Obligatory party-pooping: this doesn't really seem to be a poem about a spelling checker or corrector.
Naive spelling [auto]correctors correct by text distance, because people almost always make mistakes in text input by typing the right words but making the wrong motions to do so.
Slightly-less-naive autocorrect takes this approach further, and understands that e.g. "yjr" should become "the" because it's the same letters all shifted over by one.
And, as in the submitted article, the best autocorrectors just look at the whole sentence context and try to predict what word the input "should" have been given what it looks like—this is essentially a kind of compressed sensing (like fMRIs use!), though in practice it tends to be baked down to something like markov chains of levenstein automata with back-propagation on w>n.
...whereas this poem is more like the result of a very naive speech to text algorithm, one which parses each word independently without the context of the sentence-so-far.
(Side-note: I'm surprised that speech-to-text algorithms still work mostly in "real time" with only a limited buffer, unable to go back and change anything more than a few words in the past. It's why they fail to recognize names, for example. If speech-to-text algorithms would buffer the entire audio stream for a dictated document, showing an estimate of the output text so far, but continuing to re-estimate the entire document after every word/sentence/paragraph, they'd perform much better. The difference would be on the level of a standard web-renderer's line-at-a-time reflow algorithm, vs. TeX's whole-document reflow.)
Obligatory "Well Actually": the poem isn't trying to demonstrate the technology behind spellcheckers (or correctors). It's demonstrating a near-universal weakness of algorithmic checking/correcting that isn't shared with human proofreaders: when the mistaken word is the correct spelling of another word. The poem takes this to absurd extremes, but it's quite possible to make such mistakes through typos (e.g., "pun" instead of "pin").
Naive spelling [auto]correctors correct by text distance, because people almost always make mistakes in text input by typing the right words but making the wrong motions to do so.
Slightly-less-naive autocorrect takes this approach further, and understands that e.g. "yjr" should become "the" because it's the same letters all shifted over by one.
And, as in the submitted article, the best autocorrectors just look at the whole sentence context and try to predict what word the input "should" have been given what it looks like—this is essentially a kind of compressed sensing (like fMRIs use!), though in practice it tends to be baked down to something like markov chains of levenstein automata with back-propagation on w>n.
...whereas this poem is more like the result of a very naive speech to text algorithm, one which parses each word independently without the context of the sentence-so-far.
(Side-note: I'm surprised that speech-to-text algorithms still work mostly in "real time" with only a limited buffer, unable to go back and change anything more than a few words in the past. It's why they fail to recognize names, for example. If speech-to-text algorithms would buffer the entire audio stream for a dictated document, showing an estimate of the output text so far, but continuing to re-estimate the entire document after every word/sentence/paragraph, they'd perform much better. The difference would be on the level of a standard web-renderer's line-at-a-time reflow algorithm, vs. TeX's whole-document reflow.)