Abstract

Schema-flexible named entity recognition (NER) is essential for information extraction in domains where entity type inventories vary across tasks, datasets, or documents. However, existing schema-conditioned methods mainly rely on label names or static type descriptions, which limits their ability to capture document-specific entity realizations, span boundaries, and distinctions among semantically related types. To overcome this limitation, we introduce DCS, a dynamic conceptual semantics framework for schema-flexible NER that generates document-conditioned semantic specifications for target entity types and uses them to guide span-based extraction.

Methodology

DCS is a hybrid generative–discriminative architecture with two modules: a compact semantic describer that generates a document-conditioned description for each target entity type, and a discriminative span extractor that uses those descriptions to predict labeled spans. The describer is refined through extraction-driven feedback to improve downstream span prediction.

Results

Extensive experiments on out-of-domain, multilingual, supervised, and controlled benchmarks show that DCS consistently improves schema-flexible NER, particularly for domain-specific and semantically competing entity types. Ablation and robustness analyses demonstrate that the gains mainly arise from adaptive semantic conditioning rather than longer descriptions or increased architectural complexity.

Citation

@article{xu2026dcs,
  title={DCS: A dynamic conceptual semantics framework for schema-flexible named entity recognition},
  author={Xu, Shiting},
  journal={Neurocomputing},
  volume={704},
  pages={134871},
  year={2026},
  publisher={Elsevier},
  doi={10.1016/j.neucom.2026.134871}
}