EviLink-NA: Calibrated Open-Set Spatiotemporal–Relational Fusion for Narrator Entity Linking in Hadith Isnads

Entity linking of Hadith narrators is notoriously difficult due to homonymous names, orthographic drift, and incomplete metadata. We present EviLink-NA, a calibrated, openset framework for narrator attribution in isnads that separates retrieval from verification. First, a multi-view retriever builds a compact Top-k candidate set using orthographic and phonetic normalization, KG anchoring, and a bi-encoder reranker. Second, five independent evidential channels—textual compatibility, temporal feasibility, geospatial feasibility, path-consistency in the transmission network, and reliability priors—score each candidate as non-negative “evidence.” We then perform Dirichlet evidential fusion to produce calibrated class probabilities alongside an explicit uncertainty mass, enabling a principled abstain decision for open-set and low-evidence cases. For isnads with multiple ambiguous mentions, we add a lightweight factor-graph layer with pairwise potentials (directionality, temporal, and geospatial coherence) and run loopy belief propagation to enforce joint consistency. Across large-scale narrator corpora, EviLink-NA achieves state-of-the-art closed-set Accuracy@1 and substantially improves calibration (ECE, Brier, NLL; see Table 4), while sharply reducing false positives in near-synonym clusters. The framework is modular, missingness-robust, and delivers risk-controllable coverage through decision thresholds, supporting practical curation workflows for digital Hadith studies.