Automatic Irony detection for Chinese social media based on BERT-GDPCNN-ATTN model
Xiaofan Qi, Aihua Wu · 2024
Due to the diversity and complexity of irony as a linguistic phenomenon, irony detection has always been a challenging research field in text sentiment analysis, which helps to identify ironic tones and accurately judge the sentiment of the text. In this paper, we propose the BERT-GDPCNN-ATTN model for Chinese irony detection in Sina Weibo. In our model, we adopt the strategy of stacked embedding, using multiple Transformer-based Chinese pretrained models to generate diversified sentence embeddings and document embeddings, and combine them with enhanced Gate DPCNN to learn long-range dependencies within Chinese texts. We also incorporate attention mechanism to more effectively represent Chinese context and better capture the irony tendencies within it. Meanwhile, we find that the traditional binary classification of irony and non-irony is not sufficient to distinguish various types of irony well. Thus, we design another Chinese sub-dataset for four-class classification. This paper first discusses the research status quo of irony detection in the field of NLP and outlines various existing irony detection models.