A Decade of Semantic Enrichment: Methods, Calibration, and Human- in-the-Loop Workflows (A Systematic Survey and Taxonomy)
Keywords:
Semantic enrichment, entity linking, knowledge graphs, calibration, selective prediction, human-in-the-loop, HCI, active learning, reproducibility, surveyAbstract
We provide a comprehensive survey of semantic enrichment—spanning mention detection, candidate generation, entity linking, calibration, and human-in-the-loop review—covering 2014–2025. Building on the bibliometric baseline of Shayegan & Mohammad [1], we (i) assemble a five-layer taxonomy that organizes modeling and operations choices; (ii) quantify method adoption trends and reliability practices across 248 papers and 21 production reports; (iii) analyze dataset and metric heterogeneity across news, technical, clinical-like, and GLAM domains; and (iv) distill HCI patterns that convert calibrated uncertainty into effective workflows. Our synthesis shows convergence toward dense retrieval with cross-encoder re-ranking, growing use of temperature scaling and selective prediction, and measurable throughput gains from rationale-first UIs paired with batch triage. We release four reproducible figures and two summary tables to guide practitioners building trustworthy enrichment pipelines.
