The statistical features of words are widely used in natural language processing. This paper summarized eight types of statistical features, and studied the role of these features in extracting sentimental words and classifying product reviews. Different from the multi-dimensions of lexical elements in the vector space models(VSM), this paper only employed these 8 types of statistical features in representation of words or documents, which had the ability that could lower the VSM's dimension and could effectively derive the latent semantic space without expensive time and space complexity of SVD calculation. Sentiment words extraction result show that combining these statistical features and PoS tags of words can achieve much higher extraction accuracy than other methods with precision of 76.4%. Product reviews classification results show that in contrast with sentimental words in constructing the feature space, exclusively using these 8 kinds of statistical features can improve classification precision by 10.8%.

[2019-03-05] Printed Article /doc/2017.09.0913/v2