Sarcasm detection using sentiment conflict recognition network
Yan Liu, Song Jin, Cheng Zhang, Junxiao Wang, Hongtao Liu, Ying Luo · 2025
Sarcasm frequently appears as a form of expression in everyday communication, often used to convey the opposite of the literal meaning of a sentiment. Sarcasm detection has high application value for sentiment analysis, public opinion exploration, and so on. A common form of sarcasm is the conflict in the text description. For example, in the sentence “sprained my foot, what a happy day!”, sprain is a bad thing which makes people unhappy, but the author uses the positive word “happy” to comment on it, which further expresses his unhappiness by teasing. Aiming at this feature, we propose a text sarcasm detection model. Compared with Transformer-based approaches that require extensive computational resources and struggle with short-text contextual modeling, our model employs Long Short-Term Memory (LSTM) networks for text encoding, which demonstrates superior adaptability in processing short textual units like tweets and news headlines. LSTMs sequential processing captures local semantic dependencies (e.g., contrasting adjacent words like happy and sprained), while its linear computational complexity (O(n)) reduces overhead compared to Transformers quadratic cost (O (nš)). The LSTM architecture not only achieves comparable performance with significantly reduced computational consumption but also effectively captures local semantic dependencies crucial for detecting sentiment conflicts. The encoded representations are then fed into the Sentiment Conflict Recognition Network (SCRN) to extract internal contradiction features. We conduct extensive experiments on four benchmark datasets from Twitter and news headlines, where our model outperforms Transformer-based baselines in short-text scenarios while maintaining competitive accuracy. Experimental results show that our model has better sarcasm recognition ability than previous methods. Visualization of the attention layer in SCRN shows that the model can select related words under different sentiment tendencies, which illustrates the interpretability of our model. This combination of efficiency and efficacy makes the LSTM-based architecture particularly suitable for real-world sarcasm detection applications where short text predominates with demanding requirements for computational efficiency.