A Sarcasm Detection Method Based on the Fusion Representation of Sentiment Contrast and Deep Semantic

Sarcasm is a rhetorical form with strong emotional connotations, frequently appearing in comments and opinion texts on social media. This paper proposes a novel sarcasm detection method based on the sentiment contrast and deep semantic representation fusion. Firstly, to capture the affective signature of sarcasm, we design two context-based measures for sentiment intensity to establish 1) intra-sentential sentiment contrast representations among multiple clauses within a sentence, and 2) inter-sentential sentiment contrast representations between a sentence and its preceding context. Next, we fine-tune a large language model to get the sentence representations and design a squeeze-and-excitation module along with residual connections to achieve deep semantic representations of sentences. Finally, we build a multi-perspective fusion representation that integrates the inter-clausal sentiment contrast representation, the inter-sentential sentiment contrast representation and the deep semantic representation of the sentence to achieve effective sarcasm detection. Experimental results indicate that the proposed method achieves best performance compared with the existing baseline methods on multiple datasets.