A Study of Dependency Features for Chinese Sentiment Classification

摘要

Syntactic dependency features, which encode long-range dependency relations and word order information, have been employed in sentiment classification. However, much of the research has been done in English, and researches conducted on exploring how features based on syntactic dependency relations can be utilized in Chinese sentiment classification are very rare. In this study, we present an empirical study of syntactic dependency features for Chinese sentiment classification. First, we consider two types of feature sets (word unigrams and word-dependency relations), three commonly-used feature weighting schemes (term presence, term frequency, and TF-IDF), and two well-known learning methods (Naive Bayes and SVM) to evaluate the performance of different classifiers. Then, we use ensemble technique to combine different types of features and classification algorithms. Specifically, two types of ensemble methods, namely average combination method and meta-Iearning combination method, are evaluated for two ensemble strategies. Through a wide range of comparative experiments conducted on two widely-used datasets in Chinese sentiment classification, finally, some in-depth discussion is presented and conclusions are drawn about the effectiveness of dependency features for Chinese sentiment classification.