Sentiment analysis for tweets using deep learning

Summary

The aim of this thesis is to address the problem of sentiment analysis in Twit- ter. Twitter is a microblogging and social networking service on which users post and interact with messages known as “tweets”. Tweets were originally restricted to 140 characters but this limit was doubled. Many people use Twitter to express sentiments about a variety of subjects. Sentiment analysis refers to the problem of studying text, messages, posts and reviews uploaded by users on Internet regarding the opinion they have about a product, service, event or person. Analyzing the sentiment is very crucial for companies or other organizations in order to understand the opinion from the people that used a product or a service they produced. In this thesis, we use a Greek dataset from a paper with hotel reviews as benchmark and we compare it with our approach, which uses word embeddings, in order to surpass the accuracy of the paper. Also, we manually annotated a Greek dataset with 4600 tweets and three classes in order to create Machine learning and Deep learning algorithms. We experimented with Support Vector Machine and Logistic Regression algorithms and Deep Learn- ing models like FeedForward, LSTM, BiLSTM and CNN. We compared the Deep Learning models and achieved better results with BiLSTM approach. Moreover we compared our results with other Deep Learning methods that have been done with empasis on Greek language.