Answers are grounded in this paper and cite their sources.
Knowledge graphs (KGs) are widely used to represent and propagate semantic information, yet their structural properties can induce systematic distortions in learned representations. In this paper, we study how structural bias in a knowledge graph influences the formation of bipolar semantic embeddings when supervision is provided by two semantically opposing anchor concepts. We introduce two evaluation metrics, anchor polarity separation score (APSS) and global polarity separation score (GPSS) which quantify anchor-specific alignment and axis-wide semantic separation, respectively. Using five pairs of semantically opposing anchor concepts, along with additional anchor pairs selected according to node degree and separation-cohesion ratio, we systematically analyze the influence of degree, neighborhood composition, and graph topology on induced semantic polarity. Experimental results show that bipolar embeddings are shaped by the interaction between anchor semantics and structural properties of the graph, rather than by any single structural characteristic, and that these effects are consistently observed across different graph neural network (GNN) architectures. The results further demonstrate that APSS and GPSS provide effective and interpretable measures for assessing the quality and stability of bipolar semantic axes in knowledge graph embeddings.
No takes yet. Share an insight, caveat, or question.
Janghoon Yang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: