作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于卷积神经网络参数优化的中文情感分析

卷积神经网络模型的训练需要设计者指定大量模型参数,但因模型对各类参数的敏感度不一,导致实验效果不佳。针对上述问题,研究中文文本情感分析,以词向量维度、词向量训练规模、滑动窗口大小和正则化方法等作为不同模型的影响因素,设计单层卷积神经网络,在不同影响因素下分别进行中文情感分类实验,并根据结果得出卷积神经网络在处理中文情感分析时对各类参数的敏感程度和具体的模型参数优化建议。   

  1. (1.华南师范大学 计算机学院,广州 510631; 2.华南师范大学 软件学院,广东 佛山 528225)
  • 收稿日期:2016-08-04 出版日期:2017-08-15 发布日期:2017-08-15
  • 作者简介:王盛玉(1992—),男,硕士研究生,主研方向为自然语言处理、情感分析;曾碧卿,教授;胡翩翩,硕士研究生。
  • 基金资助:
    国家自然科学基金(61503143)。

Chinese Sentiment Analysis Based on Parameter Optimization of Convolutional Neural Network

The training of Convolutional Neural Network(CNN) model requires designers to set a large number of model parameters.Because the sensitivity of the model to various parameters is different,the experimental results are poor.To address the problem,this paper provides an analysis on Chinese text sentiment analysis and designs a one layer CNN with influence factors of different models,including the dimensionality of word vectors,the training scale of word vectors,slide window size,regularization method,and so on.Chinese sentiment classification experiment is conducted under different influence factors.According to the results,the sensitivity degree of the CNN on various parameters when dealing with Chinese sentiment analysis and the optimization of specific model parameters are proposed.   

  1. (1.School of Computer,South China Normal University,Guangzhou 510631,China;2.School of Software,South China Normal University,Foshan,Guangdong 528225,China)
  • Received:2016-08-04 Online:2017-08-15 Published:2017-08-15

摘要: 卷积神经网络模型的训练需要设计者指定大量模型参数,但因模型对各类参数的敏感度不一,导致实验效果不佳。针对上述问题,研究中文文本情感分析,以词向量维度、词向量训练规模、滑动窗口大小和正则化方法等作为不同模型的影响因素,设计单层卷积神经网络,在不同影响因素下分别进行中文情感分类实验,并根据结果得出卷积神经网络在处理中文情感分析时对各类参数的敏感程度和具体的模型参数优化建议。

关键词: 卷积神经网络, 情感分析, 参数优化, 词向量, 深度学习

Abstract: The training of Convolutional Neural Network(CNN) model requires designers to set a large number of model parameters.Because the sensitivity of the model to various parameters is different,the experimental results are poor.To address the problem,this paper provides an analysis on Chinese text sentiment analysis and designs a one layer CNN with influence factors of different models,including the dimensionality of word vectors,the training scale of word vectors,slide window size,regularization method,and so on.Chinese sentiment classification experiment is conducted under different influence factors.According to the results,the sensitivity degree of the CNN on various parameters when dealing with Chinese sentiment analysis and the optimization of specific model parameters are proposed.

Key words: Convolutional Neural Network(CNN), sentiment analysis, parameter optimization, word vector, deep learning

中图分类号: