Improving stance detection accuracy in low-resource languages: a deep learning framework with ParsBERT

Stance detection, the task of identifying the stance or viewpoint expressed in a text, plays a crucial role in understanding the sentiment and credibility of information. However, in low-resource languages such as Persian, the lack of labeled data poses a significant challenge for developing accurate stance detection models. This research article proposes a deep learning approach that leverages BERT-based embeddings and transfer learning techniques to address this challenge. Specifically, we utilize the ParsBERT model, a language-specific BERT model trained on Persian texts, for improved performance on the Persian news stance detection task. In addition, we propose an ensemble classification approach using BERT-based base learners to detect stances in Persian texts. By considering stance detection as a sentence pair classification task and using ParsBERT, we achieve higher accuracy in classifying the stance of Persian texts compared to baseline methods and simpler configurations. Experimental results on a common Persian stance dataset demonstrate the effectiveness of our proposed methods, showcasing the potential of BERT-based embeddings and transfer learning in low-resource languages like Persian. This research contributes to advancing stance detection techniques in Persian text and opens doors for further research in other low-resource languages. The source code and experimental data of this research work will be publicly available at https://github.com/vkiani/stance.