I know this is an old post. But heck.
I have done a character level sentence generator for Esperanto in HTM. See: Esperanto NLP using HTM and my findings
When asking HTM to generate sentences. HTM demonstrates that with a single layer of TM. It can learn a sub part of the Esperanto. Namely all words are commonly ended with -o, -as, -is, -n, etc… (These endings has grammatical meanings in Esperanto) but not the grammar of the entire sentence; such as the -n suffix indicates the object is being acted on.
I’m working on using the hierarchical part of HTM to improve the performance. Hopefully I can come up with better results.