Sarcasm detection integrates topic background and comment sentiment

Sarcasm is an expressive form where intent contradicts the literal meaning. Consequently, detecting sarcasm can offer insight into the speaker's underlying sentiment. Existing studies on topic-oriented sarcasm detection use topic text as the context for comments, which provides limited information due to its brevity and lack of explanations for buzzwords. In this paper, we propose a sarcasm detection method that integrates topic background and comment sentiment. Instead of employing static topic descriptions, we dynamically construct backgrounds based on frequently mentioned buzzwords to provide contextual explanations. We further leverage a pre-trained sentiment model to enhance the representation of sarcastic semantics. Finally, we propose a fusion strategy using gated low-rank bilinear pooling to integrate topic-aware features and enriched sentiment representations while minimising feature redundancy. Experimental results on ToSarcasm dataset demonstrate the effectiveness of our method.