Exploring Uncertain Samples through Active Learning To Enhance Text Emotion Classification

Abstract

The traditional supervised learning of emotion classification models requires large number of training examples, which are usually very expensive to acquire. However, without a careful selection there may exist many irrelevant and duplicate examples for the experts to annotate, which wastes the human labor and affects the learning efficiency. In this paper we propose a novel method to improve the quality of training data and enhance the efficiency of training through active learning. Specifically, we use the information entropy to measure the uncertainty of model predictions on raw text samples, and choose those with the highest uncertainty for human experts to annotate as new training samples. Learning from these examples could most significantly solve these uncertain problems and therefore efficiently improve the model. Our experiment of multi-label emotion classification on Chinese microblog shows that the proposed method can effectively find out emotion statements with high uncertainty, and with these new training examples the emotion classification model gets improved in a steady and efficient manner.