Interpolative self-training approach for sentiment analysis

摘要

Sentiment analysis has become one of the fundamental research areas with an objective of estimating the polarity of text documents. While sentiment analysis requires rich training resources, the number of available labeled documents is limited. The proposed interpolative self-training model is an extension of self-training as one of the most common semi-supervised learning algorithms. The proposed method is based on enlarging learning documents by interpolating data in both the training and the test phase. The method also includes a weighting strategy for data selection in each iteration. The method is evaluated using four Twitter datasets for the task of sentiment analysis. The results indicate that the proposed self-training model successfully outperforms the baseline and the standard self-training approach.