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

I use a bit of sentiment analysis to detect sarcastic reactions that can of course not be detected without seeing what they are a reaction to, though applying sentiment analysis on sentences without first dependency parsing is going to be less accurate.

Metrics of probability, like the use of commonly sarcastic words, are only sufficiently accurate when there are enough of them, which is why Twitter sarcasm recognition projects include hidden user data. e.g. if 60% of a user’s comments have been sarcastic in the past, there’s already a 60% chance that any of their next comments will be sarcastic.

The problem with making chatbots capable of detecting sarcasm is that they will then also get it wrong when the user is not being sarcastic. You have to wonder which is worse.

Wouldn’t it help if you kept track of the basic “position” held by the conversant?

Yes and no. Most sarcasm is triggered by the prior remark as it relies on clarity of contrast. Sarcasm relying on the other person’s memory of your real opinion tends to be limited to people who know each other very well. Chatbots generally don’t show their memory to go beyond the previous sentence, so people aren’t likely to engage in long-range sarcasm with them. Similar goes for strangers on Twitter and Reddit, the source for most computational sarcasm research.

There’s also the question of how to log one’s position. Sentiment analysis can provide a rough estimate, as do literal statements of opinion that you could pattern-match (“I like ...”), but you’d also have to extrapolate what topic the position is about. Exact positions like one’s stance on a political decision would require a fairly intricate database and topic extraction, I’d expect, while one’s word choice would probably still be a good indication of sarcasm.