Combining Unigrams and Bigrams in Semi-Supervised Text Classification

Unlabeled documents vastly outnumber labeled documents in text classification. For this reason, semi-supervised learning is well suited to the task. Representing text as a combination of unigrams and bigrams has not shown consistent improvements compared to using unigrams in supervised text classification. Therefore, a natural question is whether this finding extends to semi-supervised learning, which provides a different way of combining multiple representations of data. In this paper, we investigate this question experimentally running two semisupervised algorithms, Co-Training and Self-Training, on several text datasets. Our results do not indicate improvements by combining unigrams and bigrams in semi-supervised text classification. In addition, they suggest that this fact may stem from a strong “correlation” between unigrams and bigrams.