A BIO-Annotated Named Entity Recognition Dataset for Sorani (Central) Kurdish (V1)
Description
This dataset provides a BIO-annotated Named Entity Recognition (NER) corpus for Sorani (Central) Kurdish, comprising 420,896 tokens across 26,365 sentences. Text was drawn from authentic Sorani-language sources, including sports news (yariga.net), general news articles (the KurdishTextCorpus, Hugging Face), and other publicly available Kurdish texts. To improve coverage of rare entity categories — including Blood Type, Military, and Telephone Number — synthetic examples were generated using large language models and subsequently reviewed and corrected by native Kurdish speakers for linguistic fluency and entity boundary accuracy. Each token is labeled using the BIO (Beginning-Inside-Outside) tagging scheme across 15 entity types, mapped to 31 unique BIO tags, spanning general categories (Person, Organization, Location) and domain-specific categories (Attack, Nationality, Military, Blood Type, Compound entities). Annotation was carried out by the authors together with two Kurdish linguists following written guidelines addressing entity boundaries, multi-word spans, and disambiguation of overlapping categories, reaching a Cohen's kappa of 0.86. The file contains three columns: Sentence (sentence identifier), Word (individual token), and Tag (BIO-format entity label). Each row corresponds to one token; consecutive rows sharing a Sentence identifier form a full sentence. This dataset supports the development and evaluation of NER models for Sorani Kurdish and other low-resource, morphologically rich languages. It accompanies the Data in Brief article "A BIO-Annotated Named Entity Recognition Dataset for Sorani (Central) Kurdish.