Research On Chinese Short-text Sentiment Analysis

For the problem of analyzing the Chinese short-text sentiment automatically, the online product reviews are used and this paper mainly researches the following three aspects, i.e. the extraction of the effective text feature, the extraction of the modifiers of a certain word in the text, and the extraction of the classification method. For the problem of extracting the effective text feature, three kinds of effective text features are proposed. One combines the degree words and negative words with sentiment bearing words, forming different feature sets. The other is to consider the modifiers plus sentiment bearing words as the basic feature unit. The last is to use the part of speech information during the feature extraction procedure. In the process of constructing the domain dictionary and extracting the basic feature unit, the modifiers of some words are needed to be found. The traditional method is based on the sliding window method, but it can’t find the modifiers of certain word accurately and find all of the modifiers. So the dependency parsing method which is on the basis of analyzing the sentence structure is used. In this way, the relationships between words can be received. So, all of the modifiers of a certain word can be obtained accurately. For the problem of selecting the classification method, two kinds of methods are used. One is based on the supervised machine learning, which utilizes the SVM classifier and SVMperf classifier to complete the classification task. The other one combines the corpus and the background knowledge, and it can be regarded as an effective implementation method which combines the lexicon based method and the machine learning method. This proposed method overcomes the problems in the lexicon based methods merely, which needs different domain dictionary according to the domain of the text, and overcomes the problems in the machine learning based methods merely, which requires a large of manually marked training data consuming a lot of time and energy. Finally, the experimental analysis of the proposed methods on the feature extraction methods and classification methods are made, and good results are obtained.