Sentiment Analysis System Design And Implementation On Chinese Review Based On Product Features

With the quickening pace of e-commerce’ pace, people’s lives have already closely connected with online shopping. On the condition that the Internet has gradually developing in an open building model which is based on the principle “Take users as the center, be easy to participate”, all the merchants of e-commerce websites allow customers to comment the goods and services they buy from the websites, these comments expressed by people of all kinds of emotional color and emotional tendencies. For a businessman, analyzing the customer’s comment texts is the best way to know the customers satisfaction degree with the products and services. Potential users can also understand the public’s view of a matter or product through these subjective comments. But with the vast amount of information, both merchants and customers do not have enough time to collect and process vast amounts of goods comments. So that, it is urgent to need the computers to help users and merchants quickly obtain and sort out these evaluation related information, thus arising the Sentiment analysis technology.In this paper, the sentiment analysis methods for product comment text were studied. At present, the product reviews opinion mining in sentence granularity is mainly based on the product’ features, the main research content is as follows1) Use the collocation extraction techniques of natural language processing to extract implicit product features and then manual finish similar features clustering.2) Put forward a new data model: Chinese Sentiment treewarehouse, which contains 11855 statement parsing trees and 215154 phrases fine-grained tags which generated by the parse trees.3) Design a modified recursive neural network model to handle these new data sets. Compared with previous neural network model with random vector as input, we introduced the Word2 vec of Google to generate high quality word vectors. We make Word2 vec generate the word vector containing the context related information after it has learned thousands of articles which received participle processing. Word2 vec can not only map the word to the K dimension vector space, but also correspond word vector operations to semantics. Finally, the experiments show that our model can efficiently capture the Influence of different levels of negative words on emotional polarity, which makes considerable improvement in the positive and negative classification of simple sentences.4) Based on the work above and web crawling technology, a multi-level network sentiment analysis system is designed and implemented.