Large-Scale Language Models for Sarcasm Detection with Data Augmentation

Data is the foundation upon which deep learning models are built. It shapes their understanding, generalization abilities, and performance. However, high-quality and abundant data is usually not readily available in practice, which hinders the success of many deep learning projects. Even though there are already several data augmentation techniques available to expand the dataset, the critical concern still remains: will the augmented data be suitable for all tasks, particularly those that are highly reliant on linguistic and contextual features, such as sarcasm detection? To figure this out, in this paper, we enlarged the iSarcasmEval dataset, the first shared task to target intended sarcasm detection, using 1) manually labelled data from similar task and 2) augmented data generated with the noising data augmentation technique, and evaluated the quality of the expanded data on a variety of popular large-scale language models, such as BERT, XLNet and their variants. Experimental results demonstrated that human labelled data can significantly boost the detection accuracy of these language models, while noising data augmentation technique yields negative results.