Reducing Sparsity in Sentiment Analysis Data using Novel Dimensionality Reduction Approaches

No aspect of our mental life is more important to the quality and meaning of our existence than emotions and sentiments. Recently researches have introduced many Machine Learning approaches to analyse sentiment from public blogs, social networks, etc. Due to the sparse and high-dimensional textual datasets one needs Feature Selection before applying classifiers. The scope of my thesis are Dimensionality Reduction techniques for predicting one of the two opposite sentiments, specifically for Polarity Classification. The greatest challenge for Text Classification problems in general is data sparsity. Especially it is for Bag-of-words model, where the document is represented by the number of occurrences of each term in the vocabulary. Hence it can be hard for a classifier to understand the relationships between all the words in the initial vocabulary when training set is not large enough. In this thesis I investigate possible steps required to decrease the sparsity: setting the vocabulary, using sentiment dictionaries, choosing data representation and Dimensionality Reduction methods and their underlying strategies. I describe fast and intuitive unsupervised and supervised tf-idf scores for Feature Ranking. In addition, Word Clustering algorithm for merging the words with very close semantical meaning is introduced. By clustering semantically close words we decrease the feature space with minimum loss of information compared to Feature Selection, where we simply omit the features. Polarity Classification problem is investigated on two datasets: SemEval 2013 Twitter Sentiment Analysis and KDD Project Excitement Prediction using Extreme Learning Machine. Best performance for both datasets was achieved by using the proposed Word Clustering and supervised tf-idf score with 20 times less features than original vocabulary size