Improving Sentiment Analysis in Twitter Using Sentiment Specific Word Embeddings

Most existing continuous word representation learning algorithms usually only capture the syntactic information in the texts while ignoring the sentiment relations between words. These represenations are not sufficiently effective for sentiment analysis as, in many cases, words with similar syntactic context, having neighboring word vectors can bear opposite sentiment polarity. In this paper, we present a weighted average word embeddings method which incorporates sentiment information in the continuous representation of words based on an adapted version of the delta TFIDF measure. Majority voting was then applied to determine the final polarity involving three machine learning classifiers notably, Support Vector Machine, Maximum Entropy and Naive Bayes. Our experiments show promising results and a significant improvement over unweighted embeddings as well as traditional Term Frequency-Inverse Document Frequency (TFIDF).