Scholarlink

<p>Complex query answering (CQA) on knowledge graphs aims to find answers in incomplete knowledge bases through multi-step logical reasoning. Recent studies have proposed logical message passing mechanisms based on pre-trained neural link predictors. However, existing message passing models, when aggregating logical messages, often rely on their own modeling capacity to implicitly handle positive and negative messages, thereby overlooking the intrinsic differences between logical polarities such as affirmation and negation. To address these shortcomings, we propose the polarity-aware Griffin message passing network (PG-MPN). The core idea of this network is to differentiate and hierarchically process the logical messages delivered by the pre-trained neural link predictor. First, PG-MPN explicitly distinguishes between positive and negative messages and enhances the representation of negative messages via a nonlinear transformation. Subsequently, independent attention weights are applied to each type of message separately. At the core of the framework, the Griffin architecture is employed as an efficient sequence processor to capture the complex dependencies among the weighted positive and negative messages. Finally, an adaptive weighted aggregation is performed on the output of Griffin to generate more informative node representations. Experimental results demonstrate that modeling the polarity of logical messages, combined with a powerful sequence processing architecture, enhances the model’s ability to understand and reason over complex logical structures. The source code is available at <a href="https://github.com/HanDongqi/PG-MPN">https://github.com/HanDongqi/PG-MPN</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

Complex query answering on knowledge graphs with Griffin and polarity-weighted message passing

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Dongqi Han,
  • Yao Zhang,
  • Fanghao Li,
  • Shengli Wu,
  • Hu Lu

摘要

Complex query answering (CQA) on knowledge graphs aims to find answers in incomplete knowledge bases through multi-step logical reasoning. Recent studies have proposed logical message passing mechanisms based on pre-trained neural link predictors. However, existing message passing models, when aggregating logical messages, often rely on their own modeling capacity to implicitly handle positive and negative messages, thereby overlooking the intrinsic differences between logical polarities such as affirmation and negation. To address these shortcomings, we propose the polarity-aware Griffin message passing network (PG-MPN). The core idea of this network is to differentiate and hierarchically process the logical messages delivered by the pre-trained neural link predictor. First, PG-MPN explicitly distinguishes between positive and negative messages and enhances the representation of negative messages via a nonlinear transformation. Subsequently, independent attention weights are applied to each type of message separately. At the core of the framework, the Griffin architecture is employed as an efficient sequence processor to capture the complex dependencies among the weighted positive and negative messages. Finally, an adaptive weighted aggregation is performed on the output of Griffin to generate more informative node representations. Experimental results demonstrate that modeling the polarity of logical messages, combined with a powerful sequence processing architecture, enhances the model’s ability to understand and reason over complex logical structures. The source code is available at https://github.com/HanDongqi/PG-MPN.