Bochun: Automatically annotated stance detection dataset for Sorani Kurdish language

This Research presents the first-ever, high-quality, automatically annotated Kurdish stance detection dataset in the Sorani dialect to fill the gap of lacking annotated resources for Kurdish, a low-resource language in Natural Language Processing (NLP). The dataset consists of 2,174 Kurdish news articles—1,410 economic and 764 political—that were originally published in 2024 and 2025, which are recent and topically relevant. By selecting these texts from well-known Kurdish news agencies, content validity and linguistic purity were preserved throughout. Necessary preprocessing techniques are applied. Annotation is carried out in two steps. First, a pattern-recognition method with 2,456 phrases and keywords was applied to determine if the subject of every text fell into the economics or politics category. Next, the position of every article was annotated with an extended lexicon of 4,243 adjectives and verbs, categorized under support, oppose, and neutral. Wherever direct matches were not possible, semantic similarity and zero-shot classification were used as fallback measures. In order to verify the automatic annotation, a team of domain experts manually assessed a representative sample of the annotated texts, with a high inter-annotator agreement score confirming the validity of the approach. The dataset is made available in XLSX (Excel) format, facilitating ease of use and versatility for a variety of research tasks in NLP. Due to its annotated and organized corpus, this dataset is a solid starting point for researchers who are building Kurdish language processing models. The dataset is released publicly to allow other researchers to build upon it and push the limits of NLP system performance on low-resource languages.