Abstract
Neuro-symbolic models for Knowledge Graph Completion (KGC) often underperform due to low logical rule coverage. This paper proposes three simple rule-augmentation techniques—abduction, inversion, and random walks—combined with a quality-control filtering step. These methods significantly enhance performance, achieving gains of up to 7.1 points in Mean Reciprocal Rank (MRR) and establishing new state-of-the-art results on benchmarks like WN18RR.
Introduction
Knowledge Graphs (KGs) such as Freebase and WordNet are essential but notoriously incomplete. Knowledge Graph Completion (KGC) typically employs either neural embeddings, which are scalable but act as "black boxes," or symbolic rules, which are interpretable but difficult to scale. Neuro-symbolic (NS-KGC) models attempt to integrate both approaches to achieve the best of both worlds.
However, a critical bottleneck in NS-KGC is "coverage": the generated symbolic rules often fail to "fire" for specific queries, forcing the model to rely entirely on its neural component. This research aims to strengthen symbolic inferences by expanding the rule set through simple yet effective augmentations without sacrificing quality.
Methodology
The authors propose three primary augmentation strategies to transform small, existing rule sets into high-coverage, high-quality ones:
- Abduction: This involves "reverse-engineering" rules. If a deductive rule states $A \land B \Rightarrow C$, an abductive counterpart hypothesizes that if $C$ and $B$ are observed, $A$ might be true.
- Rule Inversion: This creates equivalent rules by utilizing the inverse of every relation in a path (e.g., a path A → B → C becomes C → B' → A').
- Random Walk Augmentation: New rules are independently generated by performing local random walks across the graph to identify frequent entity connections.
To prevent the rule set from including noisy or "garbage" rules, the authors implement Rule Filtering. They apply a PCA Score (a metric of rule confidence) and prune rules that fall below a specific threshold or lack sufficient "groundings" (examples) in the data.
Experimental Results
The approach was evaluated on four standard KGC benchmarks: WN18RR, FB15K-237, Kinship, and UMLS. The augmentations consistently outperformed baselines across various models, including RNNLogic+ and RotatE.
As shown in Figure 1, the performance gains are substantial. On the Kinship dataset, the researchers observed a massive 7.1 point increase in MRR and an 8.5 point gain in Hits@1. The results on WN18RR represent a new state-of-the-art for neuro-symbolic KGC models.
Analysis of Rule Augmentation
A key finding is that abduction acts as a "powerhouse" for rule generation.
As illustrated in Figure 3, abductive reasoning can nearly triple the number of high-quality rules in certain datasets. This allows the model to hypothesize explanations for facts that the neural component can subsequently validate. Furthermore, the study highlights that quality is more important than quantity; the PCA filtering step can reduce the final rule set by up to 70% while actually improving performance by removing noise.
While augmentations initially increased training time by 5-6x, the filtering step brought this down to a manageable 2-3x increase compared to the baseline, making the performance gains computationally viable.
Conclusion and Future Work
The proposed augmentations effectively bridge the "coverage gap" in neuro-symbolic models. However, limitations remain: the inversion and abduction techniques require the Knowledge Graph to support inverse relations, and the benefits may vary for models that do not rely on grounding entire rules. Future research could involve integrating these augmentations directly into the iterative rule-learning process to discover even more complex logical patterns.
Summary updated