Tree Kernel-Based Sentence-Level Sentiment Classification

With the rapid development of Internet, the amount of information increases in an explosive way, especially the subjective information. How to effectually mine useful information from these subjective texts is an issue currently. Sentiment classification is a way to analyze the subjective information in the text and then mine the opinion.We focus on the sentence-level sentiment classification. On the systematically analyzing the importance and difficulties of the sentence-level sentiment classification, this paper proposes a tree kernel-based approach of sentence-level sentiment classification. It employs the SVM-based convolution tree kernel to automatically capture structural information. We also composite the syntax tree/dependency tree-based features and other flat features to improve the performance.Firstly, we focus on how to apply the structure features from the syntax tree to the sentiment classification and propose a novel approach of sentence-level sentiment classification which apply the tree kernel and composite kernel to the SVM classifier. The experimental results show that the performance of our approach can achieve higher F1 measure than that of the linear kernels.Secondly, we provide two kinds of syntax tree pruning strategies: adjectives-based and sentiment words-based. As for the former, we propose a dynamic window algorithm to optimize the situation when a sentence contains more than one adjective; and for the latter, we introduce the domain-related sentiment words into the classification. The experimental results show that the latter's performance is better than that of the former. Otherwise, the experimental results also show that the tree kernel can achieve higher performance in the implicit sentiment classification. Finally,we proposed a novel approach of sentence-level sentiment classification which apply the tree kernel and dependency tree to the SVM classifier. The pruning strategy of dependency tree is based on sentiment words and denpendency relation filtering. The experimental results show that our pruning strategy is feasible.