SymbolFusion uses relation constraints in neurosymbolic graph learning to model literary structure and emotion trajectories

Abstract Literary texts contain recurring imagery, typed symbolic relations, and long-range emotion shifts that sequence models do not represent explicitly. This paper presents SymbolFusion, a relation-constrained neuro-symbolic graph framework that predicts imagery and symbolic candidates from raw text, validates three literary relation types through an auditable node-relation compatibility function, and propagates emotion only across valid graph paths. No gold symbolic or affective annotation is supplied during validation or testing. The 3,120-passage corpus is divided by literary work into 2,184 training, 468 validation, and 468 test passages, with an additional author-disjoint evaluation. Across ten seeds, SymbolFusion obtains 91.3% symbolic graph agreement, an emotion consistency score of 0.880, a symbolic-emotion alignment score of 0.820, and a transition-point F1 of 0.842. Against the strongest constraint-based baseline, the SGA gain is 2.2% points (95% bootstrap CI 1.5–2.9; p < 0.001). Controlled comparisons among unconstrained attention, soft penalties, hard masks, and permuted relation types isolate the effect of literary compatibility rules. Standard metrics, repeated latency measurements, expert-rating correlations, reproducibility settings, and error analysis delimit the method as an inspectable computational aid rather than a replacement for human literary interpretation.