Knowledge and Information Systems Взаимодействие человека и компьютера

TransDualKGR: a transformer-based dynamic representation and dual-agent framework for multi-hop reasoning over knowledge graphs

Li, Guanfeng, Wang, Shizhuo
Springer Nature · 2026-10-30 · DOI 10.1007/s10115-026-02931-x

Аннотация

To tackle the challenges posed by the rigidity of static representations, the coupled nature of the action space, and the oversimplified reward mechanisms in multi-hop knowledge graph reasoning, we introduce the TransDualKGR framework. We combine a Transformer-enhanced dynamic representation module with a dual-agent (relation and entity) architecture to enable collaborative reasoning. Specifically, the Transformer module employs a relation-conditioned attention mechanism and Graph Positional Encoding (GraphPE) to capture topological structures. Meanwhile, we adopt PageRank-based neighbor sampling to generate dynamic context-aware entity embeddings. Moreover, we innovatively devise a hierarchically decoupled dual-agent architecture. The relation agent is designed to optimize the search within the relation space, effectively compressing the action space to |R| dimensions. On the other hand, the entity agent locates the target entity based on the selected relation, restricting the action space to the subset of entities connected by that relation. These two agents collaborate through a shared LSTM state. We also propose a quadruple dynamic reward function to jointly optimize four aspects: the destination reward, the path efficiency reward, the PageRank-guided entity influence factor reward, and the diversity reward driven by dynamic embedding variance. Extensive experiments conducted on FB15k-237, WN18RR, and NELL-995 demonstrate that our framework outperforms the state-of-the-art baselines on most metrics. It improves the Mean Reciprocal Rank (MRR) by 6.0%, 2.4%, and 4.4% on these datasets, respectively. Additionally, on FB15k-237 the path diversity index reaches 0.43 at T = 3.

Читать у издателя →