摘要

In statistical text emotion recognition, semi-supervised learning that can leverage plenty of unlabeled data has drawn much attention in recent years. However, the quality of the training data is typically influenced by some mislabeled samples. In this paper, we present a novel co-training method, namely adaptive multi-view selection (AMVS), to improve labeling accuracy of unlabeled samples for semi-supervised emotion recognition. In particular, two importance distributions are proposed to construct multiple discriminative feature views. One is the distribution of feature emotional strengths, and the other is the importance distribution of view dimensionality. On the basis of these two distributions, several feature views are iteratively selected from the original feature space in a cascaded way, and corresponding base classifiers are trained on these views to build a dynamic and robust ensemble. The experimental results on the real-life dataset consisting of moods posts demonstrate the proposed AMVS outperforms conventional multi-view semi-supervised emotion recognition methods, and that abundant emotional discriminative features could be fully exploited in view selection process.