Multi-Agent Collaborative Semantic Enrichment with Calibration, Active Learning, and HCI-Guided Review
Keywords:
Semantic enrichment, multi-agent systems, entity linking, calibration, active learning, human-in-the-loop, HCI, selective prediction, interoperabilityAbstract
The literature on semantic enrichment has advanced rapidly in modeling techniques and knowledge integration, as cataloged by Shayegan & Mohammad [12], yet less attention has been paid to multi-agent collaboration and operational reliability in human-in-the-loop (HITL) settings. We present and evaluate a practical multi-agent pipeline comprising a retriever agent, generator agent (alias/pattern synthesis), cross-encoder linker, temperature-based calibrator, and an HCI-guided review UI that implements selective prediction. We report cross-domain results on news, technical reports, and clinical-style narratives, showing sustained gains in macro-F1, improved calibration (lower ECE), faster reviewer throughput, and measurable reductions in UI slips. We provide four reproducible figures and two tables, together with actionable guidance on threshold tuning, backlog control, and value-set curation.
