In terms of mapping social interactions and content algorithmically irony, sarcasm, and 'insider culture/messages' can cause communications actual messages to be hidden in plain sight from AI and skew actual meaning.
Yes,
algorithms are great at picking up concepts via text dialogues, but
things like sarcasm and irony are much harder to detect.
Here's a sarcastic statement that would have trouble being detected:
> I reallllllllllly want to study for my exam tonight…
While
the algorithm may detect and ascribe values to each word it can't
easily understand the concept. In this case the algo would likely think
the user actually wanted to study for the exam a lot.
Another example of sarcasm as a defense:
> You're soooooo smart…
Again, the algo would think that the person was calling someone smart with emphasis; when in reality it's the opposite.
The
difficult is that this type of thing leads to a gray area in terms of
detection where it's hard to actually quantify what is being meant (by
an AI). They may ascribe multiple periods (…) to sarcasm, but it
wouldn't always be correct.
-----
As for Irony:
It'd
have to be a chaining of conflicting terms across a post, I'd presume.
But, it'd also have to factor in if one of the posts was sarcastic.
For irony:
"I'm so tired of these fascist assholes crashing our parades :C"
followed by a post of
"Got this racist fuck kicked off campus! No hate here, assholes!"
Which
has irony embedded in multiple layers, as demonstrated by the
iron-fisted approach to the original problem, and the comparison of
directing a string of identifying terms like "no hate" to a directed
response of ",assholes!" while also somehow recognizing that the order
of distaste towards "fascist assholes" later links to the mentioned
"assholes" in the second comment.
It's rather interesting really,
seems like it'd be a string of codes trying to determine a certain
length that would constitute irony by identifying positive and negative
words, associating them with subjects, and then comparing those subjects
to text-speak inflections such as drawn out lettering. The biggest
problems would likely come from "txt spk" where words aren't so easily
identified as words or misspellings and can create multiple meanings
because people who use shorthand type are usually incredibly stupid.
-----
I'll beak down the first irony example for fun (trying to think in linguistic/language/AI terms):
> "I'm so tired of these fascist assholes crashing our parades :C"
I'm = Self reference
so = Emphasis on next word
tired = Connotation of worn down, energy expended
of these facist = what the self referencer is tired of
crashing = verb associated with destroying or bringing down
our parades = our herd mentality group/club that they are in opposition of
:C= memetic text face indicating a frown
The
interesting part here is that using a verb like crashing an AI would
need to associate likely synonyms based on contextual clues. I think it
could easily be done in code. Synonyms and like-word-replacement seem to
be key.