Title: The DNA of sarcasm and its implication in cross-domain tasks Authors:  Havana Rika - The Academic College of Tel Aviv - Yaffo (Israel) [presenting]
Dan Vilenchik - Ben-Gurion University (Israel)
David Ben-Michael - Open University of Israel (Israel)
Abstract: Sarcasm is a form of figurative language where the speaker intends to convey a message implicitly. The explicit meaning of a sarcastic statement is often contradictory to the implicit meaning and heavily reliant on the context. In some instances, the sarcastic intent may be accentuated by the speaker's tone of voice, which is absent in the written text. As a result, detecting sarcasm is a non-trivial task for humans, let alone for automatic methods. The problem of sarcasm detection has traditionally been approached as a binary classification task, aiming to predict whether a given text contains sarcasm or not. In recent years, there has been a growing trend to address sarcasm detection using deep neural network (DNN) models. These models have been solely evaluated through the in-domain method and, unfortunately, demonstrate poor performance in the cross-domain evaluation method (training on one and testing on another). The purpose is to explain the low cross-domain performance through the many shades of sarcasm. For example, some remarks are more humorous than others, while others may be more toxic; in short, not all sarcastic comments were born equal. The differences are identified and presented among a variety of well-known sarcasm datasets. Using these insights, a data enrichment procedure is guided that significantly improves cross-domain performance up to an additive 13\% in F1 score without requiring labelled data.