A cybernetic system doesn't have a Limbic System so "doing poetry" as a human would is impossible at the present.  There's an intriguing avenue of approach that would combine a computationally weighted response of the word, tokens, and their association pathways to words, tokens, and combinations of words and tokens to words, tokens and combinations of words and tokens as part of the standard Comparative Progession Matrix© operations.  Never had the time to explore it.  

So I'm going to collapse discussion of "poetry" and "metaphorical" just to keep this comment within reasonable limits.  

Metaphorical communication, it turns out, is a sub-group of a larger problem.  Essentially processing this mode requires the same general operations as Input Validation.  The utterance must be qualified as "Proper" - let me put it - and not an attempt to degrade the Information in the databases, an attempt to breach data security, and on on.  Without getting too much into the technical details, we can assign the "Metaphorical" label to an utterance when there is a sufficient number of previous usages of the metaphor, the words and tokens comprising the metaphor as well as the (computationally derived) Semantics of the words and tokens making "sense¹" given the context of the utterance and some other post-input validation processing.  Some of this latter can occur in Real Time.  The rest has to be down during the 'wait time' between keystrokes or during 'down time' when the system is not being bombarded by user inputs.  

The standard techniques for parsing poetry are, as you know, worthless.  The Robert Frost couplet:

"And I have promises to keep,
and miles to go before I sleep."

has been projected [I can find the cite, if you care] to take well over 2,000 human-years to figure-out the meaning doing it "the old fashioned way."  I can show you our system doing it in under two seconds.

But, again, it's not "poetry," for the reasons given above.  Rather it is a cognitive representation of the poem.  These are two different things, as I'm sure you're aware.  Our Metaphor processing is, roughly, the same: it's not the "metaphor" as such, but the computationally derived cognitive representation of the metaphor.  The system cannot "do" poetry or metaphor but it provides a damn good mimicry of so doing.

Turns out, this is Good Enough for most practical purposes.  Most people don't "do" poetry either and metaphorical usage by humans is highly stereotyped and subject to strict limitations.  (OK.  Subject to some n-dimensional - where n is 1.585, in most cases - shadowing of the phenomenological as given by the final Interpretant comprising the proper significant effect of the Representamems resulting in a final Interpretant comprising the proper significant effect of a cybernetically derived and validated semantic Representamem OF the utterance."  Ta-dah & whoopie.)  These "strict limitations" are a necessary feature as well as an emergent property of Language.  It's through semiotic, syntactical, and semantic Sorities humans overlap and it is only through these over-lappings that humans can communicate or 'range-in' to communicate.  Same with a computer.  As a last resort, our system will do what a human does ... ask "WTF are you on about?"  (Tho' more politely.  ;-)  Limited testing indicates a human user accepts this and actually expects it.  This fits the interface within the common mode of communication, it also provides a psychological benefit by making the computer seem less Other; makes the system more "human" thus less threatening ... oddly.  

The second step, is to determine the where in various hard and virtual association pathways the tokens, words, and final Representamem of the utterance lie. Without getting into details, the result is a human retrievable "emotive sense" of the metaphor as vectored through the association pathways and other final and intermediate tokens, words, and final Representamems it encounters along the way.  And their pathways and so, recursively onwards.

Amusingly, since this process is all input based, by allowing the system to input to itself the system can, through the same process, derive its own, novel, metaphors.  We've only paper tested this function and the results were ... unique ... enough to place the whole thing in the 'Whenever' file.  

¹  This would require a longer exposition than I can get into.  Suffice to say, any utterance that "makes sense" eventually cycles around to previously supplied and/or previously inferred, by the system, Cognitive Junction© -- the "narrative semantics" eventually is "validated" by the "message semiotics."  That's not quite right, much more to it, does gets the point across?  

No one could have predicted