Sarcasm Detection with BERT Based Ensembled Deep Learning Models

Abstract

Sarcasm detection has been a topic of interest in natural language processing (NLP) research for several years and the growing interest of Internet users in social media has heightened researchers' desire to mine the content available, both quantitatively and subjectively. However, because of its ambiguous nature, detecting sarcasm in textual format has been difficult in the Sentiment Analysis specially in multi-domain datasets which is a challenging and new area of study in this field. To address these issues, we propose a novel approach to overcome the limitations of individually trained models on single domain datasets. Our methodology utilizes neural techniques, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Networks (CNN), combined in ensemble models to detect sarcasm on the multi-domain dataset of News Headlines and SARC. To enhance the accuracy and measure the contextual dependencies, the dataset is prepared using BERT embeddings. This integration of models aims to enhance the generalization capabilities and effectively address the limitations encountered when using individually trained models. Our results demonstrate that the proposed ensemble models, incorporating BERT embeddings, achieves superior performance compared to other state-of-the-art models. Our best Weighted Average Ensemble model achieves an accuracy of approximately 92% which is highest in previous methods on multi-domain dataset and comparable to the traditional methods on single domain datasets. Moreover, the utilization of an ensemble model in the study enhances the stability, precision, and predictive capability of the proposed model.