Latent information mining for semi-supervised sentiment classification in catering reviews

In this paper, we aim to address the issue that semi-supervised learning is prone to be influenced by the quality and quantity of initial seeds. In order to expand the initial labeled data, we select credible samples from unlabeled data by a proposed bilateral latent information miner. The miner can extract information from unlabeled data for both positive and negative class respectively. Then we train a semi-supervised sentiment learner with the expanded labeled data. Experiments show that the proposed method can achieve a better and more robust performance than other methods, especially when initial labeled data is randomly selected in a small size.