A weakly supervised approach to Chinese sentiment classification using partitioned self-training

摘要

With the rapid evolution of documents on the World Wide Web which express opinions, there exists an increasing demand for developing such a sentiment analysis technique that can easily adapt to new domains with minimum supervision. This article introduces a novel weakly supervised approach for Chinese sentiment classification. The approach applies a variant of self-training algorithm on two partitions split from test dataset, and combines classification results of the two partitions into a pseudo-labelled training set and an unlabelled test set, then trains an initial classifier on the pseudo-labelled training set and adopts a standard self-learning cycle to obtain the overall classification results. Experiments on the four datasets from two domains show that our approach has competitive advantages over baseline approaches; it even outperforms the supervised approach in some of the datasets despite using no labelled documents.