Using Skipgrams, Bigrams, and Part of Speech Features for Sentiment Classification of Twitter Messages

In this paper, we consider the problem of sentiment classification of English Twitter messages using machine learning techniques. We systematically evaluate the use of different feature types on the performance of two text classification methods: Naive Bayes (NB) and Support Vector Machines (SVM). Our goal is threefold: (1) to investigate whether or not partof-speech (POS) features are useful for this task, (2) to study the effectiveness of sparse phrasal features (bigrams and skipgrams) to capture sentiment information, and (3) to investigate the impact of combining unigrams with phrasal features on the classification’s performance. For this purpose we conducted a series of classification experiments. Our results show that POS features are useful for this task while phrasal features could improve the performance of the classification only when combined with unigrams.