Context-aware User Sentiment Analysis In Social Media

In recent years,social media has become significant sources for people to exchange information and express personal opinion and sentiment.Meanwhile,with the development of mobile Internet technology and the widespread use of smartphones,user generated messages in social media usually contain multi-modal contents,in which different modalities are correlated to each other.Most existing research has put the importance on the point of single-modal data.Only a few works are related to multi-modal sentiment analysis,which focus on the fusion of different modalities,but rarely take into the account the impact of context information.However,the length of user generated messages is usually short and lacks explicit sentiment words,whereas context information can depict the environment attributes of user messages and provide supplemental information,which is helpful to user sentiment analysis.In view of the weakness of previous researches,a novel context-aware user sentiment analysis model is proposed from the perspective of users who publish the tweets in social media,in which the semantic correlation of different modalities and the impact of tweet context information are both involved.The main work includes:Firstly,we extracted an effective experiment dataset from a real social media dataset,and fill the missing information by webpage crawling method.On this basis,text preprocessing is carried out to reduce the irregularity of text contents of tweets.In addition,we extracted lowlevel visual features of tweet images associated with sentiment factor from the perspective of psychology.Then bag-of-visual-words model is used to further represent tweet images as visual words.Furthermore,regarding sentiment factor as latent variables hidden in tweets,we proposed a context-aware user sentiment analysis topic model.Semantic correlation is established between different modalities by constraint of overall sentiment distribution and topic in the same tweet.The dependency between tweets and context information including time neighborhood and replies of tweets are also modeled explicitly in this model.After model construction,we derivate its variable sampling rules and parameter updating rules,and designed a sampling algorithm for model inference.Finally,experiments are carried out based on a real social media dataset.Through the comparison and analysis of the experimental results,we can get the following conclusions: The context-based user sentiment analysis model proposed in this thesis can detect the sentiment of multi-modal tweets more efficiently,and both context information used in our model can bring about significant improvement of model performance,which proves the rationality and validity of the sentiment analysis model proposed in this thesis.User sentiment analysis in social media plays a key role in individual behavior analysis,and provides a basis for revealing the patterns of user behaviors.The study of user sentiment analysis has great practical significance in many fields,such as Internet word of mouth marketing,mental health care and stock market analysis.