Sarcasm Detection with Contextual-Representation Multihop-Attention network

Sarcasm is a delicate and implicit form of expression where what people say often runs counter to what they really mean. With this intentional ambiguity and complexity, sarcasm detection has always been a challenging task, even for humans. The importance of sarcasm detection has been realized by industry and academia. However, existing methodologies for automatic sarcasm detection predominantly rely on lexical and linguistic indicators, which prove to be insufficiently effective in addressing the complexities of sarcasm recognition.This paper focuses on tackling the challenging problem of sarcasm detection by leveraging a novel Contextual-Representation Multihop-Attention (CRMA) network based on bidirectional GRU. CRMA not only understands the ambiguity of sarcasm by the combination of contextual representation and word representation, but also learns the semantic complexity among sentences by the multihop attention mechanism. We perform comprehensive experimental evaluations on IAC dataset collected from popular website.The empirical findings reveal that our proposed framework achieves a substantial performance advantage over existing cutting-edge approaches, while supplementary case analyses further corroborate the robustness of our network architecture in identifying sarcastic expressions.