Proceedings of the Annual Meeting of the Cognitive Science Society

Human readers derive coherent relational understanding from information distributed across a document through a progressive, hierarchical process: local semantic associations are established first and then integrated into global coherence. How the mind achieves such robust integration while limiting combinatorial explosion and error propagation remains an open question. We propose progressive coherence building as a key computational principle, in which complex inference is constrained through sequentially layered, verifiable representations. Guided by this hypothesis, we introduce a Hierarchical Cognitive Model (HCM) for cross-context relational learning. HCM instantiates three cognitively motivated stages: (1) lexical-semantic grounding via prompt-based semantic priming; (2) local proposition formation through intra-triple attention that enforces local semantic consistency; and (3) global coherence optimization using prior-guided axial attention to integrate evidence across the narrative. This staged design stabilizes intermediate representations and mitigates error propagation. Experiments on CDR, GDA, and DocRED demonstrate state-of-the-art performance. Ablation results reveal cumulative degradation when higher stages are removed, empirically supporting the proposed hierarchical processing account.