Selective Classifiers for Part-of-Speech Tagging
Erin Renshaw, Christopher J. C. Burges, Ran Gilad-Bachrach · 2014
We investigate the use of selective classi-fiers for part-of-speech tagging (POS). The idea is to allow classifiers to abstain on hard instances, passing them to down-stream classifiers that may have more context available. In this report 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 pro-cessing. We show that a selective classi-fier that is constructed as an abstaining committee of two off-the-shelf POS tag-gers can indeed achieve very high accura-cies with modest drops in coverage. We also compute the overall accuracy when all instances are voted on by applying ma-jority vote to the abstentions, and we find that this results in state of the art accura-cies, robustly.