Research On Chinese Opinion Sentence Extraction Based On Ensemble Learning

In recent years, the rapid popularization and development of the network make that more and more people can freely express their views online. The mining and orientation analysis of these texts that contains huge potential values can provide decision support for the politics, public opinion analysis, and market shopping guide. However, these opinions in the texts are mainly made up of opinion sentences, so we aim at the research of opinion mining of sentence level in this paper. The major work of this paper includes:(1) The research of opinion sentence extraction based on ensemble learning of opinion elementsWe design an opinion ensemble classification function through analyzing the structure characteristic and the important elements of the opinion sentences from the view of empiricism. We hope that we can get a higher function value of an opinion sentence and a lower function value of a non-opinion sentence through the rational allocation of the weight of all the parameter so that we can separate the two types of data to the greatest extent. The experiment results show that we can get the better performance when we give a higher weight value to the evaluate pattern and the degree word.(2) The research of semi-supervised opinion sentence extraction based on ensemble learning of classifierThe large scale and high quality train set associated with the specific areas is very important to improve the performance of machine learning, however, it is so expensive to do the label work. In this paper, we use the BootStrapping idea to expand the train set and we try to use three classical machine learning methods from the view of rationalism that is Naive Bayes algorithm, Support Vector Machine (SVM) algorithm and Maximum Entropy algorithm combined with semi-supervised ensemble learning ideas to extract the opinion sentences. We propose the method of semi-supervised opinion sentence extraction based on ensemble learning of classifier. Experiment results show that the integrated classifier’s performance is obvious better than the single and the method of semi-supervised opinion sentence extraction based on ensemble learning of classifier can get the comparative results to the5fold cross valid experiment when using a low label rate.(3) We design an experiment system of Chinese opinion sentence extraction based on java platformWe construct a java system of Chinese opinion sentence extraction based on the two methods proposed in this paper. Users can get the corresponding result to the selected method of extraction. It can be used as an experimental platform of Chinese opinion sentence extraction, and also can be used as a sub-system in opinion mining.