Main Article Content
Emotion identification from text can help boost the effectiveness of sentiment analysis models. Sarcasm is one of the more difficult emotions to detect, particularly in textual data. Even though several models for detecting sarcasm have been presented, their performance falls way short of that of other emotion detection models. As a result, few strategies have been introduced in the paper that helped to enhance the performance of sarcasm detection models. To compare performance, the model was tested using the TweetEval benchmark dataset. On the TweetEval benchmark, the technique proposed in this paper has established a new state-of-the-art. Besides the low performance, interpretability of existing sarcasm detection models are lacking compared to other emotion detection models like hate speech and anger. Therefore, an attention-based interpretability technique has been proposed in this paper that interprets the token importance for a certain decision of sarcasm detection model. The results of the interpretability technique aid in our comprehension of the contextual embeddings of the input tokens that the model has paid the greatest attention to while making a particular decision which outperforms existing transformer-based interpretability techniques, particularly in terms of visualisations.