Enhancing Chinese Sarcasm Detection with a Large-Scale Manually Annotated Dataset and a Hybrid Deep Learning Model

Sarcasm detection is a critical aspect of sentiment analysis on social media platforms, where expressions often convey meanings that deviate from their literal interpretations. Traditional natural language processing (NLP) models struggle with sarcasm detection, as they frequently fail to capture the nuanced aspects of sentiment and contextual cues. Although Chinese is considered a high-resource language, detecting sarcasm remains challenging due to the absence of large-scale, annotated datasets and sarcasm's complex, context-dependent, and culturally nuanced nature. This paper presents a multi-dimensional approach to sarcasm detection in Chinese, effectively addressing these challenges. We introduce the Chinese TikTok Sarcasm Dataset (CTSD), a relatively large-scale, manually annotated collection of sarcastic comments sourced from Douyin (the Chinese version of TikTok), to address data scarcity. We demonstrate the effectiveness of the new dataset through extensive experiments using state-of-the-art models. Initially, we fine-tune several pre-trained language models (PLMs), and subsequently, we introduce a hybrid deep learning model, BERTWWM-CNN-BiLSTM, which combines the power of BERT-wwm for rich contextual embeddings, Convolutional Neural Networks (CNN) to capture local n-gram features, and Bidirectional Long Short-Term Memory Networks (BiLSTM) for learning long-range dependencies. Experimental results show that BERTWWM-CNN-BiLSTM outperforms baseline models, achieving an F1 score of 79.07% and an AUC of 85.51%, reflecting improvements of 3.58% and 10.63%, respectively. These findings highlight the potential of our approach in enhancing sarcasm detection in social media text.