Identification of sentiment labels based on self-training

Abstract

Traditional methods for sentiment classification based on supervised learning require a large amount of labeled data for training. However, It is hard to obtain enough labeled data because it can be too expensive compared with unlabeled data. In this paper, we propose an identification of sentiment labels based on self-training (ISLS) method that can make full use of the large number of labeled data. We extract sentiment expressions based on sentiment seeds by self-training, learn sentiment words on unlabeled data and annotate unlabeled data. The sentiment expressions include processing and extracting for the negative meaning of the text. The ISLS method avoids the subjective problems of manual annotation. Experiments validate the effectiveness of the proposed ISLS method.