Research And Implementation Of Text Sentiment Analysis Of Power Grid

With the increasing popularity of the Internet, the network has gradually become the most important information carrier in modern society. Mining the emotional information in web texts, especially the negative information, can help the government and enterprises to find the public opinion trend and take timely countermeasures.This paper takes the texts related to State Grid as the research object.We represent the text information by Vector Space Model and Word Embedding, and propose a sentiment analysis method based on multi-level features, which can reduce the feature dimension because it compromises the advantages of Principal Component Analysis and Dynamic Sentiment Lexicon. Specific word includes:We implemented a complete system, which can crawl the data,preprocess the data, classify the text information, and identify the negative information. The system can collect the data from Tencent Weibo, Sina Weibo and Tianya forum. After the preprocessing of the data,the system can classify the text information and identify the negative information.In order to identify the texts related to State Grid from the corpus,we use Vector Space Model and Word Embedding to represent the text information, and classify the different sets by Support Vector Machine,Decision Tree and Logistic Regression classifier. Aiming to solve the problem that the texts related to State Grid is relatively small, we make a research on classification learning of imbalanced corpus and use the SMOTE sampling method to carry out the corresponding experiments.From the experimental results, we found that the method needs to be further improved.This paper puts forward a method of text sentiment analysis based on multi-level features to identify the negative information in the corpus of State Grid. This method improved the dynamic sentiment lexicon method and absorbed the advantages of Principle Component Analysis to reduce the feature dimension and make the feature vector contain emotion information in different levels. The experimental results show that the proposed method can effectively reduce the dimension of the feature space, and get a better result than the improved dynamic sentiment lexicon method.