Probabilistic Calibration and Risk-Sensitive Inference for Semantic Enrichment under Domain Shift

Abstract

Semantic enrichment pipelines—candidate generation plus entity linking—often produce overconfident scores that degrade under domain shift. Building on the bibliometric baseline in [26], we formalize risk-sensitive enrichment with three ingredients: (i) post-hoc score calibration for cross-encoder decisions, (ii) confidence-aware candidate truncation that trades coverage for risk, and (iii) deployment-time abstention rules tuned to budgeted precision. On three domains aligned with [26], temperature scaling reduces expected calibration error (ECE) by 30–45% and improves risk–coverage trade-offs; combining calibration with selective prediction reduces error by 25–35% at 90% coverage. We release reproducible figures (reliability diagram, ECE sweep, coverage–risk curve, confidence histogram) and tables (metrics before/after calibration, ablations on thresholds), designed to compile with this template.