... thesemethods can often find good solutions, both are quitesensitive to the selection of auxiliary information, and making good selections requires significant in-sight.23 Conditional Random FieldsLinear-chain ... distance function ∆, a target expectationˆf, data D, a function f , and a model distribution pθ, the GE criterion objective function term is ∆ˆf, E[f(x)]. For the purposes of this paper, ... predicted label distribution on theunsupervised data match a target distribution. Notethat supervision here consists of the the full distribu-tion over labels (i.e. conditioned on the maximumentropy...