Sarcasm detection enhancement through a perspectivist features approach

Abstract

Sarcasm that uses positive language to convey a negative sentiment, criticism or disapproval is among the most challenging NLP tasks. This mismatch between form and intent forces models to rely on contextual, speaker expectations, and often cultural cues rather than literal sentiment. In this paper, we aim to detect sarcastic samples among those of the ArSarcasm dataset. The dataset covers more than $\mathbf{1 0, 5 4 7}$ samples annotated at sarcasm, sentiment and dialect levels. The annotators used True and False tags to determine the presence and absence of sarcasm within a sample. For sentiment, they have employed positive, negative and neutral labels. The dialect label covers several varieties such as MSA, Egyptian, Levantine, Gulf, and Maghrebi dialects. We have performed several experiments using a combination of a transformer-based approach with sarcastic features, leveraging both linguistic cues and deep contextual representations to more effectively identify sarcastic expressions. Our system is based on AraBERT enhanced with GANs, then we have altered the pretrained model AraBERT vector by adding characteristics to it. GANs are used to improve the distinction between various classes based on an adversarial network. Whereas, the perspectivist sentiment labels help in detecting the presence of mixed sentiment which may be a sarcasm indicator. The dialect, in turn, may be an indicator of the dominant sentiment within a specific dialect. The obtained results were pertinent; However, since the studied corpus consists of highly underrepresented sarcastic samples, we investigated the effect of the combination of conflicting sentiment labels as well as dialect. The obtained results are promising since we have reached an accuracy of 88% and a sarcastic $\mathbf{F}$-measure of 0.64, which outperforms the state of the art on the same dataset by 11 and 2 points, respectively.