Graph Based Semi-supervised Sentiment Classification

With the rapid development of the Internet today, it has become an urgent problem to use big data better in the field of information technology. Also, Text data, the carrier of human knowledge, is significant to human beings. Then the way to use a large number of unlabeled samples to improve the accuracy of sentiment classification has become more and more important. This paper mainly studied semi-supervised classification algorithm based on graph from the following three aspects:(1) A semi-supervised sentiment classification algorithm based on Clustering Kernel was proposed. We firstly constructed a weighted undirected graph in the labeled samples and unlabeled samples, computed the kernel functions, and used this for training SVM classifier to complete the classification work. This method directly fused the information contained by unlabeled samples into the kernel, without multiple classifiers. Experiments show that the algorithm is better than these semi-supervised sentiment classification algorithms based on Self-learning SVM or Co-training SVM in classification accuracy, and is adaptive on different datasets.(2) A semi-supervised sentiment classification algorithm based on clustering kernel and Graph Mincuts Algorithm(GMA) was given. Firstly a graph in all samples was projected, with different weights for labeled and unlabeled points. By solving the Clustering Kernel on the graph, a new data representation was achieved. To get better performance, added the new form into the GMA algorithm. This algorithm fits the basic assumptions of semi-supervised learning more, with the effective use of the information contained in the unlabeled samples. Experiments show that it is superior to other semi supervised sentiment classification algorithm in classification accuracy, and has good results in different data sets.(3) Two improved semi-supervised sentiment classification algorithms based on Graph-of-words model were raised. In the text preprocessing, we used Graph-of-words text representation model to vector the text and solved the sentiment classification problem by combining the classification algorithm proposed above. The algorithm further improves the classification accuracy of semi-supervised sentiment classification algorithm by considering the effect of text representation method for sentiment classification algorithm. Experimental results show that the semi-supervised sentiment classification algorithms based on Graph-of-words model are better than these algorithms listed in classification accuracy.