← Back to feed
2026-05-25reasoningdatacode

Neural Scalable Symbolic Search Framework for Complex Logical Queries with Multiple Free Variables

Weizhi Fei, Hang Yin, Zihao Wang, Shukai Zhao, Wei Zhang, Yangqiu Song

PDF preview unavailable
Read on arXiv →

Key claim

NS3 improves joint ranking for complex queries in KGs.

The NS3 framework offers a novel approach to answering complex queries over knowledge graphs by approximating joint rankings without exhaustive enumeration. It improves joint ranking performance while maintaining strong accuracy on marginal queries. This advancement is particularly valuable for practitioners dealing with multi-variable queries in knowledge representation.

In plain English

The NS3 framework offers a novel approach to answering complex queries over knowledge graphs by approximating joint rankings without exhaustive enumeration. It improves joint ranking performance while maintaining strong accuracy on marginal queries. This advancement is particularly valuable for practitioners dealing with multi-variable queries in knowledge representation.

Novelty
8.0/10

The proposed NS3 framework significantly advances the approach to joint ranking in complex query answering over knowledge graphs.

Reliability
8.0/10

The paper provides strong experimental results across multiple datasets and releases code for reproducibility.

Deep reliability assessment

The methodology supports improved joint ranking for EFOk queries through a budgeted framework, but it may overclaim effectiveness by not addressing all complexities of larger k-variable queries. The results are promising but may not generalize across all types of knowledge graphs or query structures.

Reproducibility

Yes, the source code is provided at https://github.com/HKUST-KnowComp/NS3_KDD2026.

Discussion questions

  1. 1.What assumptions about the completeness of knowledge graphs underlie the effectiveness of the proposed method?
  2. 2.How can the findings be applied to real-world applications in industries like finance or security?
  3. 3.What would happen if the knowledge graph had a significantly higher degree of incompleteness than tested?

Key figure

Figure 1 visualizes an EFO2 query related to fraudulent activities, showing its logical formula and query graph structure.

Benchmark results

FB15k-237HIT@10: 9vs FIT+2.3%SOTA
FB15kHIT@10: 61.3vs FIT+2.5%SOTA
NELLHIT@10: 32.8vs FIT+4.6%SOTA
GitHub1 repo
HKUST-KnowComp/NS3_KDD2026Official