Combining Semantic and Prior Polarity for Boosting Twitter Sentiment Analysis

Twitter sentiment analysis offers organizations an ability to monitor public feeling towards the products and events related to them in real time. Most existing researches for Twitter sentiment analysis are focused on the extraction of sentiment feature of lexical and syntactic feature that are expressed explicitly through words, emoticons, exclamation marks etc, although sentiment implicitly expressed via latent contextual semantic relations, dependencies among words in tweets are ignored. In this paper, we introduce distributed representation of sentence that can capture co-occurrence statistics and contextual semantic relations of words in tweets, and represent a tweet via a fixed size feature vector. We used the feature vector as sentence semantic feature for the tweet. We combined semantic feature, prior polarity score feature and n-grams feature as sentiment feature set of tweets, and incorporated the feature set into Support Vector Machines(SVM) model training and predicting sentiment classification label. We used six Twitter datasets in our evaluation and compared the performance against n-grams model baseline. Results show the superior performance of our method in accuracy sentiment classification.