Performance Evaluation of Pre-trained Models in Sarcasm Detection Task

Sarcasm is a widespread phenomenon in social media such as Twitter or Instagram. As a critical task of Natural Language Processing (NLP), sarcasm detection plays an important role in many domains of semantic analysis, such as stance detection and sentiment analysis. Recently, pre-trained models (PTMs) on large unlabelled corpora have shown excellent performance in various tasks of NLP. PTMs have learned universal language representations and can help researchers avoid training a model from scratch. The goal of our paper is to evaluate the performance of various PTMs in the sarcasm detection task. We evaluate and analyse the performance of several representative PTMs on four well-known sarcasm detection datasets. The experimental results indicate that RoBERTa outperforms other PTMs and it is also better than the best baseline in three datasets. DistilBERT is the best choice for sarcasm detection task when computing resources are limited. However, XLNet may not be suitable for sarcasm detection task. In addition, we implement detailed grid search for four hyperparameters to investigate their impact on PTMs. The results show that learning rate is the most important hyperparameter. Furthermore, we also conduct error analysis by means of several sarcastic sentences to explore the reasons of detection failures, which provides instructive ideas for future research.