... paper,we analyze the following three model families:In the HMM, the input x is a sequence of wordsand the output y is the corresponding sequence of part -of- speech tags.In the PCFG, the input x ... that the first iteration of EM reinforces the systematic mis-takes of the supervised initializer. In the first E-step, the posterior counts that are computed summarize the predictions of the supervised ... system. If thesematch the empirical counts, then the M-step does notchange the parameters. But if the supervised systempredicts too many JJs, for example, then the M-stepwill update the parameters...