Sarcasm Detection as a Catalyst: Improving Stance Detection with Cross-Target Capabilities

--Stance Detection (SD) in social media has become a critical area of interest due to its applications in social, business, and political contexts, leading to increased research within Natural Language Processing (NLP). However, the subtlety, nuance, and complexity of texts sourced from online platforms, often containing sarcasm and figurative language, pose significant challenges for SD algorithms in accurately determining the author's stance. This paper addresses these challenges by employing sarcasm detection as an intermediate-task transfer learning approach specifically designed for SD. Additionally, it tackles the issue of insufficient annotated data for training SD models on new targets by conducting many-to-one Cross-T arget SD (CTSD). The proposed methodology involves fine-tuning BERT and RoBERT a models, followed by sequential concatenation with convolutional layers, Bidirectional Long Short T erm Memory (BiLSTM), and dense layers. Rigorous experiments are conducted on publicly available benchmark datasets to evaluate the effectiveness of our transfer-learning framework. The approach is assessed against various State-Of-The-Art (SOT A) baselines for SD, demonstrating superior performance. Notably, our model outperforms the best SOT A models in both in-domain SD and CTSD tasks, even before the incorporation of sarcasm-detection pre-training. The integration of sarcasm knowledge into the model significantly reduces misclassifications of sarcastic text elements in SD, allowing our model to accurately predict 85% of texts that were previously misclassified without sarcasm-detection pre-training on in-domain SD. This enhancement contributes to an increase in the model's average macro F1-score. The CTSD task achieves performance comparable to that of the in-domain task, despite using a zero-shot fine-tuning approach, curtailing the lack of annotated samples for training unseen targets problem. Furthermore, our experiments reveal that the success of the transfer-learning framework depends on the correlation between the lexical attributes of the intermediate task (sarcasm detection) and the target task (SD).