Exploring the impact of training datasets on Turkish stance detection

Authors: MUHAMMED SAİD ZENGİN, BERK UTKU YENİSEY, MÜCAHİD KUTLU

Abstract: Stance detection has garnered considerable attention from researchers due to its broad range of applications, including fact-checking and social computing. While state-of-the-art stance detection models are usually based on supervised machine learning methods, their effectiveness is heavily reliant on the quality of training data. This problem is more prevalent in stance detection task because the stance of a text is intimately tied to the target under consideration. While numerous datasets exist for stance detection, determining their suitability for a specific target can be challenging. In this work, we focus on Turkish stance detection and explore the impact of training data on the model performance. In particular, we fine-tune BERT model with various datasets and assess their performance when the test data is the same/different compared to the training data in terms of target and domain. In addition, given the scarcity of resources for Turkish stance detection, we investigate i) whether we can use existing datasets in other languages in a cross-lingual setup, and ii) the effectiveness of data augmentation with simple automatic labeling methods. In order to conduct our experiments, we also create new Turkish stance detection datasets for various targets in different domains. In our comprehensive experiments, our findings are as follows. 1) Using training data with multiple targets in the same domain yields high performance as the model is able to learn more characteristics of expressing stance with additional data. 2) The domain of the training data plays a crucial role in achieving high performance. 3) Automatically generated data enhances performance when combined with manually annotated data. 4) Training solely on Turkish data outperforms training with the combination of Turkish and English data. Overall, our study points out the importance of creating Turkish annotated datasets for different domains to achieve high performance in stance detection.

Keywords: Stance detection, natural language processing, Turkish

Full Text: PDF