Sarcasm is hard to detect for chatbots.
Does modeling of conversation context help in sarcasm detection?
“...conceptually, it is important to keep the Sarcastic Turns and the context (respective Prior Turns and/or Succeeding Turn) separate for as long as possible, as the model is designed to recognize a possible inherent incongruity between each.
Long Short-Term Memory networks with sentence-level attention achieved significant improvement when using Prior Turn as context for all the datasets. Using the Succeeding Turn did not prove to be helpful for our datasets.”
In other words, using prior volley and inherent incongruity metric with the current volley is useful in sarcasm detection. The key is finding a method for “inherent incongruity” analysis to discriminate between a simple non sequitur and actual sarcasm related to the Prior Turn.
Link to pdf