BARD10: Bangla Authorship Recognition Dataset

BARD10: Bangla Authorship Recognition Dataset BARD10 is a curated corpus of Bangla-language texts collected from publicly available online Bangla blogs and opinion articles.The dataset was developed for authorship recognition and stylistic analysis within the field of natural language processing (NLP).It contains writings from ten Bangla authors, drawn primarily from digital prose and opinion-style blog posts.The dataset maintains a balanced distribution across all authors, ensuring reliable and unbiased evaluation of both classical and transformer-based language models. Dataset Structure The dataset is distributed in CSV format with three files: BARD10.csv – Complete dataset with all author–text pairs BARD10_train.csv – 80% training split (stratified by author, random_state = 42) BARD10_test.csv – 20% test split (same stratification) Columns: author — The author’s name or label (categorical class). text — The Bangla text sample written by that author. Data Sources Texts were collected from public Bangla blogs, literary archives, and open-access online sources.Each sample was manually verified for author authenticity and curated to ensure balanced representation across the ten authors. Preprocessing Information The released dataset is raw and has not been preprocessed.However, for the experiments described in the accompanying publication, the following preprocessing pipeline was applied (using the BNLP library) to prepare model inputs: Unicode normalization URL and HTML tag removal Punctuation removal Stopword removal (BNLP Bangla stopword list) Whitespace normalization Intended Use BARD10 can be used for: Authorship attribution and verification Stylometric feature analysis Transformer benchmarking (BanglaBERT, mBERT, DistilBERT) Cross-genre and domain adaptation Low-resource language evaluation Dataset Summary Language: Bangla (Bengali) Number of Authors: 10 Genres: Fiction and Opinion File Format: CSV Split Ratio: 80/20 (stratified) Version: 1.0 License: CC BY 4.0 Ethical Statement All texts were collected from publicly accessible sources intended for general readership.No private, confidential, or personally identifying content was included.The dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license for academic and non-commercial research use.