John Rentoul told us in yesterday’s Independent that he had found a tool which analysed what was being said about party leaders on Twitter, taking into account whether the comment was sarcastic or not.
I spoke to Karo Moilanen at the company about how “sentiment analysis” works. He told me the algorithm detects positive and negative sentiments associated with the leaders, and can even recognise a double negative as a positive, for example, “kill bacteria”.
What about sarcasm, I asked, thinking about how Twitter works. “We have a rudimentary sarcasm detector,” he said. “There are patterns which tend to correlate with sarcasm.” But how accurate is it? “Sarcasm is hard for people to detect. Human accuracy can be as low as 40 per cent.”
TheySay “trains” its computer programme by feeding it texts that humans have marked as being sarcastic. “Algorithms can hence learn that sarcasm tends to involve cases in which someone likes something negative,” said Moilanen, “or conflicting or abrupt changes of sentiment between strongly positive and negative words and phrases.” He said that computer algorithms can detect sarcasm between 55 and 95 per cent of the time, depending on the study, with an average of 77 per cent.