结合主动学习和自动标注的评价对象抽取方法

Opinion target extraction with active-learning and automatic annotation

提出了结合主动学习和自动标注的评价对象抽取方法。具体实现过程中,首先,利用少量的已标注样本训练分类器,对非标注样本进行测试,获取自动标注结果及其置信度;其次,通过置信度计算每个样本的整体置信度,挑选出低置信度即不确定性高的样本待标注;最后,对待标注样本中置信度低的词语进行人工标注,而置信度高的部分则采用自动标注结果。实验表明,该方法可以在确保抽取性能的同时有效地减小人工标注语料的开销。

An opinion target extraction method combined active-learning and automatic annotation is introduced. Firstly, the results of automatically annotation with the confidence are obtained by using a few of labeled corpus to train the classifier to test the unlabeled samples; secondly, the samples of low confidence is annotated by calculating the confidence of every sample; finally, the words of low confidence in the selected samples is annotated manually, while the others are adopted the results of automatic annotation. The empirical results demonstrate that the proposed method effectively reduces the annotation cost and achieves good performance on opinion target extraction.