Unlabeled Text Classification Optimization Algorithm Based on Active Self-Paced Learning
Tingyi Zheng, Li Wang · 2018
This paper aims to introduce an algorithm for learning from unlabeled and very few labeled text based on the combination of convolutional neural network (CNN), one-vs-all SVM classifier and Active Self-Paced Learning(ASPL). The algorithm first initialize the classifier using a few annotated samples, and extract text features using CNN. It then rank the unlabeled samples according to their importance weight v. The top-ranked samples will be get, and form these samples into high-confidence sample set. In addition, we consider that a few annotations may contain incorrectly annotated samples, then, we re-rank them and select Top-6 ones with lowest prediction scores to verify these annotations. Experimental results shows that the accuracy of optimized text classifier can be improved by unlabeled or very few labeled text data. This algorithm can provide a more effective way for unlabeled or few labeled text samples classification.