Active learning with bagging for NLP tasks

Abstract

Supervised classifiers are limited by the annotated corpora available. Active learning is a way to circumvent this bottleneck, reducing the number of annotated examples required. In this paper, we analyze the benefits of active learning combined with bagging applied to Quotation Start, Noun Phrase Chunking and Text Chunking tasks. We employ query-by-committee as query strategy to actively select examples to be annotated. By using these techniques, we achieve reductions up to 62.50% on the annotation effort depending on the task to obtain the same quality as in passive supervised learning. © 2012 Springer-Verlag GmbH.