简介：Asimpleandeffectivegreedyalgorithmforimageapproximationisproposed.Basedonthematchingpursuitapproach,itischaracterizedbyareducedcomputationalcomplexitybenefitingfromtwomajormodifications.First,ititerativelyfindsanapproximationbyselectingMatomsinsteadofoneatatime.Second,theinnerproductcomputationsareconfinedwithinonlyafractionofdictionaryatomsateachiteration.Themodificationsareimplementedveryefficientlyduetothespatialincoherenceofthedictionary.Experimentalresultsshowthatcomparedwithfullsearchmatchingpursuit,theproposedalgorithmachievesaspeed-upgainof14.4～36.7timeswhilemaintainingtheapproximationquality.

简介：ThisletterpresentsanewchunkingmethodbasedonMaximumEntropy(ME)modelwithN-foldtemplatecorrectionmodel.Firsttwotypesofmachinelearningmodelsaredescribed.Basedontheanalysisofthetwomodels,thenthechunkingmodelwhichcombinestheprofitsofconditionalprobabilitymodelandrulebasedmodelisproposed.Theselectionoffeaturesandruletemplatesinthechunkingmodelisdiscussed.ExperimentalresultsfortheCoNLL-2000corpusshowthatthisapproachachievesimpressiveaccuracyintermsoftheF-score:92.93%.ComparedwiththeMEmodelandMEMarkovmodel,thenewchunkingmodelachievesbetterperformance.