Scholarlink

<p>NER comprises two sub-duties: finding the entity boundaries and sorting them into preset divisions. Most pioneering tactics do not pay much attention to the contribution of location information and fail to capture and effectively represent the traits of the text span, leading to poor generalization to real-world situations. This investigation mainly proposes a spatially sensitive attribute aggregation tactic for entity identification. A new method is recommended that makes the scheme sensitive to traverse length and boundary, with the addition of rotational position coding that characterizes relational spatial information more precisely. Furthermore, based on the combined deep mechanisms developed, such as CLN, dilated convolution, and the multiplicative attention mechanism, traits of spans will be extracted and fused for richer representation. Extensive explorations investigated the efficacy of the recommended tactic based on the Sina Weibo and Onto Notes 4.0 databases, which improved the performance significantly, reaching F1-scores of 68.67% and 80.92%, respectively. These surpass some classic standards such as BERT-SoftMax and BERT-CRF, thus demonstrating the efficacy of the recommended tactic for flat and nested NER duties. Therefore, this study has important implications for the development of robust, flexible schemes for recognizing entities in various languages at different levels.</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

A NER method based on location-aware multi-feature fusion

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Wen Zhou

摘要

NER comprises two sub-duties: finding the entity boundaries and sorting them into preset divisions. Most pioneering tactics do not pay much attention to the contribution of location information and fail to capture and effectively represent the traits of the text span, leading to poor generalization to real-world situations. This investigation mainly proposes a spatially sensitive attribute aggregation tactic for entity identification. A new method is recommended that makes the scheme sensitive to traverse length and boundary, with the addition of rotational position coding that characterizes relational spatial information more precisely. Furthermore, based on the combined deep mechanisms developed, such as CLN, dilated convolution, and the multiplicative attention mechanism, traits of spans will be extracted and fused for richer representation. Extensive explorations investigated the efficacy of the recommended tactic based on the Sina Weibo and Onto Notes 4.0 databases, which improved the performance significantly, reaching F1-scores of 68.67% and 80.92%, respectively. These surpass some classic standards such as BERT-SoftMax and BERT-CRF, thus demonstrating the efficacy of the recommended tactic for flat and nested NER duties. Therefore, this study has important implications for the development of robust, flexible schemes for recognizing entities in various languages at different levels.