Social Media Text Sentiment Analysis

In recent years,Internet technology is still developing rapidly,resulting in a large number of Internet applications,including social networking.The social networking applications have a large number of users at all times involved in the figures,products,events and other related social media data.The sentiment analysis technique is used to mine the subjective sentiment information in the text.Sentiment analysis on micro-blog as the representative of the social media can develop the potential commercial and social value and it's widely used in the applications such as product information feedback,commodity recommendation algorithm,public opinion monitoring,hot event tracking and so on.This paper mainly focuses on the problem of sentiment classification for social media.The first two chapters introduce the current situation and basic technology of this problem in detail.Then,according to the shortcomings of the related research,we propose our sentiment classification methods in the next three chapters.(1)We propose a sentiment classification method based on machine learning and semantic rules fusion.We use some semantic rules to improve the precision of sentiment orientation measurement of traditional lexicon-based approach.Then we add the rule features to the basic feature set after converting and expanding them.Our method is superior to the general fusion method in both sentiment analysis granularity and feature representation.Experiments show that the method has achieved a greater performance improvement.We ranked first in task of micro-blog sentiment classification in Chinese Opinion Analysis Evaluation-2015(COAE-2015).(2)This paper uses the distant annotating approach to help solve the sentiment classification problem.In Chapter 4,based on the sentiment lexicon construction method of neural network model,we improved this model by using semantic rules and sample weights.The dictionary constructed by the improved method outperforms dictionaries constructed by other methods or manual annotation.We won the first place in task of sentiment words extraction in Chinese Opinion Analysis Evaluation-2016(COAE-2016)based on this dictionary.(3)We propose to improve the performance of classification by ensemble learning methods with distant supervised social media data.First,the Bagging ensemble learning model is proved to be superior to any single model in stability and generalization ability.On this basis,the Stacking ensemble learning model is proposed.This model combines the distant supervised data with the manually supervised data through training on the prediction result of base classifiers and the lexicon features.The experimental results show that the classification performance of this model outperforms the model which only uses lexicon features.