Turkish Stance Detection on Social Media Using BERT Models: A Case Study of Stray Animals Law

Recently, social media has transformed into an essential platform for information dissemination, allowing individuals to articulate their opinions and apprehensions on a wide array of subjects. Stance detection, which refers to the automated examination of text to ascertain the author’s perspective regarding a specific proposition or subject, has emerged as a significant area of research. Within the scope of this study, a Turkish-labeled dataset was created to determine the stances of social media users regarding the Stray Animals Law and various pre-trained BERT models were fine-tuned on this dataset, four of which were Turkish (BERTurk 32k and 128k, ConvBERTurk and ConvBERTurk mC4), one multilingual (mBERT) and one base (BERT-Base). The BERTurk 128k model outperformed other BERT models by achieving a remarkable accuracy rate of 87.10%, along with 87.11% precision, 87.10% recall, and 87.10% F1 score. In conclusion, this study has accomplished a contribution in the limited field of Turkish stance detection research by comparing various BERT models in the context of Turkish texts that has not been previously undertaken to our knowledge. The promising results that were obtained from this and similar studies could contribute to the automatic extraction of public opinions, thereby assisting policymakers in formulating efficient policies.