How about apply this in concept / keyword / issue / opinion extraction?

Observations: the observed terms (nouns, verbs, etc.) in a sequence
Hidden states: the concepts / keywords / issues / opinions we wanna retrieve.

First, use an annotated (concepts / keywords / issues / opinions) corpus to train and get a HMM.
Second, in testing data, use the observed term sequence to predict / extract most possible concepts / keywords / issues / opinions.

Reference: HMMs on wikipedia, PowerPoint Slides (from 盧文祥, CSIE, NCKU), PowerPoint Slides (from Gideon Dror)