| 引用本文： |
-
王纪恬,陈艳平,黄蓉,黄瑞章,秦永彬.结合位置感知的命名实体识别方法[J].广西科学,2025,32(1):96-105. [点击复制]
- WANG Jitian,CHEN Yanping,HUANG Rong,HUANG Ruizhang,QIN Yongbin.Named Entity Recognition Methods Combined with Location-Aware[J].Guangxi Sciences,2025,32(1):96-105. [点击复制]
|
|
| |
|
|
| 本文已被：浏览 19次 下载 13次 |
码上扫一扫！ |
| 结合位置感知的命名实体识别方法 |
|
王纪恬1,2,3, 陈艳平1,2,3, 黄蓉1,2,3, 黄瑞章1,2,3, 秦永彬1,2,3
|
|
|
| (1.贵州大学文本计算与认知智能教育部工程研究中心, 贵州贵阳 550025;2.贵州大学公共大数据国家重点实验室, 贵州贵阳 550025;3.贵州大学计算机科学与技术学院, 贵州贵阳 550025) |
|
| 摘要: |
| 命名实体识别(Named Entity Recognition，NER)的性能影响自然语言处理中诸多下游任务。跨度分类是命名实体识别常用的方法，由于其需要枚举每一个跨度，因此存在高复杂度和大量负实例问题。此外，对每个跨度的独立预测不仅忽略了词与词之间的依赖关系和位置信息，而且导致模型获取的语义信息较为单一，从而忽略了全局信息。针对上述问题，本文提出结合位置感知的命名实体识别方法。具体来说，首先使用位置编码增强词与词之间的位置特征，序列融合了绝对位置信息和相对位置信息，从而得到关注语序的语义信息，预测可能的实体边界；然后对候选实体边界进行匹配组合并过滤生成带有标签信息的候选实体实例；最后使用具有局部信息感知的标签注意力机制和多层感知机联合判断候选实体的标签。实验结果表明，本文提出模型在ACE2005、GENIA和CoNLL-2003数据集上的F1分数分别达到90.02%、81.33%和94.52%，该结果充分验证了所提模型在不同数据集上的有效性，进一步证明了其在命名实体识别任务中的优越性能。 |
| 关键词: 命名实体识别 嵌套命名实体识别 边界检测 位置编码 神经网络 |
| DOI：10.13656/j.cnki.gxkx.20250521.011 |
| 投稿时间：2023-08-31修订日期：2023-09-22 |
| 基金项目:国家自然科学基金项目(62166007)和贵州省科技支撑计划项目(〔2022〕277)资助. |
|
| Named Entity Recognition Methods Combined with Location-Aware |
|
WANG Jitian1,2,3, CHEN Yanping1,2,3, HUANG Rong1,2,3, HUANG Ruizhang1,2,3, QIN Yongbin1,2,3
|
| (1.Engineering Research Center of Text Computing and Cognitive Intelligence, Ministry of Education, Guizhou University, Guiyang, Guizhou, 550025, China;2.State Key Laboratory of Public Big Data, Guizhou University, Guiyang, Guizhou, 550025, China;3.School of Computer Science and Technology, Guizhou University, Guiyang, Guizhou, 550025, China) |
| Abstract: |
| The performance of Named Entity Recognition (NER) influences the research on multiple downstream tasks in the natural language field.Span classification is a commonly used method for named entity recognition.The need to enumerate each span results in high complexity and a large number of negative instances.Since each span is predicted separately,the dependency relationship and position information of words in the sequence are ignored,resulting in the semantic information acquired by the model being single and local.To address these problems,a named entity recognition method that incorporates positional encoding.First,position encoding was used to enhance the positional features between words.The sequence of sentences was fused with both absolute and relative positional information.The model thus obtains attention-based representations and predicts potential entity boundaries.Then,the candidate entity boundaries were matched and filtered to generate candidate entity instances with label information.Finally,a label attention mechanism with local information awareness and a multi-layer perceptron was employed to jointly determine the labels of the candidate entities.The F1 score of the model on the ACE2005,GENIA,and CoNLL-2003 datasets reached 90.02%,81.33%,and 94.52%,respectively,which verified the effectiveness of the model on different datasets and the excellent performance in named entity recognition tasks. |
| Key words: named entity recognition nested named entity recognition boundary detection position encoding neural networks |
|
|
|
|
|
