Sentiment classification of online reviews: using sentence-based language model

Abstract

With the development of social media, the increasing online reviews of products are greatly influencing the electronic market, making sentiment classification the topic of interest for both industry and academia. This paper develops a sentence-based language model to perform sentiment classification at a fine-grained sentence level. The proposed approach applies a machine learning method to determine the sentiment polarity of a sentence at first, then designs statistical algorithm to compute the weight of the sentence in sentiment classification of the whole document and at last aggregates the weighted sentence to predict the sentiment polarity of document. Besides, experiments are carried out on corpuses in different evaluation domains and languages, and the results demonstrate the effectiveness of the sentence-based approach in obtaining a more accurate result of sentiment classification across different reviews. Furthermore, the experimental results also indicate that the position and the sentiment of a sentence have great impact on predicting the sentiment polarity of document, and corpuses with different evaluative objects, languages and sentiments also greatly influence the performance of sentiment classification. It is believed that these conclusions will be a good inspiration for similar researches.