International Conference on Computational Linguistics · 2012 · 15 citations · 18 references
ComputingComputer ScienceComputational LinguisticsMachine LearningNatural Language ProcessingDimensionality ReductionLanguage ModelingLanguage ModelsPart-of-speech TaggingSpeech RecognitionWord Sense DisambiguationAutomated Text ClassificationGradient DescentStochastic Gradient DescentSubspace LearningText Classification
Inspired by work on robust optimization we introduce a subspace method for learning linear classifiers for natural language processing that are robust to out-of-vocabulary effects. The method is applicable in live-stream settings where new instances may be sampled from different and possibly also previously unseen domains. In text classification and part-of-speech (POS) tagging, robust perceptrons and robust stochastic gradient descent (SGD) with hinge loss achieve average error reductions of up to 18% when evaluated on out-of-domain data.
18