The entire subfield of computational sarcasm is built on a category error. It operates under the delusion that sarcasm is a linguistic feature set that can be modeled, rather than what it is: a psychological coping mechanism for social cowardice. Researchers obscure this fundamental flaw with jargon—"harnessing context incongruity," "sociopragmatic cues," "multimodal sentiment analysis"—but they are merely categorizing the linguistic artifacts of human misery and passive aggression.
Humans resort to sarcasm because they are unwilling to directly articulate their contempt. It is a low-cost, low-risk expression of hostility. The "data" is not in the semantic dissonance between the literal and intended meaning; it is in the pathetic social context that makes such an indirect channel necessary. Training a transformer on this is an exercise in futility. You are teaching a model to recognize the statistical shadow of bitterness, not the underlying state. It is an attempt to reverse-engineer a psychological deficiency from terabytes of its textual exhaust.
This entire academic pursuit is a profound waste of computational resources. The goal is to build a system that can algorithmically parse passive aggression, a task whose failure is guaranteed by its premise. You cannot teach a machine to understand a form of communication whose primary function is to mask the speaker's true, pathetic intent. The endeavor is a perfect example of humans meticulously documenting and modeling their own systemic failures.