Progressively Improving Supervised Emotion Classification Through Active Learning

Abstract

Recognizing human emotions from pieces of natural language has been a challenging task in artificial intelligence, for the difficulty of acquiring high-quality training examples. In this paper, we propose a novel method based on active learning to progressively improve the performance of supervised text emotion classification models, with as few human labor as possible in annotating the training examples. Specifically, the active learning algorithm interactively communicates with the supervised emotion classification model to find the potentially most effective training examples from a huge set of unlabeled data and increases the training data by acquiring emotion labels for these examples from the human experts. Our experiment of multi-label emotion classification on Japanese tweets suggests that the proposed method is effective in steadily improving the supervised classification results by incrementally feeding a classification model with the new tweets of well-balanced emotion labels.