Stacking Ensemble-Based Approach for Sarcasm Identification with Multiple Contextual Word Embeddings

Prior research has emphasized the efficacy of pre-trained word embedding techniques in gauging and identifying emotions conveyed in text documents. However, relying exclusively on a singular word embedding approach can hinder our capacity to capture the intricate interconnections between words within texts. To address this issue, this paper introduces a stacking ensemble model for sarcasm detection that harnesses the strengths of multiple contextualized word embeddings. The proposed model accomplishes sarcasm recognition through the utilization of a trio of state-of-the-art contextual word embedding techniques, namely XLNet, BERT, and RoBERTa. The model utilizes these three sets of contextualized word embeddings to train a stacking ensemble classifier. This ensemble comprises base-level classifiers, including CNN, BiLSTM, and BiGRU, and a meta-level classifier based on SVM. The effectiveness of the proposed method has been examined using two datasets of different domains: the self-annotated Reddit corpus of political comments (SARC 2.0 pol) and news headlines. Experimental findings indicate that the proposed methodology achieves an accuracy of 80.52% on the SARC 2.0 pol dataset, representing a notable increase of 2.02%. Furthermore, on the news headlines dataset, the proposed approach attains an accuracy of 94.63%, signifying a substantial improvement of 2.73% over existing state-of-the-art methods.