Scholarlink

<p>Complex Query Answering (CQA) on Knowledge Graphs aims to find answers in incomplete knowledge bases through multi-step logical reasoning. While existing message-passing approaches have made significant progress, they predominantly rely on a static, single-pass aggregation paradigm. This paradigm treats logical messages as isolated pieces of evidence fused in a one-off manner, overlooking the intricate dependencies among messages and failing to establish a precise logical consensus, particularly in scenarios involving conflicting logic (e.g., the coexistence of affirmative and negating evidence). To address this limitation, we propose the Iteration-Guided Logical Injection Network (ILIN). This model reshapes the reasoning process from static aggregation into a dynamic, multi-round message refinement and negotiation process. First, we introduce an Iterative Message Refinement mechanism. By leveraging the synergy between Multi-Head Self-Attention and Gated Recurrent Units (GRUs), messages representing different logical constraints can perceive each other within a global context and iteratively evolve, thereby progressively resolving logical ambiguities. Second, to accurately handle logical polarity, we design a Relation-Conditioned Logical Injection module, which provides differentiated, context-aware non-linear transformations for positive and negative messages during iteration. Furthermore, to address the diversity of query structures in CQA, we propose a Dynamic Task Weighting training strategy that adaptively balances the multi-task learning process and effectively prevents overfitting. Experiments on FB15k, FB15k-237, and NELL995 datasets demonstrate that ILIN significantly outperforms mainstream baselines like CLMPT, especially on queries involving complex negation and multi-hop logic. Visualization analysis further confirms that the model achieves a dynamic convergence from an ambiguous state to a precise semantic localization. The source code are publicly available at <a href="https://github.com/HanDongqi/ILIN">https://github.com/HanDongqi/ILIN.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ILIN: Iteration-Guided Logical Injection Network for complex query answering on knowledge graphs

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Dongqi Han,
  • Jinghua Yu,
  • Hu Lu

摘要

Complex Query Answering (CQA) on Knowledge Graphs aims to find answers in incomplete knowledge bases through multi-step logical reasoning. While existing message-passing approaches have made significant progress, they predominantly rely on a static, single-pass aggregation paradigm. This paradigm treats logical messages as isolated pieces of evidence fused in a one-off manner, overlooking the intricate dependencies among messages and failing to establish a precise logical consensus, particularly in scenarios involving conflicting logic (e.g., the coexistence of affirmative and negating evidence). To address this limitation, we propose the Iteration-Guided Logical Injection Network (ILIN). This model reshapes the reasoning process from static aggregation into a dynamic, multi-round message refinement and negotiation process. First, we introduce an Iterative Message Refinement mechanism. By leveraging the synergy between Multi-Head Self-Attention and Gated Recurrent Units (GRUs), messages representing different logical constraints can perceive each other within a global context and iteratively evolve, thereby progressively resolving logical ambiguities. Second, to accurately handle logical polarity, we design a Relation-Conditioned Logical Injection module, which provides differentiated, context-aware non-linear transformations for positive and negative messages during iteration. Furthermore, to address the diversity of query structures in CQA, we propose a Dynamic Task Weighting training strategy that adaptively balances the multi-task learning process and effectively prevents overfitting. Experiments on FB15k, FB15k-237, and NELL995 datasets demonstrate that ILIN significantly outperforms mainstream baselines like CLMPT, especially on queries involving complex negation and multi-hop logic. Visualization analysis further confirms that the model achieves a dynamic convergence from an ambiguous state to a precise semantic localization. The source code are publicly available at https://github.com/HanDongqi/ILIN.