Reproduce “Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yisen Gao, Jiaxin Bai, Tianshi Zheng, Qingyun Sun, Ziwei Zhang, Xingcheng Fu, Jianxin Li, Yangqiu Song

The CtrlHGen framework introduces controllable logical hypothesis generation for abductive reasoning in knowledge graphs, enabling users to guide hypothesis creation by specifying semantic content or structural complexity. This approach yields improved semantic similarity and adherence to control conditions, outperforming large language models in generating precise, structured explanations.

Key insights

  • Controlling both the semantic content and structural complexity of generated hypotheses is feasible and significantly enhances the practical utility of abductive reasoning in knowledge graphs.
  • Learning complex logical structures is greatly facilitated by decomposing them into simpler, foundational sub-logical patterns during training, improving generalization.
  • A multi-metric semantic reward, coupled with a specific condition adherence reward, provides a more stable and effective learning signal for reinforcement learning in generating complex logical queries than a single, strict metric.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.