An algorithm created by computer science researchers detects sarcasm in tweets better than humans. And it used emojis and millions of tweets to learn how to do it.
The machine brain, called DeepMoji, could help flag hate speech online, according to MIT Technology Review. It could also help researchers understand how information and influence filters through the web.
“What’s more, as machines become smarter, the ability to sense emotion could become an important feature of human-to-machine communication,” wrote MIT Technology Review.
DeepMoji learned to discern the emotions conveyed in a text from more than a billion tweets, according to a paper published Aug. 1. It links emojis to the words and phrases that often accompany them. It then identifies whether a text expresses positivity, negativity or love.
The researchers tested their model against an online human workforce group called Mechanical Turk. The human group appropriately identified emotions in messages 76.1 percent of the time. DeepMoji accurately read the emotions 82.4 percent of the time.
“It is better at capturing the average human sentiment-rating than a single (human) MTurk rater,” the authors wrote.
DeepMoji can recognize slight nuances in language. For example, it applied positive emojis to the sentence “I love mom’s cooking.” About half the time, DeepMoji gave that sentence a savoring-food emoji. It also used a heart and a smiling face with heart eyes.
But for a very similar sentence, DeepMoji ascribed negative emojis. “I love how you never reply back,” a sarcastic sentence, was given emojis such as an angry face, a broken heart and a disappointed face.
Similarly, it understands that the phrase “this is sh--” is negative, but it appropriately recognizes “this is the sh--” as positive.
DeepMoji is still learning to glean emotion, and volunteers with an internet access can help it. An MIT page created for the project allows users to input any sentence and submit it to DeepMoji, which then ascribes five emojis to ascertain the emotions behind the sentence. Users can input how they felt when writing a submitted sentence, providing DeepMoji with more data to learn from.
Bjarke Felbo, an MIT student and co-author of the paper, wrote on Medium that DeepMoji is a small step to more sophisticated emotional analysis. In the future, he and other researchers hope to create more nuanced labels beyond the positive/negative categorizing.
Researchers would also like to distinguish between the emotional content conveyed by a text and the actual emotions felt by an author of a text.
Felbo encouraged people to help DeepMoji gain more knowledge using the MIT page, the data from which will be shared with the research community.
“You can help us improve the field of emotion analysis by telling us how you felt when tweeting,” he wrote.