Erin Renshaw, Christopher J.C. Burges, and Ran Gilad-Bachrach
16 May 2014
We investigate the use of selective classifiers for part-of-speech tagging (POS). The idea is to allow classifiers to abstain on hard instances, passing them to downstream classifiers that may have more context available. In this paper we focus on just the first stage of such a cascade, and ask whether selective classifiers attain the accuracies needed on those instances they accept, given that such instances will not be revisited by downstream processing. We show that a selective classifier that is constructed as an ab-staining committee of two off-the-shelf POS taggers can indeed achieve very high accuracies with modest drops in coverage. We also compute the overall accuracy when all instances are voted on by applying majority vote to the abstentions, and we find that this results in state of the art accuracies, robustly.