Detecting Negative Sentiment on Sarcastic Tweets for Sentiment Analysis

Sentiment Analysis (SA) is a fundamental and practical research problem in the field of natural language understanding(NLU). Meanwhile, sarcasm detection is a task to detect sarcasm in textual data. Previous works solve these two problems independently and neglect the fact that sarcasm is omnipresent and non-negligible during sentiment analysis. To explore this issue, in this paper, we formulate a general sentiment Analysis (GSA) problem where sarcastic data could be input and point out the limitations of current mainstream frameworks by systematic investigation. To address the GSA problem, we propose a sarcasm-perceivable SA (Sp-SA) training framework to train a model that is robust to sarcasm and able to achieve state-of-the-art performance. Extensive experiments and detailed analysis demonstrate our Sp-SA framework's effectiveness and interpretability. Code and dataset will be publicly available for future research.