一种构建情感标签均衡语料库的主动学习算法

AN ACTIVE LEARNING ALGORITHM FOR CONSTRUCTING CORPUS WITH BALANCED EMOTION LABEL

为提高构建的情感语料库中情感分布的均衡性,提出一种基于主动学习的算法以保持新构建训练集中情感标签的均衡。综合信息性、代表性、多样性和互补性标准于一体,通过文本的情感预测概率和特征统计量逐层筛选样本,利用互补性准则中的标签平衡措施抽取候选样本。该算法可以抑制模型选择高频次情感标签的样本,并促进低频次情感标签的样本选择,以达到情感标签平衡的目的。多标签情感分类实验表明,该算法能有效构造情感标签均衡的文本训练集,并通过所构造的训练集逐步提高文本情感分类的效果。

In order to improve the balance of emotional distribution in the constructed emotion corpus,an active learning based algorithm is proposed to maintain the balance of emotional labels in the newly constructed training.It integrated the informational,representative,diversity and complementarity standards,and selected the samples layer by layer through the emotional prediction probability and feature statistics of the text.The candidate samples were extracted by the label balancing measures in the complementarity criterion.The algorithm can suppress the model to select the samples of the high-frequency emotional labels,and promote the sample selection of the low-frequency emotional labels to achieve the purpose of emotion label balancing.Multi-label emotion classification experiments show that this algorithm can effectively construct a text training set with balanced emotion label,and gradually improve the effect of text sentiment classification through the constructed training set.