Currently, social networks are an important part of people's lives. Through them, society can be kept informed of the events that occur in the day to day, spend hours of entertainment, and even issue opinions on various issues. In fact, we are constantly giving our opinion through publications, comments, and tweets, and in doing so we use figures of speech that, probably, many times Internet users do not fully grasp, one of these figures is sarcasm.
Sarcasm is often used to imply the opposite of what is said, often with the intention of insulting or ridiculing. However, sarcasm is often not detected even when we are using it in conversation and much less when it is used in texts since it is based mainly on vocal tones, facial expressions, and gestures.
Faced with the difficult task of detecting sarcasm in written texts, DARPA's Information Innovation Office collaborated with researchers at the University of Central Florence to develop a deep learning model that can detect sarcasm in precisely written texts.
According to a press release published by DARPA, the researchers managed to create this sarcasm detector, which identifies words included in tweets or online messages that trigger sarcasm, with the help of recurring neural networks and attention mechanisms. The model tracks dependencies between reference words and then generates a ranking score, which helps to identify whether or not there is sarcasm in the text.
"Essentially, the researchers' focus is on uncovering patterns in the text that indicate sarcasm. Identify the reference words and their relationship to other words that are representative of sarcastic expressions or statements," explained Brian Kettler, program manager at the DARPA Information Innovation Office.
In research published by Entropy magazine, the authors of this work, Ramya Akula and Iván Garibay explained that they experimented with online platforms, social networks, and discussion forums. They chose a series of publications and headlines, which were classified with the help of the sarcasm detector into two categories: those that contained sarcasm and those that did not, thus obtaining extraordinary results.
The accuracy it obtained was varied for each platform it examined, for example, on Twitter it scored 98.7 for accuracy, on Reddit a score of 81.0, and in the headlines a score of 91.8. However, it had some problems when examining the Sarcasm Corpus V2 Dialogues platform, where it only achieved a score of 77.2 for accuracy.
The highly accurate results obtained were mainly due to the model learning to pay attention to specific words that trigger sarcasm, but it also identified punctuation marks that are also important markers of this rhetorical figure.
Given the efficiency with which the detector was able to identify sarcasm, it is hoped that this technology can help to more easily understand the feelings with which opinions are transmitted through social networks. "We would like to understand the sentiment. When people like something or don't like something, and sarcasm can really fool feeling detection ... It is an important technology and allows the machine to better interpret what we are seeing online "Brian Kettler noted.
So far they have indicated that they will continue to work on further improving the deep learning model and it is hoped to be able to use it in languages other than English, so their next challenge is to face the linguistic resources that have been implemented in the language in line.