A novel method for semantic action recognition through learning a pose lexicon is presented in this paper. A pose lexicon comprises a set of semantic poses, a set of visual poses and a probabilistic mapping between visual and semantic poses. This paper assumes that both visual poses and mapping are hidden and proposes a method to simultaneously learn a visual pose model that estimates the likelihood of an observed video frame being generated from hidden visual poses, and a pose lexicon model that establishes the probabilistic mapping between the hidden visual poses and the semantic poses parsed from textual instructions. Specifically, the proposed method consists of two-level hidden Markov models. One level represents the alignment between the visual poses and semantic poses. The other level represents a visual pose sequence and each visual pose is modelled as a Gaussian mixture. An Expectation-maximization algorithm is developed to train a pose lexicon. With the learned lexicon, action classification is formulated as a problem of finding the maximum posterior probability of a given sequence of video frames that follows a given sequence of semantic poses, constrained by the most likely visual pose and the alignment sequences. The proposed method was evaluated on MSRC- 12, WorkoutSU-10, WorkoutUOW-18, Combined-15, Combined- 17 and Combined-50 action datasets using cross-subject, crossdataset, zero-shot and seen/unseen protocols.