Sarcasm Detection in Chinese and English Text with Fine-Tuned Large Language Models

Sarcasm detection through automatic methods has been explored by some researchers. While these works cover different scenarios of language usage and leverage multiple methods from conventional machine learning to large language models, the performance of them is usually dependent on specific datasets. In this work, in order to accurately detect sarcasm in a more open setting, Low-Rank Adaption (LoRA) is utilized to fine-tune large language models (LLMs) with sarcastic datasets of different sources and different languages. The fine-tuned models outperform their original versions with more than 20 percent improvement on F1-score. This makes it possible to tailor LLMs for sarcasm detection in a more open setting and apply them in a more practical way to solve specific tasks. The performance of a fine-tuned BERT model is also studied for comparison purposes. In conclusion, the method in this work enables LLMs to be able to accomplish more open sarcasm detection tasks at a practical level.