TextCNN模型优化与文本分类性能提升的研究
Research on TextCNN Model Optimization and Performance Enhancement for Text Classification
焦雪平 1刘晓群1
作者信息
- 1. 河北建筑工程学院,河北 张家口 075000
- 折叠
摘要
针对TextCNN模型在处理文本时会对关键信息捕捉不足从而影响分类的准确性的问题,提出了一种基于注意力机制的改进方案来提高其性能,并使用多尺度卷积核用于捕获文本中不同粒度的特征,从而提高分类的准确性,随后在注意力加权后的输出上应用dropout操作.通过这种方式,dropout能够在保留注意力信息的同时,减少特定部分对最终输出的过度影响,提高模型的泛化能力.实验结果表明,改进后的TextCNN模型对文本进行分类时准确率达到了92.27％,精确率达到了 92.32％,召回率达到了 92.27％,F1 值达到了 92.28％,优于原始的 TextCNN、TextRNN 和TextRNN-attention模型,且各类别的评价指标都有所提升.
Abstract
A solution based on attention mechanism was proposed to address the issue of insuffi-cient capturing of key information by the TextCNN model,which affects the classification ac-curacy.Multiple-scale convolutional kernels were employed to capture features of different granularities in the text,thereby enhancing the classification accuracy.Subsequently,dropout was applied to the attention-weighted output.Through this approach,dropout could reduce the ex-cessive influence of specific parts on the final output while retaining attention information,thus improving the models generalization ability.Experimental results demonstrate that the improved TextCNN model achieved an accuracy of 92.27％,precision of 92.32％,recall of 92.27％,and Fl score of 92.28％in text classification,surpassing the original TextCNN,TextRNN and Tex-tRNN-attention models.Moreover,performance metrics for each category were also enhanced.
关键词
TextCNN/卷积神经网络/注意力机制/文本分类Key words
TextCNN/convolutional neural network/attention mechanism/text classification引用本文复制引用
出版年
2024
NETL
NSTL
万方数据