A Sentiment Classification Model Based on Deep Learning and Two Level Stacking SVM

EasyChair Preprint 10467

Abstract

Short text contains limited context information, which is too sparse to extract useful features using machine learning algorithm, and short text does not always follow the grammar of written language. Therefore, how to extract short text features effectively and design a better classifier to improve the accuracy of sentiment classification is of great significance. In order to solve above problems, in this paper we proposed sentiment classification method based on deep Learning and two-level stacking SVM ( two-level stacking support vector machine, TS-SVM). In this method, first, CNN (convolutional neural network) based on attention mechanism (CNN-ATT) is used to extract the local features, the Bi-LSTM (Bi-directional Long Short-Term Memory model) based on attention mechanism (Bi-LSTM-ATT) is used to extract the global features, and the fusion features are used as input of the classifier; secondly, in order to further improve the classification accuracy, softmax algorithm is no longer used in this paper, but the TS-SVM classification algorithm is used. The results show that the accuracy of CNN/Bi-LSTM-ATT-TS-SVM is 13.19% and 7.9% higher than that of traditional CNN model on Chinese and English datasets, and 12.55% and 7.68% higher than that of traditional Bi-LSTM model on Chinese and English datasets, respectively. By introducing attention mechanism model the average accuracy of sentiment classification of Chinese and English datasets is improved by 5.05% and 2.83%, and the accuracy of sentiment classification of Chinese and English datasets by introducing two-stage stacking SVM is improved by 7.82% and 4.96%, and the recall rate and F value are also better than other models.

Keyphrases: Attention Mechanism, Bi-LSTM, CNN, SVM, ensemble learning, sentiment classification

@booklet{EasyChair:10467,
  author    = {Fei He and Baozhong He and Can Wu and Maobao Yang},
  title     = {A Sentiment Classification Model Based on Deep Learning and Two Level Stacking SVM},
  howpublished = {EasyChair Preprint 10467},
  year      = {EasyChair, 2023}}