Emojis may be the key to algorithms finally being able to grasp sarcasm. It could mean more effective ways to track public moods and tastes or spot hate speech.
Online marketers and have for several years been using social media posts as a way to track how people are talking about brands and products. Originally that just meant seeing how many times something was mentioned, which had the obvious limitation that it didn’t distinguish between positive and negative mentions.
That led to sentiment analysis, which meant looking at the context of a term being mentioned. Basic versions simply assigned the mention as positive, negative or neutral, with more sophisticated versions trying to decipher specific moods and emotions that the brand or product may have evoked.
The big problem has always been the use of sarcastic and ironic speech, particularly when written in a dry tone. Fortunately a lot of people lack confidence in their own use of sarcasm and undermine it by attaching an emoji.
Researchers at MIT developed an algorithm that was ‘trained’ with 1.2 billion tweets that used at least one of a selection of 64 popular emoji. They then put it up against some standardized sentiment analysis benchmark tests and found that not only was it more accurate than existing algorithms that didn’t take account of emoji, but that it may even have been better at spotting tone than humans.
That claim came from a comparison between ‘DeepMoji’ spotting sarcasm correctly in 82 percent of cases compared with the 76 percent achieved by humans recruited via Mechanical Turk. One big note of caution there is that many of the would-be workers on that site come from countries where English isn’t the primary language, so unless the researchers corrected for that, the workers may have been at a disadvantage in analysing tweets.