Recent developments in Complex Query Answering (CQA) highlight a crucial trade-off between logical soundness and computational efficiency when working with incomplete Knowledge Graphs (KGs). This paper discusses the Grounding-Skolemization dichotomy, which sheds light on existing methodologies that often fail to balance these two requirements. The Grounding-based methods typically face combinatorial explosion, while Skolemization-based methods may overlook important aspects of logical consistency.
To overcome these limitations, the authors propose the Logic-constrained Vector Symbolic Architecture (LVSA), a novel neuro-symbolic framework that combines a differentiable Skolemization module with a neural negator, facilitated by a logical constraint-driven optimization process. This innovative approach theoretically guarantees universality for all Existential First-Order (EFO1) queries. Empirical evaluations show that LVSA not only outperforms existing Skolemization methods but also achieves a dramatic reduction in inference costs compared to traditional Grounding-based approaches.
👉 Pročitaj original: arXiv AI Papers
