Comparative Study of Hybrid and Calibrated Entity Linking for Semantic Enrichment Across Domains

Abstract

Bibliometric evidence [17] indicates rapid growth in semantic enrichment, yet comparative studies that integrate calibration and human-in-the-loop (HITL) operation remain scarce. We present a cross-domain comparison of four systems: (i) BM25+metadata, (ii) SBERT bi-encoder, (iii) domain-adapted bi-encoder+cross-encoder, and (iv) a hybrid that pairs (iii) with calibrated thresholds and lightweight HITL review. Across news, technical reports, and clinical-like narratives, the hybrid achieves the best end-to-end macro-F1 and reliability (lower ECE) with acceptable latency. We contribute four reproducible figures (PR curves, reliability, accuracy–throughput trade-off, error taxonomy) and tables for dataset statistics, metrics, and latency.