| High-quality Named Entity Recognition (NER) in the legal domain is often hindered by the scarcity of reliable datasets and the structural noise inherent in legal corpora. To address this, we present a semi-automated Human-in-the-Loop (HITL) architecture designed for the rapid development/enhancement of legal NER datasets. Our pipeline integrates a hybrid named entity extraction engine, combining heuristic-based approaches with a fine-tuned Transformer, and a semantic similarity module within an active learning loop to facilitate named entity annotations. Crucially, a Large Language Model (LLM) framework acts as a neural adjudicator assistant to resolve conflicts and enforce strict entity constraints. We demonstrate the usefulness of the proposed method by building an improved version of the Greek Legal NER dataset expanding the named entity volume by 209\% while significantly reducing the time cost of manual annotation. Finally, we report evaluation results using the new dataset and a diverse set of NER methods under fully supervised and zero/few-shot scenarios. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.