On a whim, I fed the plain text from The Faerie Queene into a recurrent neural network (RNN), then generated 500,000 characters from the resulting model. This was no great technical feat; I simply downloaded some software, and let it run for three or four days.
The results were both surprisingly good, and yet still nonsense. An example:
And euill dyde, tell I, that yet dismayd
Surprisd, and death appriall deadly swayd,
Can of the good Sans-peares, and gamzelled cleft
Against them passing, when he found in stone
Vpon the same, to be her pourtracinesse,
My liefest eyes, your weapon man abide,
Vntill the whiles the ground his ciristere
Of his owne ruled wemen still none redrest,
That in straunge brassed mantle words he to haue felt.
The model gets a lot right: it tends to make well-shaped FQ stanzas, although they don’t rhyme like they should; more lines end with some kind of punctuation than not, and stanzas end with terminal punctuation, as I would expect them to; and it seems to want to land the last line of each stanza with the thud of some weighty conclusion.
But for all that, the stanzas just aren't readable. In fact, they're less readable than The Mutable Stanzas. And they offer no help with the rhyme, which seems like the main difficulty of FQ stanzas.
Around the office, we often imaging the possibility of generating the last six (or eighteen, depending on how you figure the “complete” FQ) books of the poem. I'm not sure an RNN is a useful approach, although I suppose it might be possible to take the lines generated by the model, none of which actually occur in FQ, and rearrange them in a manner similar to The Mutable Stanzas.
But perhaps there is value in the output of the model, not at the level of the line or the stanza, but at the level of the word, where it generates suggestive words which, although they don't occur in FQ, point toward themes in the poems. Like,