AI learned to recognize sarcasm on Twitter by emoticons

The creators of The Simpsons, as in many other cases, predicted the appearance of this invention long before its creation by real scientists.

Some human actions remain incomprehensible to the computer. But some of the behavioral features of computer systems have learned to recognize, and very well. For example, the IBM Watson cognitive system can define the emotional tone of a letter. Teaching a computer to identify emotions is quite difficult, but possible. Recently, another system was introduced that is capable of understanding sarcasm. In this case, we are talking about sarcasm in messages from Twitter. The developers of this system claim that it can determine the emotional content of the messages of various users better than in most cases the person himself does.

Why is all this necessary? First of all, in order for companies to determine the attitude of social network users to their products and to themselves. Now for this purpose, keywords and some other methods are used. But if the computer can determine the emotions of people sending messages, then this can significantly improve the efficiency of companies. In addition, if the machines confidently determine the emotions of people, it will help users to understand what emotions another person used, who sent, for example, an e-mail message.

From the very beginning, the developers of the "emotional" algorithm wanted to create a system that can detect posts of racist content on Twitter. But soon after the algorithm was ready, the project team realized that it gives a lot of false positives. That is, the machine did not understand, for example, comic or sarcastic messages, and took everything at face value. So there was a need to teach the AI ​​to recognize at least sarcasm.
')
This algorithm uses deep learning, which is becoming more common. As key indicators, experts decided not to use words, but emoticons. Yes, they are now contained in most of the messages, so that you can roughly understand the mood of the tweet itself or another user message in the social network. As soon as the researchers were able to achieve the desired, they realized that the AI ​​is working with a bang.

“Because in the online mode we are not able to use non-verbal cues to help us understand what we are talking about, we began to use emoticons. And the neural network was trained to understand the connection between the emotional content of the tweet and the emoticons contained in this message, ”said Iyad Rahvan, processor from MIT Media lab, head of the research group.



Learning neural network was really big. In order for the algorithm to learn how to correctly identify the relationship between the emotional content of the message and Emoji, the scientists collected a base of 55 billion tweets, then highlighting 1.2 billion messages with emoticons (there are 64 types of popular emoticons). Initially, they taught the system to predict which emoticon should be used in one message or another, depending on whether it is funny, sad, or any other. After that, the system began to train to identify sarcasm by the presence in the message already studied by computer patterns.


As it turned out, the neural network learned to define sarcasm much better than the specialists who developed it. The system correctly identifies sarcasm in 82 cases out of a hundred. Man, on average - in 76 cases out of a hundred.

The neural network was also taught to add emoticons to messages with a certain emotional attitude. And the computer coped well with this task as well. You can try out the algorithm here on this site . Here you can help scientists improve their algorithm by sending tweets with the right emoticons. This makes the work of the neural network even more accurate.

Some experts, who managed to familiarize themselves with the work of scientists, said that using emoticons to train a neural network and create an “emotional algorithm” is a great idea. As for sarcasm, there may be a problem - the fact is that not so many people are able to identify sarcasm. Some do not even know what it is. Therefore, it is difficult to say how useful a neural network can be, which is able to define sarcasm. But since it is already there, it means that someone needs it. In addition, on the basis of this work it is possible to develop other projects, more universal.