AI reads your tweets and spots when you’re being sarcastic
ok thanks for being a great caring person!
ah got to love that iPhone battery lol
@BernieSanders and obama doing a great job #sarcasm
Without a helpful hashtag, picking up on sarcasm online can be hard even for humans. For literal-minded computers, it’s often a major headache. But now a machine learning system can automatically recognise when individuals are being sarcastic.
Mining people’s comments on social media is big business. Advertisers track people’s attitudes and moods, companies and governments follow public opinion. But people being sarcastic and saying the opposite of what they actually is super tricky to pick up on. So concerned is the US Secret Service that it listed sarcasm detection as a desired feature in a 2014 tender for a social media analytics service.
Clued up or clueless?
Computers can exploit small textual clues, such as use of exclamation marks, to detect sarcasm with some degree of accuracy. But without context, it is hard identify the tone of a comment. “Isn’t Obama great!!” clearly means different things coming from a Republican or a Democrat.
Looking at information like the relationship between a comment’s author and audience or where the comment is posted online makes a big difference, pushing the accuracy up to around 80 per cent. But coding these features by hand is laborious, and selecting which to use depends largely on intuition.
Now Silvio Amir at the University of Lisbon, Portugal, and colleagues have turned to machine learning. They have trained their system to identify sarcasm on Twitter simply by looking at a user’s past tweets. “We can get away without looking at all this external information,” says Amir.
Who you are
Using just these tweets, the system builds up a picture of a person that is rich enough to guess when they are being sarcastic. “It intuitively makes sense,” says Amir. “Tell me what you talk about and I can tell you who you are.”
Amir’s system predicts sarcasm with an accuracy of 87 per cent, slightly better than existing approaches. However, by learning to detect sarcasm without human input, the system should be very easy to use.
Amir also says the approach should work for any language and any online platform where posting history is available. “The key innovation is realising you can build a model of the user merely based on what they have said in the past,” says Amir. The team will present their findings next week at CoNLL, a Google-sponsored conference on natural language processing in Berlin.
Mark Carman at Monash University in Melbourne, who studies sarcasm detection, thinks it would be relatively straightforward to integrate the approach with other types of social media analysis, such as tracking people’s emotions or stock market trends.
What’s more, dealing with sarcasm would be a great help for marketers and customer service teams, says Carman – not to mention virtual assistants like Apple’s Siri.