An Ensemble Method Based on Confidence Probability for Multi-domain Sentiment Classification

Abstract

Multi-domain sentiment classification methods based on ensemble decision attracts more and more attention. These methods avoid collecting a large amount of new training data in target domain and expand aspect of deploying source domain systems. However, these methods face some important issues: the quantity of incorrect pre-labeled data remains high and the fixed weights limit accuracy of the ensemble classifier. Thus, we propose a novel method, named CEC, which integrates the ideas of self-training and co-training into multi-domain sentiment classification. Classification confidence is used to pre-label the data in the target domain. Meanwhile, CEC combines the base classifiers according to classification confidence probabilities when taking a vote for prediction. The experiments show the accuracy of the proposed algorithm has highly improved compared with the baseline algorithms.