning sampling strategy reduces the labeling amount by 10%-20% compared with other strategies and dramatically reduces the labeling cost.