Understanding Sarcasm Through Dataset-Level Linguistic and Semantic Profiling: A Comparative Study of iSarcasm and MMSD2.0

With the widespread use of social media, Natural Language Processing systems tend to lose sight of the actual sentiment of a sarcastic text. Previously conducted sarcasm detection studies primarily aimed at improving the design of model architecture. There is less emphasis on the internal aspects of the datasets. In order to fill this gap, our research proposes a comprehensive framework to analyze sarcasm dataset in terms of five dimensions: Lexical Diversity and Complexity, Sentiment Conflict Intensity, Syntactic Marker Frequency, Contextual Incongruity Density and Pragmatic Cue Disparity. With the help of these measures, we are able to compare two popular sarcasm datasets, iSarcasm and MMSD2.0, to gain insight into their linguistic characteristics. We examined 2,000 samples and found the differences in the lexical richness (0.342 vs 0.381), positive sentiment rates (0.151 vs 0.262), readability rates (0.5081 vs 0.5586) and explicit cues of sarcasm (0.043 vs 0.091), which demonstrates that both datasets represent two different styles of sarcasm and should be handled with different modelling strategies to be reliably detected.