An Alternative Method for Sentiment Classifcation with Expectation Maximization and Priority Aging
Yeliz Yengi, Mehmet Karayel, Sevinç İlhan Omurca · 2017
Sentiment classification has been an active research topic in recent years due to its potential impact on semantic based text retrieval. In sentiment classification problem, semi-supervised techniques are more important because of the huge and nonsense unlabeled data in the texts. Since the class labels are manually as- signed by experts and the text data are usually difficult to distinguish positive la- beled ones from negatives, which of the unlabeled data points should be labeled before gaining importance. Basically in semi-supervised classification, data are partially labeled and the task is to label the remaining data. In this paper, a novel two-stage semi-supervised learning model for sentiment classification which is based on Ex- pectation Maximization (EM) and priority aging algorithms is proposed. In the pro- posed approach, the priority degrees initially assigned to unlabeled data and then these are labeled due to their priorities by EM. Namely, the most suitable unlabeled data points are joined primarily to the set of labeled data by using the priority aging algorithm. The effectiveness of the proposed approach is demonstrated by using the IMDB dataset. The experiments show that the more desired results are obtained with regard to the results of conventional EM algorithm.