关键词: 卷积神经网络, 文本情感分析, 词向量, 字向量, 卷积核

Abstract: The single channel Convolutional Neural Network (CNN) cannot fully study the feature information of text with a single perspective. In order to solve the problem, a new Dual-Channel CNN (DCCNN) algorithm was proposed. Firstly, the word vector was trained by Word2Vec, and the semantic information of sentence was obtained by using word vector. Secondly, two different channels were used to carry out convolution operations, one channel was the character vector and the other was the word vector. The fine-grained character vector was used for assisting word vector to capture deep semantic information. Finally, the convolutional kernels of different sizes were used to find higher-level abstract features within the sentence. The experimental results show that, the proposed DCCNN algorithm can accurately identify the sentiment polarity of text, its accuracy and F1 value are above 95%, which are significantly improved compared with the algorithms of logistic regression, Support Vector Machine (SVM) and CNN.

Key words: Convolutional Neural Network (CNN), sentiment analysis, word vector, character vector, convolutional kernel