As previously described,
people speaking face to face can rely upon other cues than language ones;
speakers can roll their eyes, place heavy stress upon certain words, slow their speaking rate etc. This is
obviously not available when reading from text. This leads to the problem of only having lexical factors as
clues to work with, both for the reader to pick up on and for a NLP system to detect. As noted previously,
ironic sarcasm is non-literal, which not only gives problems to designing a system to detect it, but also for
readers to pick up on it. Assuming the author does not explicitly state that the phrase is meant sarcastically,
readers have few clues to pick up on the sarcasm, such as the context of the situation, known as common
ground (Clark, 1996), and the words being used to pick up on the sarcasm. An example of some of the clues
that readers and systems can pick up on are extreme adjectives and adverbs, such as 'absolutely fantastic',
which Utsumi (2000) suggests as being a way of implicitly displaying a negative attitude. As you can tell
from the example I used, this is commonly used to express ironic sarcasm.