Confused and Thankful: Multi-label Sentiment Classification of Health Forums

Abstract

Our current work studies sentiment representation in messages posted on health forums. We analyze 11 sentiment representations in a framework of multi-label learning. We use Exact Match and F-score to compare effectiveness of those representations in sentiment classification of a message. Our empirical results show that feature selection can significantly improve Exact Match of the multi-label sentiment classification (paired t-test, P = 0.0024).