Co-training Based on Multi-type Text Features

Sentiment classification is intended to classify the sentiment color categories expressed by the text. This paper illustrates the sentiment classification method based on the semi-supervised algorithm that aims to improve performance by using unlabeled data. This paper proposes a novel co-training style semi-supervised learning algorithm in order to improve semi-supervised learning ability. In our algorithm, there are three classifiers trained on the original labeled data, where the text representation for each classifier is unigram, bigram, and word2vec, respectively. And then these classifiers can use unlabeled data to update themselves. In detail, any of two classifiers have the same label, then add the new labeled data to a training set of the third classifier. By combining different types of features, our algorithm can extract text information from multiple views which contribute to sentiment classification. In addition, this algorithm doesn’t require redundant and sufficient perspectives. Experiments show that our algorithm is superior to traditional co-training algorithm and partial semi-supervised learning algorithm.