Abstract
Emojis play a nuanced role in digital communication and have a potential to convey sarcastic intent as they often offer non-explicit and sometimes ambiguous cues. This ambiguity has a potential to fuel hate-speech, trolling, or cyber-bullying under the guise of sarcasm. There have been numerous studies that employ modalities like audio, images, videos, emojis or a combination of modalities to detect sarcasm in online text. There is limited research that focuses solely on the impact of emojis in discerning sarcasm. Therefore, in this work we use popular attention networks to capture if sarcasm classification can be improved when emojis are present in text. We experiment with LSTM, Bi-LSTM, and attention networks and compare the results with the fine-tuned benchmark DeepMoji model. Our experiments demonstrate that the emojis can help improve sarcasm classification. These models outperform the benchmark DeepMoji model on two different test datasets on Matthew's correlation coefficient and Area under the curve metrics. Our proposed models surpass DeepMoji by an increase in 0.22 and 0.25 when compared for MCC and an increase in 13.3 % and 14.76 % for the ROC-AUC metric.