Sarcasm Detection Base on Adaptive Incongruity Extraction Network and Incongruity Cross-Attention

Sarcasm is a linguistic phenomenon indicating difference between literal meanings and implied intentions. It commonly used on blogs, e-commerce platforms, social media. Numerous NLP tasks, such as opinion mining sentiment analysis systems, are hampered by its nature in detection. Traditional techniques concentrated mostly textual incongruity. Recent research demonstrated that the addition of commonsense knowledge into sarcasm detection an effective new method. However, existing cannot effectively capture sentence “incongruity” information or take good advantage external knowledge, resulting imperfect performance. In this work, modules proposed for maximizing utilization text, their interplay. At first, we propose adaptive incongruity extraction module to compute distance each word text knowledge. Two applied respectively, which can obtain two attention matrixes. Therefore, words sequence receives representation with enhanced semantics. Secondly, cross-attention extract corresponding thereby allowing us pick useful addition, improved gate feature fusion determines how much should be considered. Experimental results publicly available datasets demonstrate superiority our method achieving state-of-the-art performance three well enjoying interpretability.