To understand how the HMM learns grammar, we consider the POS. An HMM has the characteristic of clustering. We expect that the same characteristic words in a natural language has the same state in an Ergodic HMM. Therefore, this word grouping is strongly related to the POS.
Therefore, the following method is used to analyze the Ergodic HMM.
Change words into POSs.
( We already know the relationship between words and POSs.)
Sum the probabilities of identical POS's.
To keep the diagram tidy, show only those POSs whose proportion exceeds 5%.