融入注意力机制的双通道神经网络命名实体识别
Dual Channel Neural Network Named Entity Recognition Integrated with Attention Mechanism
陶露1
作者信息
- 1. 皖江工学院 电气信息工程学院,安徽 马鞍山 243000
- 折叠
摘要
针对深度学习方法识别命名实体缺乏丰富语义信息及冗余信息对命名实体识别的影响问题,提出一种融入注意力机制的双通道神经网络命名实体识别模型(BW-ATT-NERM).首先使用Word2vec和BERT两种语言模型将文本转换成相应的向量表示形式作为模型输入;然后采用BiGRU网络提取文本特征向量,文本特征向量利用注意力机制生成特征向量的加权语义表示;最后利用CRF训练和学习文本特征向量与输出标签之间的关系,预测和输出最佳标签序列.实验结果表明:BW-ATT-NERM模型平均准确率、平均召回率、平均F1 值达到 95.97%,94.26%,95.11%,与基准识别模型(LSTM-CRF)相比,识别效果更加明显.
Abstract
Aiming at the problem of the lack of rich semantic information and the influence of redundant informa-tion on named entity recognition by deep learning methods,a dual channel neural network named entity recogni-tion model(BW-ATT-NERM)incorporating attention mechanism is proposed.Firstly,use Word2vec and BERT language models to convert text into corresponding vector representations as model inputs;then,the BiGRU net-work is used to extract text feature vectors,which generate weighted semantic representations of feature vectors u-sing attention mechanisms;finally,using CRF to train and learn the relationship between text feature vectors and output labels,predict and output the optimal label sequence.The experimental results show that the average accu-racy,average recall,and average F1 value of the BW-ATT-NERM model reach 95.97%,94.26%,and 95.11%,respectively.Compared with the benchmark recognition model(LSTM-CRF),the recognition effect is more signif-icant.
关键词
命名实体识别/双通道/双向GRU/注意力机制Key words
named entity recognition/dual channel/bidirectional GRU/attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据