Sentiment Uncertainty Measure and Classification of Negative Sentences

Sentiment classification is a powerful technology for social media big data analysis. It is of great importance to predict the sentiment polarity of a sentence, especially a negative sentence that is often used. The negation words and sentiment words play equally important roles in the sentiment classification of negative sentences. A negation word is important when it modifies a sentiment word; but it can also have sentimental implication on its own. The existing methods only consider the negation words when they modify sentiment words. In this paper, a unified classification model based on decision-theoretic rough sets is proposed to deal with the sentiment classification of negative sentences. First, the sentiment value of each clause in a sentence is calculated by several lexicons and the inter-sentence relations. A novel measure of sentiment uncertainty for a sentence is given based on Kullback-Leibler divergence. Then, the negative sentences are represented in terms of four features (initial polarity, sentiment uncertainty, successive punctuations, and sentence type) and especially two negation-related features: single negation and salient adverb. Finally, a novel attribute reduction algorithm based on the decision correlation degree is used to generate the decision rules for sentiment classification of negative sentences. The experimental results show that this model is effective and the sentiment uncertainty measure is helpful to sentiment classification.