A Hybrid Soft–Hard Rule Learning Framework With Layout-Type Constraints for Poster-Level Spatial Relation Extraction

Poster-level relation extraction (PosterRE) identifies spatial and semantic relationships among layout elements. While recent rule-based approaches improve model effectiveness, pipeline methods suffer from error propagation and end-to-end methods risk learning spurious rules. A critical limitation is neglecting element type information during rule acquisition. Building upon established neural-symbolic learning principles, we present Soft-Hard Rules with Layout Type Constraints (SH-LTRs), a framework that adapts hybrid soft-hard rule learning to spatial reasoning domains. Our approach combines a Hard Layout Type Rules Module (H-LTRs) that establishes rigid type-aware constraints through spatial pattern mining with a Soft Layout Type Rule Reasoning Module (S-LTRs) that parameterizes rule inference via neural networks. Through a rule consistency loss function, S-LTRs learns under hard rule guidance, preventing incorrect rule acquisition. Extensive experiments across three benchmark datasets and two baseline architectures demonstrate state-of-the-art performance with F1 and IgnF1 scores of 87.39 and 82.43, surpassing existing rule learning frameworks for layout understanding.