Leveraging Emotional Consistency for Semi-supervised Sentiment Classification

In this work, we exploit the emotional consistency between label information obtained by label propagation and distant supervision to leverage tweet-level sentiment analysis. Existing methods are either relied heavily on sufficient labeled data or sentiment lexicon resources, which are domain-specific in social media. We propose a three-phase approach to build a semi-supervised sentiment classifier for social media data. Our framework leverages on both labeled, unlabeled tweets and social relation graph data. First, we use label propagation to learn propagated labels for unlabeled tweets and partition all tweets into two clusters. Our label propagation is inspired by social science about emotional behaviors of connected users, who tend to hold similar opinions. Second, using sentiment lexicon resources, we use an unsupervised method to obtain noisy labels, which is utilized to train a distant supervision classifier. Next, we determine the relevance of each classifier to the unlabeled tweets, using the label consistency between the clustering given by the propagated tweet labels and the clustering given by these trained sentiment classifiers. Third, we trade-off between using relevance-weighted trained classifiers and the labeled tweet data. Our method outperforms numerous baselines and a social networked sentiment classification method on two real-world Twitter datasets.