Convolutional Neural Network For Sentiment Classification Based On Sentiment Special Word Embeddings

Two approaches became popular in sentiment classifcation:dictionary-based meth-ods and statistics-based methods.The dictionary-based methods rely heavily on the es-tablishment of sentiment dictionaries which are costly and can not cope with complex information.Therefore,statistics-based methods have gradually become popular.There are two key steps in dealing with sentiment classification problems for statistics-based methods:quantifying unstructured text data and selecting classifiers for training.Because the traditional method of quantifying structured text data can not capture the sentiment information in the text,we adopt a better method of specific-sentiment word embeddings.We first use a method of unsupervised learning to train word embeddings on a large-scale data set,and then use the specific-sentiment word embedding model to add sentiment to the resulting embeddings.Because the sentiment information belongs to a higher level of information in human language,the traditional classifier is not good enough.Otherwise,convolution operation can extract more abstract and advanced information in text data,so we choose convolution neural network as the classifier.The contribution of this paper is to combine the isolated specific-sentiment word embedding and the convolution neural network.We design a convolution neural network containing one search layer,one convolution layer,one pooling layer and one output layer.We find that the classification results of this method is better than that of the traditional word embeddings or the traditional classification method.