Scholarlink

<p>Entity Extraction, as a core task in NLP, its core goal is to accurately extract entities with specific meanings from unstructured text information and classify and define the categories to which they belong. Chinese NER remains challenging due to ambiguous boundaries, strong contextual dependence, and nested entities. We develop DBBG-Net, which equips GlobalPointer with a Boundary-aware Bidirectional Feature Enhancement (BBFE) module for Chinese NER. The model adopts DeBERTa as the backbone encoder for contextual representation learning, employs BiGRU to capture sequential features, and leverages BBFE to strengthen boundary-aware representations. GlobalPointer is then used to predict entity spans directly. Tests on the CLUENER2020 and Weibo NER benchmarks show that our method consistently performs better than strong baselines. On CLUENER2020, it achieves higher overall scores than traditional sequence-labeling models across several metrics. On Weibo, it also boosts both precision and recall when the text contains colloquial expressions and internet slang, suggesting that our boundary-aware bidirectional feature enhancement works well and generalizes across domains. We make the DBBG-Net project open source, and the repository is here: <a href="https://github.com/dp4094/DBBG-net/">https://github.com/dp4094/DBBG-net/</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DBBG-Net: a boundary-aware bidirectional feature enhancement model for Chinese named entity recognition

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Bin Zhao,
  • Pan Dong,
  • Yilong Gao,
  • Chuhang Ye,
  • Yunlong Ji

摘要

Entity Extraction, as a core task in NLP, its core goal is to accurately extract entities with specific meanings from unstructured text information and classify and define the categories to which they belong. Chinese NER remains challenging due to ambiguous boundaries, strong contextual dependence, and nested entities. We develop DBBG-Net, which equips GlobalPointer with a Boundary-aware Bidirectional Feature Enhancement (BBFE) module for Chinese NER. The model adopts DeBERTa as the backbone encoder for contextual representation learning, employs BiGRU to capture sequential features, and leverages BBFE to strengthen boundary-aware representations. GlobalPointer is then used to predict entity spans directly. Tests on the CLUENER2020 and Weibo NER benchmarks show that our method consistently performs better than strong baselines. On CLUENER2020, it achieves higher overall scores than traditional sequence-labeling models across several metrics. On Weibo, it also boosts both precision and recall when the text contains colloquial expressions and internet slang, suggesting that our boundary-aware bidirectional feature enhancement works well and generalizes across domains. We make the DBBG-Net project open source, and the repository is here: https://github.com/dp4094/DBBG-net/.