This Algorithm Was Trained to Know When You’re Being Sarcastic on Twitter

It was found to be more accurate than actual people


Social media has surely allowed us to communicate with people with ease and the use of emoji has enabled us to share our sentiment about a post. However, a typical human conversation has a wide spectrum where some aspects of speech are difficult to express fully in text form.

Sarcasm is one of those remarks people use in general speech that can be easily identified when someone speaks due to the change in tone or facial expression, but can be difficult when passing it along text. People have found a workaround by using certain emojis that give the message a sarcastic tone and researchers are hinging on that.

MIT researchers have developed an algorithm that analyzes tweets that detect sarcasm and emotional subtext in general apparently better than most people.

The algorithm, Deep Moji, uses deep learning to train the neural network to recognize patterns using a large amount of data. They collected 55 billion tweets and selected 1.2 billion which contained a combination of 64 popular emoji.

They hinged on emojis to help the system read tweets for emotion and this gave it a headstart in teaching it to recognize sarcasm. They first trained the system to predict what emoji would be used with a particular message (happy, sad, funny etc) and then later, it was taught to identify sarcasm using an existing data set of labelled examples.

The above gif shows how they used the system to test how it would act while being used by humans. They found out that it was 82% accurate at identifying sarcasm, which was higher than the score for the human volunteers (76%).

The researchers also have come up with a website that shows the emoji part of the system. when you type in a sentence, it shows the words in the sentence that it thinks have an emotional impact and gives you the appropriate emoji.

It is kind of cool that computers are now gradually getting better at sensing human emotion, which has been one of the biggest stumbling block for AI for a while. Knowing the right sentiment being shared by people in your network can be useful for ad targeting campaigns for example and this kind of research points to a future where that would be the norm.