Chinese Product Sentiment Classification Based On Sentiment Strengths

In recent years, with the rapid development of e-commerce and mobile Internet, mobile surfingand shopping online have become a part of people’s daily life. Several online shopping companies such as Taobao, Jingdong, Amazon have showed a good development trend. After a receipt of the online shopping goods, people usually will make an evaluation for the shopping. The evaluation objects are almost the merchant’s goods and services. Those product reviews often express people’s joy, anger, sadness and other emotions. Product reviews that include sentiment have important value for the application of product survey analysis, information forecasting, and so on.We focus on two aspects in this paper: the extensionof sentiment lexicon and the sentiment classification for product reviews.(1) The extension of sentiment lexiconCurrently there are some limitations in the manuallymanaged dictionary, such as identifying some irregular deformation words and new words. Word presentation can solve this problem. We can use a large-scale unlabeled data set (the corpus from Baidu’ baike) to train word presentation, and then get a set of low-dimensional word vectors. It is easy to find similar words for a given word by word presentation.(2) The sentiment classification based on sentiment strengths for product reviewsCurrently a number of dictionary-based approacheshave the following deficiency: the emotional phrase extractiondoes not consider the nonlinear interaction of adverbs,the search for a negator, and the ambiguityof sentiment words. We have introduced a Chinese sentiment strength lexicon and a lot of grammar information and propose a sentiment word strength based method called sentiDP.The method can solve the problem of ambiguous sentiment wordsand has a good semantic compositionality.The experimental results show that the sentiDP method’s precision and F１value are better than the other two methodswith the recall rate about the book data7% higher than the baseline method. This means that sentiDP can effectively deal with product reviews in the field of hotel, notebook and book. In terms of the extension of sentiment lexicon, the accuracy of the word presentationbased approach is18%higher than the NGD based one.