Investigating stance detection in turkish texts with large language models

In today's world, social media platforms have become a significant source of data where individuals can instantly express their opinions on various topics. Due to the vast amount of data shared on these platforms, manual information extraction has become impractical. As a result, the need for automating the process of information extraction is becoming increasingly important. In the field of Natural Language Processing (NLP), stance detection is a task aimed at identifying the stance or attitude of the author or speaker towards a specific topic, event, or discussion within a text or speech. Stance detection is generally approached as a classification problem. In this study, a Turkish-labeled dataset was created to analyze whether a person expresses a positive, negative, or neutral stance towards a given topic. The dataset was collected from platform X and consists of 8779 data points. The dataset includes two separate targets, Target A and Target B, and for each target, the data is labeled as 'Supports', 'Does Not Support', or 'Neither'. The created dataset was used for the stance detection task and evaluated using machine learning methods, including LSTM, GRU, and BERT models. Due to the class imbalance between the targets in the dataset, data augmentation and undersampling techniques were also applied alongside these models. The results revealed that the BERT model outperformed the other models. This study aims to provide valuable insights for understanding user tendencies in social media analysis.