Computers are getting better at understanding human languages, thanks partly to algorithms that can analyse sentences for positive or negative sentiments, says Ari Rappoport of the Hebrew University of Jerusalem, Israel. But picking up on sarcasm is still a problem. 鈥淚n many cases, sarcasm is difficult even for people to recognise鈥�, never mind computers, he says.
Rappoport and colleagues wrote a sentiment-analysing program. They then trained this software to recognise sarcasm by feeding it sentences that had been flagged up by human reviewers as likely to contain sarcasm.
The team used the trained program to analyse a selection of product reviews on Amazon.com and a random selection of posts on Twitter. Three human volunteers were asked to rate the same material for sarcastic content.
Advertisement
The algorithm agreed with the volunteers 77 per cent of the time for Amazon.com product reviews and 83 per cent of the time for the tweets.
Nice work, really
鈥淚t is a very exciting paper, because it attacks a problem that I didn鈥檛 really think we were ready to make headway on,鈥� says , a natural language processing expert at Cornell University in Ithaca, New York, who was not involved with the study.
, a language researcher at the University of Southern California in Marina del Rey, says that getting people to agree about what constitutes sarcasm from single utterances with little context is difficult. 鈥淩eally all we can say is that for these cases computers are about as good as people at agreeing with other people about what is sarcastic.鈥�
The research was presented at the in Washington DC this week. The tool could be used by marketers to track online public sentiment surrounding brands.
