In any language, sarcasm is nuanced. In emoji, it can be even more confusing, but DeepMoji can help you make sense of that eye-roll or upside down smiley face.
Using a dataset of 1.2 billion tweets containing 64 of the most commonly used emojis, five researchers at MIT trained a deep learning model to understands the nuances of how language is used to express emotions.
In their recently published paper, the researchers explained how DeepMoji’s algorithm was built through interpreting emojis on Twitter. Now trained, the model can translate words into emojis, predicting what emotionally-fuelled emoji was likely included with a particular tweet.
“Just by examining the predictions of our model on the test set it is clear that the model does have an understanding of how the emojis are related,” said Bjarke Felbo, one of the MIT researchers behind DeepMoji, in a Medium post.
Felbo explained if the model can predict which emoji was paired with a given sentence, then it can understand the context-specific emotional content.
“We beat the state of the art [models] across benchmarks for sentiment, emotion and sarcasm detection,” Felbo said.
DeepMoji was put to the test against 10 English-speaking volunteers to see if a machine or a human was better at identifying sarcasm in text. The deep learning model scored 82 per cent, where as the average accuracy of the human volunteers was 76 per cent.
The emoji-based model can capture emotion, sentiment, sarcasm and even slang, although researchers said the model has “many limitations” when it comes to more difficult concepts. While the model learns to group emojis into overall categories, at times they can generate conflicting results without the additional context of tone.
DeepMoji wasn’t just made for fun. As for practical applications, the researchers said AI-powered chatbot services that communicate with humans may benefit from having a more refined understanding of emotional content.
This model has far reaching implications for platforms like Twitter as DeepMoji has the capability of detecting hate speech and racism on social media.
Now the MIT researchers are asking for the public to contribute to the scientific research project by rating the emotions they felt when writing their last three tweets.
To test DeepMoji, researchers have released an online demo where users can type a sentence to see its emotions as emojis.