A Cross-Language Analysis on Sarcasm Detection

Cross-language sarcasm detection is valuable for assessing if a model effective in one language works equally well in another, potentially saving significant time and effort in developing language-specific models. This study proposes the following null hypothesis: Performance variations in Sarcasm Detection models due to single model component changes (e.g., embeddings or classifiers) are independent of the target language. We evaluated 27 models on English and Bangla datasets, two languages with distinct sentence structures, and analyzed performance differences using the Wilcoxon test. Based on the F1-score, the test revealed p value of 0.21, implying that there are no significant difference between the two datasets, failing to reject the hypothesis. Nonetheless, further research on cross-language sarcasm detection is needed to ensure a comprehensive grasp of this field and its potential applications.