基于旋转注意力的命名实体识别模型

A Named Entity Recognition Model Based on Rotational Attention

针对命名实体识别中实体词对关系建模分类性能不高的问题,提出一种基于旋转注意力的命名实体识别方法,该方法首先用BERT模型和Bi-LSTM进行编码,然后用卷积神经网络对输入的文本进行特征提取,接着将提取的特征序列输入旋转注意力模型中计算输出概率,最后使用MLP层对输出进行分类。实验结果表明,本文提出的方法在CADEC、GENIA和CoNLL2003等标准英文数据集上均取得了较好的性能,验证其在命名实体识别任务中的有效性和可行性。

Addressing the issue of inadequate classification accuracy in named entity recognition via entity word-to-relationship modeling,we propose a method using rotational attention.Firstly,the text is encoded using the BERT model and Bi-LSTM,followed by extraction of features from input text using convolutional neural networks.Subsequently,the extracted feature sequence is inputted into the rotational attention model for output probability calculations,and the MLP layer is used for output classification.The study's outcomes affirm the efficacy and feasibility of the technique proposed in this paper,as it successfully yields superior results on mainstream English databases including CADEC,GENIA and CoNLL2003,for named entity recognition exercise.