摘要
提出了基于XLNet的双通道特征融合文本分类(XLNet-CNN-BiGRU, XLCBG)模型。相对于单模型通道,XLCBG模型通过融合XLNet+CNN和XLNet+BiGRU这2个通道的特征信息,能提取更加丰富的语义特征。XLCBG模型对融合后的特征信息分别采用了Maxpooling、Avgpooling和注意力机制等处理方式,分别提取全局中特征值最大的向量、全局中的均值特征向量、注意力机制的关键特征来代替整个向量,从而使融合特征处理的方式多样化,使最优分类模型的可选择性增多。最后,将当前流行的文本分类模型与XLCBG模型进行了比较实验。实验结果表明:XLCBG-S模型在中文THUCNews数据集上分类性能优于其他模型;XLCBG-Ap模型在英文AG News数据集上分类性能优于其他模型;在英文20NewsGroups数据集上,XLCBG-Att模型在准确率、召回率指标上均优于其他模型,XLCBG-Mp模型在精准率、F1指标上均优于其他模型。
A two-channel feature-fusion text classification(XLNet-CNN-BiGRU,XLCBG)model based on XLNet is proposed.Compared with the single model channel,the XLCBG model can extract richer semantic features by integrating the feature information of XLNet+CNN and XLNet+BiGRU channels to diversify the methods of feature-fusion processing and increase the selectivity of the optimal classification model.The XLCBG model adopts Maxpooling,Avgpooling,and attention mechanism to extract the vector with the largest feature value in the global,the feature vector with the mean value in the global,and the key features of attention mechanism to replace the whole vector,respectively.Finally,the current popular text classification models are compared with the XLCBG model.The experimental results show that the XLCBG-S model exhibits better classification performance than other models on the Chinese THUCNews dataset.At the same time,the XLCBG-Ap model exhibits better classification performance than other models on the English AG News data set.In the English 20NewsGroups data set,the XLCBG-Att model is superior to other models in accuracy and recall rate,and the XLCBG-Mp model is superior to other models in accuracy rate and F1.
作者
孟金旭
单鸿涛
黄润才
闫丰亭
李志伟
郑光远
刘一鸣
石昌通
MENG Jinxu;SHAN Hongtao;HUANG Runcai;YAN Fengting;LI Zhiwei;ZHENG Guangyuan;LIU Yiming;SHI Changtong(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;College of Information Technologhy,Shanghai Jianqiao University,Shanghai 201306,China;Shanghai Qingxi Intelligent Technology Co.,Ltd.,Shanghai 201800,China)
出处
《山东大学学报（理学版）》
CAS
CSCD
北大核心
2023年第5期36-45,共10页
Journal of Shandong University(Natural Science)
基金
国家自然科学基金资助项目(61803255)。
