EUVIMEDEUVIMED entdecken
Crossref

A Fine‐Grained Geographic Named Entity Recognition Model Integrating Boundary Smoothing and Adaptive Pre‐Training

Haiqi Wang, Baozhong Wang, Yufei Shan, Xueying Li, Yawen Ou, Tong Liu, Jun He, Yuanhao Cao

Transactions in GIS · 2026-06 · 30 (4)

DOI: 10.1111/tgis.70313 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Transactions in GIS, online erschienen am 16.06.2026. Verfasst von Haiqi Wang und 7 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Gefördert von: Natural Science Foundation of Shandong Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the increasing use of social media in sudden disaster events, the accurate extraction of fine‐grained geographic named entities is crucial for improving disaster response efficiency. However, geographic named entity recognition in Chinese faces several challenges, including ambiguous word boundaries, entity ambiguity, and nested entities. To address these issues, this paper proposes BSAP‐GNER, a Chinese fine‐grained geographic named entity recognition method that integrates adaptive pre‐training, boundary smoothing, and adversarial training. Specifically, the model first introduces disaster‐related textual information through adaptive pre‐training to enhance its understanding of domain‐specific semantics. It then adopts a span‐level joint prediction method to collaboratively model the boundary information and semantic features of candidate text spans, thereby improving the recognition of nested entities and entities with complex boundaries. Furthermore, boundary smoothing and adversarial training are incorporated to mitigate the bias caused by hard‐label supervision around ambiguous boundaries and to enhance the robustness of the model against noisy expressions and local perturbations, respectively. Comparative and ablation experiments on FireNER, TrafficNER, and AddressNER demonstrate that: (1) BSAP‐GNER consistently achieves the best performance, yielding F1 improvements of 2.74%–5.90% over traditional sequence classification methods and 0.67%–3.27% over other span‐based classification methods. (2) Each module makes an important contribution to the performance improvement of BSAP‐GNER, while the contribution of each module varies across datasets. Specifically, boundary smoothing has the greatest impact on FireNER, whereas adaptive pre‐training contributes most significantly to TrafficNER and AddressNER.

Förderung

  • Natural Science Foundation of Shandong Province

Zitieren

APA

Wang, H., Wang, B., Shan, Y., Li, X., Ou, Y., Liu, T., He, J., & Cao, Y. (2026). A Fine‐Grained Geographic Named Entity Recognition Model Integrating Boundary Smoothing and Adaptive Pre‐Training. Transactions in GIS, 30(4). https://doi.org/10.1111/tgis.70313

Vancouver

Wang H, Wang B, Shan Y, Li X, Ou Y, Liu T, et al. A Fine‐Grained Geographic Named Entity Recognition Model Integrating Boundary Smoothing and Adaptive Pre‐Training. Transactions in GIS. 2026;30(4). doi:10.1111/tgis.70313

BibTeX
@article{Wang2026,
  author = {Wang, Haiqi and Wang, Baozhong and Shan, Yufei and Li, Xueying and Ou, Yawen and Liu, Tong and He, Jun and Cao, Yuanhao},
  title = {A Fine‐Grained Geographic Named Entity Recognition Model Integrating Boundary Smoothing and Adaptive Pre‐Training},
  journal = {Transactions in GIS},
  year = {2026},
  volume = {30},
  number = {4},
  doi = {10.1111/tgis.70313}
}
Kostenlos starten →