Case-Sensitivity of Classifiers for WSD: Complex Systems Disambiguate Tough Words Better

We present a novel method for improving disambiguation accuracy by building an optimal ensemble (OE) of systems where we predict the best available system for target word using a priori case factors (e.g. amount of training per sense). We report promising results of a series of best-system prediction tests (best prediction accuracy is 0.92) and show that complex/simple systems disambiguate tough/easy words better. The method provides the following benefits: (1) higher disambiguation accuracy for virtually any base systems (current best OE yields close to 2% accuracy gain over Senseval-3 state of the art) and (2) economical way of building more effective ensembles of all types (e.g. optimal, weighted voting and cross-validation based). The method is also highly scalable in that it utilizes readily available factors available for any ambiguous word in any language for estimating word difficulty and defines classifier complexity using known properties only.