A Semi-supervised Learning Approach for Microblog Sentiment Classification

Abstract

Most sentiment classification for microblogs are based on supervised learning methods. The performance of these methods heavily relies on carefully chosen training datasets. These datasets usually cannot be too small. This is cumbersome and makes these methods less attractive for practical use. To address this problem, approaches to automatically generate training datasets have been proposed. However, these approaches are usually rule-based, hence they cannot guarantee the diversity of the training datasets. In particular, the huge imbalance between the subjective classes and objective classes in the sentiment of tweets makes it especially difficult to obtain good recall performance for the subjective class. To address this issue, this paper proposes a semi-supervised learning approach for tweet sentiment classification. Experiments show that the performance of our proposed method is significantly better than the previous work.