基于多核卷积神经网络(BERT+Multi-CNN+CRF)的水产医学嵌套命名实体识别

刘巨升, 于红, 杨惠宁, 邵立铭, 宋奇书, 李光宇, 张思佳, 孙华

大连海洋大学学报 ›› 2022, Vol. 37 ›› Issue (3) : 524-530.

PDF(3220 KB)
PDF(3220 KB)
大连海洋大学学报 ›› 2022, Vol. 37 ›› Issue (3) : 524-530. DOI: 10.16535/j.cnki.dlhyxb.2022-047

基于多核卷积神经网络(BERT+Multi-CNN+CRF)的水产医学嵌套命名实体识别

  • 刘巨升,于红*,杨惠宁,邵立铭,宋奇书,李光宇,张思佳,孙华
作者信息 +

Recognition of nested named entities in aquature medicine based on multi-kernel convolution (BERT+Multi-CNN+CRF)

  • LIU Jusheng,YU Hong*,YANG Huining,SHAO Liming,SONG Qishu,LI Guangyu,ZHANG Sijia,SUN Hua
Author information +
文章历史 +

摘要

为解决水产医学命名实体识别中存在的嵌套实体识别准确率不高的问题,提出一种基于多核卷积的命名实体识别模型(BERT+Multi-CNN+CRF),采用多核卷积神经网络提取嵌套实体特征,通过BERT(bidirectional encoder representations from transformers)方法对输入语料进行预训练,丰富嵌套实体位置向量信息,获得嵌套实体输入特征矩阵,将提取特征矩阵与输入特征矩阵融合,以增强嵌套实体的特征表示,并进行不同模型的对比试验。结果表明,本文中提出的BERT+Multi-CNN+CRF模型,在水产医学嵌套命名实体识别任务中的准确率、召回率和F1值分别为88.04%、88.92%和88.48%,与识别准确率较高的BERT+BiLSTM+ATT+CRF模型相比,分别提高了2.25%、3.23%和2.74%。研究表明,本文中提出的BERT+Multi-CNN+CRF模型可有效解决水产医学嵌套实体识别准确率不高的问题,是一种有效的水产医学嵌套命名实体识别方法。

Abstract

To address the problem of low accuracy of nested named entity recognition in named entity recognition in aquature medicine,a named entity recognition method is proposed based on multi-core convolutional neural networks (BERT+Multi-CNN+CRF).Multi-core convolutional neural networks are used to extract nested entity features,while BERT (bidirectional encoder representations from transformers) is used to pre-train the input corpus to obtain nested entity input feature matrix and enhance the position vector of nested entity.Then the extracted feature matrix is fused with the input feature matrix to enhance the feature representation of nested entities.Comparison experiments of different models verified the recognition effectiveness of the proposed model,with the accuracy of 88.04%,recall of 88.92% and F1 values of 88.48% in the proposed BERT+Multi-CNN+CRF model in the nested NER task in aquature medicine.Compared with the BERT+BiLSTM+ATT+CRF model which has good recognition ability,the accuracy,recall,and F1 value of the proposed model are improved by 2.25%,3.23%,and 2.74% respectively.The finding shows that the model proposed in this paper can effectively solve the problem of low accuracy of NER in aquaculture medicine,and is an effective method for nested NER recognition in aquatic medicine.

关键词

水产医学 / BERT / 嵌套命名实体识别 / 卷积神经网络 / 多卷积核

Key words

aquaculture medicine / BERT / nested named entity recognition / convolutional neural network / multiple convolution kernel

引用本文

导出引用
刘巨升, 于红, 杨惠宁, 邵立铭, 宋奇书, 李光宇, 张思佳, 孙华. 基于多核卷积神经网络(BERT+Multi-CNN+CRF)的水产医学嵌套命名实体识别[J]. 大连海洋大学学报, 2022, 37(3): 524-530 https://doi.org/10.16535/j.cnki.dlhyxb.2022-047
LIU Jusheng, YU Hong, YANG Huining, SHAO Liming, SONG Qishu, LI Guangyu, ZHANG Sijia, SUN Hua. Recognition of nested named entities in aquature medicine based on multi-kernel convolution (BERT+Multi-CNN+CRF)[J]. Journal of Dalian Fisheries University, 2022, 37(3): 524-530 https://doi.org/10.16535/j.cnki.dlhyxb.2022-047
中图分类号: S 932.2    TP 391   

基金

设施渔业教育部重点实验室开放课题(2021-MOEKLECA-KF-05);国家自然科学基金(61802046)

PDF(3220 KB)

9384

Accesses

0

Citation

Detail

段落导航
相关文章

/